Mrutyunjaya Panda · Satchidananda Dehuri ·

Manas Ranjan Patra · Prafulla Kumar Behera ·

George A. Tsihrintzis · Sung-Bae Cho ·
Carlos A. Coello Coello (Eds.)

Communications in Computer and Information Science 1737

Innovations in
Intelligent Computing
and Communication
First International Conference, ICIICC 2022
Bhubaneswar, Odisha, India, December 16–17, 2022
Proceedings
Communications
in Computer and Information Science 1737

Editorial Board Members

Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Ashish Ghosh
Indian Statistical Institute, Kolkata, India
Raquel Oliveira Prates
Federal University of Minas Gerais (UFMG), Belo Horizonte, Brazil
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at https://link.springer.com/bookseries/7899
Mrutyunjaya Panda · Satchidananda Dehuri ·
Manas Ranjan Patra · Prafulla Kumar Behera ·
George A. Tsihrintzis · Sung-Bae Cho ·
Carlos A. Coello Coello (Eds.)

Innovations in
Intelligent Computing
and Communication
First International Conference, ICIICC 2022
Bhubaneswar, Odisha, India, December 16–17, 2022
Proceedings
Editors
Mrutyunjaya Panda Satchidananda Dehuri
Utkal University Fakir Mohan University
Bhubaneswar, Odisha, India Balasore, Odisha, India

Manas Ranjan Patra Prafulla Kumar Behera

Berhampur University Utkal University
Berhampur, Odisha, India Bhubaneswar, Odisha, India

George A. Tsihrintzis Sung-Bae Cho

University of Piraeus Yonsei University
Piraeus, Greece Seoul, Korea (Republic of)

Carlos A. Coello Coello

CINVESTAV-IPN
Mexico City, Mexico

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-3-031-23232-9 ISBN 978-3-031-23233-6 (eBook)
https://doi.org/10.1007/978-3-031-23233-6

© The Editor(s) (if applicable) and The Author(s), under exclusive license
to Springer Nature Switzerland AG 2022
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, expressed or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
 Preface

A hearty welcome to the proceedings of the International Conference on Innovations

in Intelligent Computing and Communications (ICIICC 2022) which was held at Utkal
University, Vani Vihar, Odisha, India, during December 16–17, 2022.
The aim of ICIICC 2022 was to provide a platform for researchers from industry and
academia, practitioners, and scientists across the globe to discuss the latest solutions,
innovative and scientific methods, and applications in intelligent computing, commu-
nications, machine learning, and data analytics to solve the real issues of the today’s
world.
ICIICC 2022 received a total 78 submissions from authors in several counties, and
each submission was reviewed by at least three reviewers in a single-blind process. Based
on the reviewers’ recommendations, finally 31 papers were selected for presentation at
the conference with an acceptance rate of 40%. The conference proceedings are published
in Springer’s CCIS series, which is indexed in Scopus, DBLP, etc. The papers in this
proceedings are divided into three tracks:

• Track 1 - Intelligent Computing, consisting of 12 papers,

• Track 2 - Communications, consisting of 6 papers, and
• Track 3 - Machine Learning and Data Analytics, consisting of 13 papers.

ICIICC 2022 was organized successfully due to support and hard work of many
people. First and foremost, we thank all the authors for their research contributions to
the conference, and for their presentations and discussion during the conference. Our
thanks go to the Technical Program Committee members and reviewers for their sincere
efforts in carefully reviewing the submitted articles. We would also like to thank SERB,
Government of India, and WB-OHEPEE, Utkal University, for providing seminar grants
to conduct the conference successfully. Our special thanks go to the following keynote
speakers for exciting and motivational keynote addresses:

• Ajith Abraham, Machine Intelligence Research Laboratories, USA

• Sung Bae-Cho, Yonsei University, South Korea
• Carlos A. Coello Coello, CINVESTAV-IPN, Mexico
• Aniket Mohanty, University of Auckland, New Zealand
• Umamaheswar A Kakinada, Charter Communications, USA
• Ganapati Panda, CVR Global University, India
• Siba K. Udgata, University of Hyderabad, India
• Saroj Ku. Meher, Indian Statistical Institute, Bangalore, India

We express our sincere thanks to the session chairs and local organizing commit-
tee members for helping us to formulate a rich technical program. We are immensely
grateful to our Honorable Vice-Chancellor, Registrar, Comptroller of Finance, Direc-
tors, and other administrative officers of Utkal University for allowing us to organize the
conference.
vi Preface

Finally, our sincere gratitude goes to Lakhmi C. Jain, Founder of KES International,
UK, for his continuous support and kind guidance in the process of technical collabora-
tions, and to the Springer CCIS series team, for their guidance in bringing the conference
proceedings in a timely manner and with good quality.
We hope that the readers of this proceedings will find it enjoyable.

December 2022 Mrutyunjaya Panda

Manas Ranjan Patra
Satchidananda Dehuri
Prafulla Kumar Behera
George A. Tsihrintzis
Sung-Bae Cho
Carlos A. Coello Coello
 Organization

Chief Patron
Sabita Acharya Utkal University, India

Patrons
Durga Sankar Pattnaik Utkal University, India
Avaya Kumar Nayak Utkal University, India
Shri Gautama Pradhan Utkal University, India

Convenor
Mrutyunjaya Panda Utkal University, India

Finance Committee
Biswojit Nayak Utkal University, India
Sanjaya Kumar Sarangi Utkal University, India

Publishing Committee
Prafulla Kumar Behera Utkal University, India
Manas Ranjan Patra Berhampur University, India
Satchidananda Dehuri Fakir Mohan University, India
Haraprasad Naik Utkal University, India
Sanjaya Kumar Sarangi Utkal University, India
N. Adhikari Utkal University, India

Organizing Committee
Prafulla Kumar Behera Utkal University, India
Mrutyunjaya Panda Utkal University, India
N. Adhikari Utkal University, India
Lalatendu Muduli India
Biswojit Nayak Utkal University, India
Haraprasad Naik Utkal University, India
Purna Chandra Sethi Rama Devi Women’s University, India
viii Organization

Omm Prakash Jena Ravenshaw University, India

Sanjaya Kumar Sarangi Utkal University, India

Advisory Committee
Ajith Abraham Machine Intelligence Research Labs, USA
Ganapati Panda IIT Bhubaneswar, India
P. K. Meher NTU, Singapore
S. K. Patra IIIT Vadodara, India
Milli Pant IIT Roorkee, India
S. K. Udgata University of Hyderabad, India
A. K. Bisoi Utkal University, India
J. Dandpat Utkal University, India
P. K. Hota Utkal University, India
N. Das Utkal University, India
S. K. Pradhan Utkal University, India
B. S. Panda IIT New Delhi, India
B. Majhi VSSUT, India
P. K. Mohapatra Utkal University, India
Prasanta K. Jana IIT(ISM) Dhanbad, India
Srikanta Patnaik SOA, India
B. K. Tripathy VIT, India

Program Chairs
Mrutyunjaya Panda Utkal University, India
Satchidananda Dehuri Fakir Mohan University, India
Prafulla Kumar Behera Utkal University, India
Manas Ranjan Patra Berhampur University, India
George A. Tsihrintzis University of Piraeus, Greece
Sung-Bae Cho Yonsei University, South Korea
Carlos A. Coello Coello Cinvestav-IPN, Mexico

Technical Program Committee

Aniket Mahanti University of Auckland, New Zealand
Nashwa El-Bendary Arab Academy of Science, Technology and
Maritime Transport, Egypt
Milli Pant IIT Roorkee, India
D. P. Mohapatra National Institute of Technology, Rourkela, India
Lalatendu Muduli Utkal University, India
B. Acharya National Institute of Technology, Raipur, India
 Organization ix

S. K. Panda National Institute of Technology, Warangal, India

S. K. Meher Indian Statistical Institute, Bangalore, India
Abhinav Tomar Netaji Subhas University of Technology, New
Delhi, India
Pratyay Kuila National Institute of Technology, Sikkim, India
Nabajyoti Mazumdar International Institute of Information Technology,
Allahabad, India
P. K. Biswal International Institute of Information Technology,
Bhubaneswar, India
Rahul Pramanik Birla Institute of Technology, Dubai, UAE
S. K. Mishra Birla Institute of Technology, Mesra, India
Md Azaruddin Aliah University, India
Partha Sarkar JIS College of Engineering, India
P. M. Khilar National Institute of Technology, Rourkela, India
S. K. Udgata University of Hyderabad, India
Nibedita Adhikari Utkal University, India
Bijayananda Patnaik National Institute of Technology, Raipur, India
Pulak Sahoo Silicon Institute of Technology, India
Sourav Bhoi Parala Maharaja Engineering College, India
Harish Kumar Sahoo VSSUT, India

Keynote Speakers
Ajith Abraham Machine Intelligence Research Labs, USA
Sung-Bae Cho Yonsei University, South Korea
Carlos A. Coello Coell Cinvestav-IPN, Mexico
Aniket Mohanty University of Auckland, New Zealand
Ganapati Panda IIT Bhubaneswar, India
Umamaheswar A. Kakinada Charter Communications, USA
Saroj Kumar Meher Indian Statistical Institute, Bangalore, India
Siba Ku. Udgata University of Hyderabad, India
Bhawani Sankar Panda IIT Delhi, India
Sadhna Rana Tata Consultancy Services, India
x Organization

Sponsors
Science and Engineering Research Board (SERB), Government of India

Odisha Higher Education Programme for Excellence and Equity (WB-OHEPEE), Utkal
University, India
Keynote Address
Industry 4.0 Meets Data Science: The Pathway for Society
5.0

Ajith Abraham

Machine Intelligence Research Labs (MIR Labs), Scientific Network

for Innovation and Research Excellence, USA

Abstract. We are blessed with the sophisticated technological artifacts

that are enriching our daily lives and the society. Industry 4.0 is the current
trend of automation and data exchange in manufacturing technologies,
which also includes a close integration of cyber-physical systems, the
Internet of things and cloud computing. In this talk, the concept of Indus-
try 4.0 and Society 5.0 will be presented and then various research chal-
lenges from several application perspectives will be illustrated. Some real
world applications involving the analysis of complex data/applications
would be the key focus.
Continual Learning for Intelligent Systems in Changing
Environments

Sung-Bae Cho

Yonsei University, South Korea

Abstract. Intelligent systems are expected to learn new knowledge incre-

mentally without forgetting in an ever-changing environment. Due to the
critical limitation accompanied by amazing success of deep learning,
a lot of continual learning methods have been proposed with different
approaches. In this talk, I will give the general definition and three key
approaches of continual learning, and present an idea of exploiting extra
memory for working out the catastrophic forgetting of the knowledge
learned with the previous tasks. Experimental results with several bench-
mark datasets such as MNIST, CIFAR-10 and CIFAR-100 will show
the possibilities and challenges of the continual learning in changing
environments.
Where is the Research on Evolutionary Multi-objective
Optimization Heading to?

Carlos A. Coello Coello

Investigador Cinvestav 3F, CINVESTAV-IPN, Mexico

Abstract. The first multi-objective evolutionary algorithm was published

in 1985. However, it was not until the late 1990s that so-called evolution-
ary multi-objective optimization began to gain popularity as a research
area.
Throughout these 37 years, there have been several important
advances in the area, including the development of different families of
algorithms, test problems, performance indicators, hybrid methods and
real-world applications, among many others. In the first part of this talk
we will take a quick look at some of these developments, focusing mainly
on some of the most important recent achievements.
In the second part of the talk, a critical analysis will be made of the
by analogy research that has proliferated in recent years in specialized
journals and conferences (perhaps as a side effect of the abundance of
publications in this area).
Much of this research has a very low level of innovation and almost
no scientific input, but is backed by a large number of statistical tables
and analyses.
In the third and final part of the talk, some of the future research chal-
lenges for this area, which, after 37 years of existence, is just beginning
to mature, will be briefly mentioned.
Designing a Software Framework Based on an Object
Detection Model and a Fuzzy Logic System for Weed
Detection and Pasture Assessment

Aniket Mohanty

University of Auckland, New Zealand

Abstract. Artificial Intelligence (AI) has provided advanced and pre-

cise tools for processing agricultural data called agreprecision. Pasture
processing is one of the main applications of agriprecision to provide
automation and saving labour tasks for dairy farmers. By weeds as the
main issue of pastures, dairy farmers need to invest a tremendous amount
of time for pasture monitoring for weed destroying. This paper proposes
designing a software framework based on an object detection model and
a fuzzy logic system for weed detection and pasture assessment. Weed
density and empty spots are two main factors of reducing pasture pro-
ductivity. By considering these two factors as pastoral input variables,
we ingest them to our system to process the pastures and score the state
of pasture’s productivity. With the aid of our software system, we can
produce 2D weed density maps, 2D bareness maps, and scoring maps,
which provides a much better insight into the pastures. The types of 2D
maps and the yield score can help and support dairy farmers to schedule,
organize, and manage pastoral weeds.
An Overview of Machine Learning Based Intelligent
Computing and Applications

Ganapati Panda1,2

1
IIT, Bhubaneswar
2 C.V. Raman Global University, Bhubaneswar, India

Abstract. In recent years, Machine Learning Techniques have played a

significant role for decision making, reasoning and finding appropriate
solution by analyzing the available data from experiments, communica-
tion receivers, finance, agriculture, health care and manufacturing. Pre-
processing of the data comprising of de-noising and normalization plays a
vital role. Subsequently, features are extracted from the raw data and then
selection/reduction of features is carried out. The selected features are
then used as inputs to the intelligent decision making process for classi-
fication, prediction, pattern recognition, inverse modeling, detection and
optimization, operation. The proposed talk will cover all these aspects in
details. In addition, various important soft computing, evolutionary com-
puting and machine learning techniques (supervised and unsupervised)
will be dealt in brief. The different application areas pertaining to each
intelligent operation will also be illustrated.
Semisupervised Learning with Spatial Information
and Granular Neural Networks

S. K. Meher

Indian Statistical Institute, Bangalore, India

Abstract. The effectiveness of a classification model depends primarily

on two things. One is the accurately labeled samples, and the other is the
number that should be significant. Without these, the model’s estimation
of posterior probability and class conditional probability density func-
tions becomes erroneous. However, getting a large number of accurately
labeled samples for real-time applications, e.g., remote sensing, health-
care, and telecommunication, is cumbersome. To address these issues,
semisupervised learning strategies are adopted that can exploit unlabeled
and labeled samples in the learning process and lead to the performance
improvement of a classification model. Here, we discuss a semisuper-
vised classification model with spatial information-based self-learning
methodology for pattern classification. The model uses a granular neural
network (GNN) as the base classifier because of its customizable network
architecture that is functionally interpretable and costs less computational
complexity. The architecture of GNN is governed by fuzzy if-then rules
generated from fuzzy granulation of input feature space. The model has
used an improved spatial neighborhood learning method to understand
data distribution in a semisupervised framework better. The technique
collects the information with collaborative opinions of two independent
information extraction approaches, i.e., based on mutual neighborhood
criteria and class maps of unlabeled samples. A case study on remote
sensing images is provided at the end.
IoT Based General Purpose Sensing Application
for Smart Home Environment

Siba K. Udgata

WiSECom Lab, School of Computer and Information Sciences,

University of Hyderabad (Institute of Eminence), Hyderabad, India

Abstract. General Purpose Sensing is an approach towards developing

a universal sensor used to detect different facets of an environment. In
this approach a single enhanced sensor is used which can monitor a vast
context without requiring direct object instrumentation. The most natural
and traditional form of sensing relies on direct and distributed sensing
wherein one single sensor is utilized to monitor a single aspect of an
environment. General purpose sensing helps in developing and creating
smart sensing environments. This also overcomes many of the issues
faced by earlier methods and helps a normal environment to become
smart environments. In this, we discuss our experience on developing a
general purpose sensing application to detect combination of different
day-to-day events and combination of events in a room environment. We
also discuss how general purpose sensing can be used to estimate the
energy consumption in a room. For this, we use different sensor modules
like MPU6050, HMC5833L, BH1750, INMP441, HC-SR501, DHT11
together with ESP8266 controller to generate a dataset for linear acceler-
ation, angular acceleration, magnetic values along the 3 axis coordinates
(X, Y, Z), light intensity, audio frequencies, movement detection, tem-
perature and humidity values while performing combination of different
general activities like Hand shake and Face touch, Raise hand and Face
touch, Raise hand and hand shake, Face touch and hand shake, Face
touch and Raise hand, Shake hand and Raise hand among a few others.
The developed module is also used to generate a data set for combination
of events in a room environment like fan on, lights on, AC on, person
walking, computer on and combination of these activities. We also esti-
mated the energy consumption in a room environment using the general
purpose sensor module which records sensor values while performing
certain activities which are then converted and analyzed by a model to
classify the activities. This is accomplished by using adaptive threshold
based data segmentation and different classification methods.
 Emerging Topics in Wireless and Network
Communications – A Standards Perspective

Umamaheswar A. Kakinada

Emerging Technologies - Research and Development,

Charter Communications, Colorado, USA

Abstract. We probably are in the initial stages of many promising evolv-

ing technologies such as AR/VR/XR, Metaverse, AI/ML based various
network functions such as network automation, quality of experience, end
user privacy and security and federated learning is just to name a few.
Many of these technologies are in the early stage of development and
many of the essential enablers for these technologies such as low latency,
high throughput, ultra-reliable fail-safe communications, and various sen-
sory devices are still in the various stages of development and realization.
The standards, protocols, and reference architectures developed by vari-
ous Standards Development Organizations (SDOs) play an important role
in development of these technologies, also enabling harmonious coexis-
tence of different components of this somewhat disparate ecosystem; to
realize these ambitious goals. In this presentation we will explore a few
of these being developed by 3GPP and other SDOs.
 Emerging Topics in Wireless and Network
Communications – A Standards Perspective

B. S. Panda

Indian Institute of Technology Delhi, Hauz Khas, New Delhi, India

Abstract. With the boom in technologies and mobile networks in recent

years, online social networks have become an integral part of our daily
lives. Nowadays, various Online Social Networks (OSNs) like Twitter,
Facebook, Instagram, etc., have enabled the world into a hyper-connected
global world and provide virtual platforms on which various social inter-
actions occur between the users. These OSNs offer users an ideal platform
to share and promote new ideas, products, or information. An online
social network can be modeled as a graph, where V denotes a set of
people or entities present in the network, and E represents edges. The
edge between two entities on a social network corresponds to a virtual
acquaintance like friendship and follow-followee relationships. Due to
the presence of many users from different parts of the world on these
social networks, they help facilitate the dissemination of information at a
faster rate. Virtual networks connect people worldwide and provide them
excellent platforms for promoting their products and ideas. It is often
the case that certain users are more influential than others present on
social networks. The process of efficiently recognizing influential users
to maximize a particular piece of information across a network is known
as Influence Maximization (IM). Influential nodes are the nodes with high
spreading capability, and a piece of information originating from them
can reach a maximum extent in the network by triggering an information
diffusion cascade. Identifying influential nodes for the task of influence
maximization is one of the popular research topics in the field of network
science. Viral marketing is one of the popular applications of influence
maximization where companies intend to reach a large number of users
from initially chosen influential seed nodes through diffusion cascade. In
this talk, we will discuss some recent solutions to influence maximization
in complex networks.
 Contents

Intelligent Computing

Ensemble Learning Model for EEG Based Emotion Classification . . . . . . . . . . . . 3

Sanjit Kumar Dash, Sambit Subhasish Sahu, J. Chandrakant Badajena,
Sweta Dash, and Chinmayee Rout

Foundation for the Future of Higher Education or ‘Misplaced Optimism’?

Being Human in the Age of Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Ashraf Alam and Atasi Mohanty

AI Enabled Internet of Medical Things Framework for Smart Healthcare . . . . . . 30

Jyoti Srivastava and Sidheswar Routray

Metaverse and Posthuman Animated Avatars for Teaching-Learning

Process: Interperception in Virtual Universe for Educational
Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Ashraf Alam and Atasi Mohanty

Tuning Functional Link Artificial Neural Network for Software

Development Effort Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Tirimula Rao Benala and Satchidananda Dehuri

METBAG – A Web Based Business Application . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Nannapaneni Akshaj and B. K. Tripathy

Designing Smart Voice Command Interface for Geographic Information

System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Shivam Pant and Narayan Panigrahi

Smart Garbage Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

Aviral Jain, Vidipt Khetriwal, Hitesh Daga, and B. K. Tripathy

Optical Sensor Based on MicroSphere Coated with Agarose for Heavy

Metal Ion Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Asesh Kumar Tripathy and Sukanta Kumar Tripathy

Influential Factor Finding for Engineering Student Motivation . . . . . . . . . . . . . . . 137

Sourajit Ghosh, Aritra Sinha, Arpita Roy, and Biswarup Neogi
xxiv Contents

Prediction of Software Reliability Using Particle Swarm Optimization . . . . . . . . 148

Getachew Mekuria Habtemariam, Sudhir Kumar Mohapatra,
and Hussien Worku Seid

An Effective Optimization of EMG Based Artificial Prosthetic Limbs . . . . . . . . . 157

Dwarika Nath Choudhury and Narayan Nayak

Communications

Performance Analysis of Fading Channels in a Wireless Communication . . . . . . 175

Pradyumna Kumar Mohapatra, Saroja Kumar Rout,
Ravi Narayan Panda, Anudeep Meda, and Banoj Kumar Panda

Power Conscious Clustering Algorithm Using Fuzzy Logic in Wireless

Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Sanjaya Kumar Sarangi, Arabinda Nanda, Manas Ranjan Chowdhury,
and Subhadra Mishra

Cryptanalysis on “An Improved RFID-based Authentication Protocol

for Rail Transit” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
Suresh Devanapalli and Kolloju Phaneendra

A Novel Approach to Detect Rank Attack in IoT Ecosystem . . . . . . . . . . . . . . . . . 204

Amardeep Das, Nibedita Adhikari, Pradeep Kumar Bhale, and S. K. Rath

Energy Efficient Adaptive Mobile Wireless Sensor Network in Smart

Monitoring Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Seli Mohapatra and Prafulla Kumar Behera

Orthogonal Chirp Division Multiplexing: An Emerging Multi Carrier

Modulation Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
Mohit Kumar Singh and Ashish Goel

Machine Learning and Data Analytics

COVID-19 Outbreak Estimation Approach Using Hybrid Time Series

Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Soham Chakraborty, Sushruta Mishra, and Hrudaya Kumar Tripathy

Analysis of Depression, Anxiety, and Stress Chaos Among Children

and Adolescents Using Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . 261
Satyananda Swain and Manas Ranjan Patra
 Contents xxv

Sentiment Analysis from Student Feedbacks Using Supervised Machine

Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Preeti Routray, Chinmaya Kumar Swain,
and Rakesh Chandra Balabantaray

Machine Learning Algorithms for Diabetes Prediction . . . . . . . . . . . . . . . . . . . . . . 289

Sudipta Priyadarshinee and Madhumita Panda

CRODNM: Chemical Reaction Optimization of Dendritic Neuron Models

for Forecasting Net Asset Values of Mutual Funds . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Sarat Chandra Nayak, Satchidananda Dehuri, and Sung-Bae Cho

Machine Learning Approaches and Particle Swarm Optimization Based

Clustering for the Human Monkeypox Viruses: A Study . . . . . . . . . . . . . . . . . . . . . 313
Akshaya Kumar Mandal, Pankaj Kumar Deva Sarma,
and Satchidananda Dehuri

Machine Learning-Based Hybrid Feature Selection for Improvised

Network Intrusion Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Pritimayee Satapathy and Prafulla Kumar Behera

Application of Efficient Feature Selection and Machine Learning

Algorithms in Mental Health Disorder Identification . . . . . . . . . . . . . . . . . . . . . . . . 348
Sumitra Mallick and Mrutyunjaya Panda

A Survey on Sentimental Analysis of Student Reviews Using Natural

Language Processing (NLP) and Text Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
J. Jayasudha and M. Thilagu

Rice Yield Estimation Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379

Niyati Mishra, Sushruta Mishra, and Hrudaya Kumar Tripathy

Sentiment Analysis on Movie Review Data Using Ensemble Machine

Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Om Prakash Jena, Alok Ranjan Tripathy, Manas Prasad Rout,
and Partha Sarathi Pattnayak

Comparative Analysis of COVID 19 Detection from Cough Speech Using

Machine Learning Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Soumya Mishra, Tusar Kanti Dash, Ganapati Panda, Amit Kumar,
and Sushant Kumar Singh
xxvi Contents

Classification of High-Resolution Satellite Image with Content Based

Image Retrieval and Local Binary Pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Rajalaxmi Padhy, Laxmipriya Samantaray, Sanjit Kumar Dash,
and Jibitesh Mishra

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417

Intelligent Computing
 Ensemble Learning Model for EEG Based
Emotion Classification

Sanjit Kumar Dash1(B) , Sambit Subhasish Sahu1 , J. Chandrakant Badajena1 ,

Sweta Dash1 , and Chinmayee Rout2
1 Odisha University of Technology and Research, Bhubaneswar, Odisha, India
[email protected]
2 Ajay Binay Institute of Technology, Cuttack, Odisha, India

Abstract. Emotion and feelings are recently becoming popular concepts in the
everyday life. It not only affects human health but also plays an essential role in
the decision-making processes. For this reason, emotion classification is one of the
important aspects to deal with the problems like mental disorders, suicidal activ-
ities and judgmental process. Electroencephalogram (EEG) signal is one of the
physiological signals which can be collected from the human brain activity while
a person performing various mental and physical task. In this paper, the DEAP
dataset has been implemented with the deep learning model for the classification
process. In the process of developing models, for the extraction of the impor-
tant features from the unprocessed EEG signals, Fast Fourier Transformation is
used. Three ensemble deep learning models are tested and compared to get the
best accuracy result for emotion classification. Furthermore, by the best model we
can classify four emotional regions in the valance-arousal plane: HVHA, HVLA,
LVHA and LVLA can be classified. The experimental results show that among
all the three deep learning models, 1D-CNN-GRU achieved the highest training
accuracy of 96.54% as compared to LSTM and 1D-CNN, which is the best model
to classify emotions in this context.

Keywords: EEG · Emotion classification · Feature extraction · LSTM · GRU ·

Fast fourier transformation

1 Introduction
Now a days several industries have adopted the implementation of emotion recognition
and classification on their employees. When the human resources fields start gain the
emotional state of the employee, they could take the advantages from it to improve the
quality of taking decision regarding their organization and it will also help to make judg-
ment regarding employees’ work. Emotion is physiological process triggered due to the
environmental and personal conditions, which affects the mood, behavior, character and
motivations. It also plays an important role in the daily conversation and behavior which
affects the decision-making process, personal and professional workflow. Nowadays,
the emotional state of a person takes a very crucial position in life, which is associated
with their mental and physical health, daily lifestyle, decisions about the future and

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 3–16, 2022.
https://doi.org/10.1007/978-3-031-23233-6_1
4 S. K. Dash et al.

many more. The Valence, Arousal, and Dominance (VAD) dimensions are another way
to divide up emotion. A measure of pleasantness known as valence can originate from
very good (pleasure) or very negative emotions (displeasure). Arousal, on the other hand,
is the level of feeling that a situation elicits and can range from positive excitement to
calmness (negative). Last but not least, dominance is the degree of control demonstrated
in response to a stimulus [1]. The popular Circumplex Model of Affect is depicted in
Fig. 1 and categories emotions into two groups: Low valence or negative valence emo-
tions (such as fear, tension, rage, sadness, boredom, etc.) and High valence or pleasant
valence emotions (happy, calm, delighted, excited, etc.).

Fig. 1. Graphical representation of circumplex model

In the present scenario, it has become increasingly significant to analyze a person’s

emotional state which can be apprehended from social networking websites, blogging
platforms like twitter, online review sites and so on. However, in order to facilitate
emotion classification, the use of EEG data is an emerging field of study. The EEG
is (Electroencephalogram) which is used on the scalp surface to record electric current
using metallic electrodes and conductive media [2]. By using the electrodes, the different
electric waves forming according to different mental states inside the brain could be
collected as EEG signals. This can help to have a better understanding of a person’s
mental state.
Recently, various researchers are using different data sets like AMIGOS,
DREAMER, Seed according to their field of experiment, but for emotion recognition
DEAP dataset is highly recommended. The EEG dataset is not used for emotion clas-
sification but also different fields of BCI like wheel chair, Telemedicine, stock market
prediction etc. In different studies different researchers are using different processing
and feature extraction methods and different machine learning models. In our paper we
are using Fast Fourier Transformation as the feature extraction model before feed into
the Deep learning model. CNN and RNN are used for the classification purpose. We also
propose 3 models, likely 1D-CNN, LSTM and a combination of two neural networks
CNN and RNN as 1D-CNN-GRU.1D-CNN is chosen from CNN and LSTM and GRU
 Ensemble Learning Model for EEG Based Emotion Classification 5

are chosen from RNN. All the models are tested in the training and testing process to
check the accuracy and loss value and will be compared.
The remainder of the paper is structured as follows. Section 2 discusses the related
works which helped us to get the fundamental knowledge about the research work.
Section 3 describes System Model and Methodology. Section 4 describes the Dataset
description used in the paper. In Sect. 5, the complete environmental setup and results
were elaborated, including the model accuracy and data loss. Section 6 concludes the
paper with detailed future work and limitations.

2 Related Works

The C-RNN deep learning hybrid model, which combines CNN & RNN, was introduced
by Xiang Li et al. (2016) for the aim of recognizing emotions [3]. The data is pre-
processed using continuous wavelet transformation & body production before being
taught inside the model. For the arousal and valence dimensions, this test’s overall
performance is 74.12% and 72.06%, respectively. In their deep learning architecture,
they used DEAP dataset [4] for emotion recognition, Alhagry et al. (2017) recommended
LSTM. According to this method, the average accuracy for the arousal, valence, and
liking classes is 85.65%, 85.45%, and 87.99%, respectively.
Lin et al. performed emotional state classification in CNN using end -to-end learning
method and DEAP Dataset [5]. The datasets were turned into six grayscale images that
included frequency and time information, and the previously identified characteristics
were then taught using the AlexNet version. For arousal and valence, this study gives
accuracy of 87.30% and 85.50%, respectively. Li et al. (2017) used the DEAP dataset to
perform their emotion recognition task using a hybrid CNN and LSTM RNN (CLRNN)
[6]. Already converted into a series of Multidimensional Feature Images is the dataset.
With the hybrid neural networks suggested inside the study, every event may be accu-
rately classified as having a common emotion 75.21% of the time. In order to accomplish
emotion classification tasks, Acharya et al. (2020) examined the general performance
of CNN and LSTM models and performed feature extraction using FFT [7]. The test
outcome was excellent for LSTM and CNN model with an accuracy of 88.6% and 87.2%
respectively for the liking emotion.
Zhang et al. (2020) discussed different deep learning models like CNN, DNN, LSTM,
and combination of CNN and LSTM model along with their applications to the research
field of EEG-based emotion classification [8]. This study made substantial use of the
DEAP dataset, and many capabilities, including mean, maximum value, standard devi-
ation, minimum value, skewness, and kurtosis, were retrieved from it. The CNN model
with 90.12% accuracy and CNN-LSTM model with 94.17% accuracy shows excellent
ability to complete this task. Anubhav et al. (2020) investigated the EEG signals with
the goal of creating a headgear version for tracking real-time emotions [9]. From the
DEAP dataset band energy and frequency domain were recovered, and the accuracies
for valence and arousal dimensions were calculated with accuracy of 94.69% 93.13%
respectively using LSTM.
A 2D-CNN structure was suggested by Dar et al. (2020) to systematic EEG indica-
tions for emotion recognition [10]. The DREAMER & AMIGOS datasets are utilized
6 S. K. Dash et al.

for this test and before being input into CNN, each statistic is converted into 2D function
matrix (PNG format). The multi-modal emotion reputation system also uses different
peripheral physiological markers, such as ECG and GSR, in addition to EEG. Only
76.65% accuracy can be attained using the EEG modality, and multi-modal fusion is
required to achieve the overall maximum accuracy of 90.8% and 99.0% for DREAMER
and AMIGOS dataset respectively.

3 System Model and Methodology

Figure 2 shows the system model for emotion classification. There are 4 phases in
the process, initially collected DEAP dataset with proper approval which is publicly
available. Then it will be passed through the Feature Extraction process in the second
phase, which extract the main features from the unprocessed EEG data. Thirdly the
data will split into Train and Test data, training data will be passed through the three
different models; 1D-CNN, LSTM,1D-CNN-GRU to check testing accuracy and all other
parameters. In the end phase the best model will be implemented as the classification
model for the classification of four emotional regions from valance arousal plane.

Fig. 2. System model

3.1 Feature Extraction

For feature extraction, Fast Fourier Transformation (FFT) was used which is best in all
tradition feature extraction models to extract the important features for the classification
purpose, which reduces the number of computation data needed for an experiment from
actual data of size N and the final result was achieved (58560,70) dimension from
(40,40,8064) which allows for faster training and higher accuracy.
The extracted features contain five frequency bands: Delta, Theta, Alpha, Beta and
Gamma with frequency 1–4, 4–8, 8–14, 14–31, and 31–50 Hz respectively. The PyEEG
python package was used to extract these five out of 70 characteristics in total. The
signal domain is transformed from time to frequency on the x-axis using FFT. This is
based on the idea of discrete fourier transform (DFT) on time series data. Calculating
the DFT coefficients in an iterative approach will reduce both computational time and
complexity. It also helps to reduce round-off errors in computations.
 Ensemble Learning Model for EEG Based Emotion Classification 7

3.2 Deep Learning Model Implementation

In this section, three deep learning models have been implemented: 1D-CNN (One
Dimensional Convolutional Neural Network), LSTM (Long Short-Term Memory), and
a hybrid model of 1D-CNN and GRU (Gated Recurrent Unit). Preprocessed version of
the DEAP dataset used in the models. The models were trained to categorise each emotion
individually using different train-test splits, such as arousal, valence, dominance, and
liking. The following is a description of model implementations using Keras (Chollet
(2015) [11]):

A. 1D-CNN Model
1D-CNN is used to extract the significant features from the DEAP dataset. CNN
works good on the time series data which are the 1D signals. In Conv1D the kernel
slides along one dimension only. This is one of the key justifications for using
1D-CNN in our research. As shown in Fig. 3, out of the 10 classes, we employed
three conv1D, three dense layers that were completely coupled, and one dense
layer with SoftMax activation. The first convolution layer employs the Rectified
linear unit (ReLU) as its activation function and has 164 filters and a kernel of
size 3. After hyperparameter tuning, optimization with Grid Search and manual
adjustments, the number and size of filters are determined. A shape of (70,1) with
the same padding and stride of one is feed into Conv1D’s first layer. To reduce
network overfitting, dropout on the dense layer outputs is implemented with 0.2
dropout probability. Following this dense layer of 21 ReLU-activated neurons with
a 0.2 dropout probability is a layer of 42 Tanh-activated neurons with a 0.2 dropout
probability. At the end Dense layer of 10 neurons with a SoftMax activation function
produces the network’s final output.

Fig. 3. 1D - CNN architecture

B. LSTM Model

Long Short-Term Memory Networks (LSTMs) is one type of recurrent neural net-
work (RNN), first introduced in 1997 by Hochreiter and Schmidhuber [12]. It solves
the problem of short-term memory. It has a built-in gate that can recognize which
information in a sequence has to be kept and which information can be discarded.
8 S. K. Dash et al.

Fig. 4. LSTM architecture

Figure 4 shows the model architecture of LSTM, which consists of two dense lay-
ers, four LSTM layers, and one bi-directional layer. The initial bi-directional LSTM
layer contains 164 units. It involves adding a second LSTM layer on top of the first
one in the network. The first receives the input sequence, and the second receives
a reverse copy to the next layer. The dropout layer, which has a probability of 0.6,
comes next. Inputs are randomly set to 0 which helps to prevent overfitting. The
following layer is a 256-neuron LSTM layer, followed by a 0.6 dropout layer. Two
LSTM layers with 82 neurons each make up the next four layers, which are then fol-
lowed by a dropout layer. There were 0.6 and 0.4 percent dropout rates, respectively.
42 neurons make up the final LSTM layer with a dropout layer of 0.4 following that
21 units of dense layers is applied. ReLU is the activation method employed in this
case. The SoftMax activation function is then applied to dense layer of 10 classes,
which gives a multiclass probability distribution. Using argmax find out the class
output is done after knowing the probability of all the classes.

C. 1D-CNN-GRU Model
This variant is a hybrid model of the 1D-CNN and GRU of deep learning architec-
tures. The network’s first input size is 256 units, followed by two seconds of the
time-stamped signal with 256 data points, 128 units of convolutional filters, and
a kernel size of 3. Rectified linear unit (ReLU) activation function is used in the
first convolution layer. The precise number and size of filters are identified after
extensive hyperparameter optimization using Grid Search and manual adjustments.
Conv1D’s first layer receives as its input a shape of the form (70,1) with the same
padding. The outputs of the first layer are normalized using a batch normalized
layer, which has a mean value zero and a standard deviation value of one. The input
is down sampled in the following layer, dropout of 0.2 after a Max pooling 1D layer
with a pool size of 2 and the second convolutional layer, which is same as the first
one. The implementation of GRU comes after the convolutional layers with an input
length of 256 units and 32 units. At the end of every GRU layers a dropout layer of
0.2 has been set. A flattening operation is implemented to send the features into the
 Ensemble Learning Model for EEG Based Emotion Classification 9

1D feature vector prior to dense layer. The dense layer is set to 32 units and ReLU is
used as the activation function. To represent the 4 labels of classification 4 units of
dense layer has been set. The activation function for the dense layer is SoftMax. In
this model 379,594 units of trainable parameter used. The details model architecture
of 1D-CNN-GRU is given in Fig. 5.

Fig. 5. 1D - CNN?+?GRU architecture

4 Dataset Description
The DEAP dataset [13] used in the experiment is publicly available for researchers for
their experiments. This data set contains both EEG and EMG signals. To collect the data
32 participants as engaged. In the dataset the physiological recordings and participant
evaluations of 32 individuals (s01-s32) are covered. These physiological clips are already
in BioSemi.bdf format and have not been processed. For each of them, 40 films were
presented which makes 40 channels in total. Depending on the rating, the emotion is
either stronger or weaker; the stronger the emotion, the higher the rating. Table 1 contains
each subject’s information, which contains two arrays: data and label, which contains
the array shape and the content of each file.

Table 1. Pre-processed dataset description

Array name Array shape Array contents

Data 40 ‫נ‬40 ‫נ‬8064 video/trial * Channel * Data
Labels 40 ‫נ‬4 video/trial * label (valence, arousal, dominance, liking)

The data was filtered by a band-pass filter with a bandwidth of 4–45 Hz and down
sampled by 128 Hz. In addition to these, the collection includes listings and links to
YouTube music videos. Questions to ask before testing are contained in the participant
questionnaire file, each trial or video was given a rating on a scale of 1 to 9. The dataset is
10 S. K. Dash et al.

divided into four classes, which is subsequently labelled as: High-Valence Low-Arousal
(HVLA), Low-Valence High-Arousal (LVHA), and Low-Valence Low-Arousal (LVLA)
are four different types of arousal The Table 2 contains all the four classes and two labels,
the threshold value for this classification is 5. If the value is greater than 5, then it will
be classified as high and low if it is less than 5.

Table 2. Label classification

Label HVHA HVLA LVHA LVLA

Valence >5 >5 <=5 <=5
Arousal >5 ?<=5 >5 <=5

In our paper we have selected 14 channels AF3, AF4, F3, F4, F7, F8, FC5, FC6,
T7, T8, P7, P8, O1, O2 and 5 bands 4, 8, 12, 16, 25, 45 to reduce the computational
cost as well as for the better result for emotion classification [14]. This entire channel
selection process is based on the significance of the brain regions that make up emotional
states. When we divided the label into the four categories of HVHA, HVLA, LVHA, and
LVLA, the paper was considered to be complete. But for now, as a result, models will
be compared on the basis of an increase in training accuracy while decreasing validation
loss for the selection of a better classification model.

5 Experimental Setup and Results

Google Collaboratory was used to compute 1D – CNN, LSTM, GRU classifiers because,
it uses Jupyter notebook service that needs no installation and gives unrestricted access
to computing tools, such as GPUs. Python is the python version (3.7.13). The current
version of TensorFlow is 2.8.0. Pandas: 1.3.5, numpy: 1.21.6, sklearn: 1.0.2, plotly: 5.5.0.
A laptop is used to run the code. Nvidia K80s, T4s, P4s, and P100s are common GPUs
seen in CoLab. To get an amazing result, hyperparameter should be tuned adequately.
Conv1D layers are utilized in the CNN architecture because they are best suited for
data in time series. Both maximum and average pooling were used, however maximum
pooling produced better results, as predicted by the literature. For the CNN, LSTM,
and 1D-CNN-GRU architectures, the finalized epoch size is 200 and batch size of 256.
The models are trained on 80–20 train–test splits and 10-fold cross validation is also
employed to determine the best metrics-accuracy. To update the weights during back-
propagation, used Adam as the optimizer and category cross entropy as the loss function.
In both cases, SoftMax performs the activation of the last layer. Separate decisions were
made for each of the three models, including the number of layers, hidden layers, filter
size, number of filters, pool size for the CNN model, and hidden neurons, dropout rates,
and layers for the LSTM and GRU models separately. Both grid search and manual
testing are used to finalize everything.
 Ensemble Learning Model for EEG Based Emotion Classification 11

Overall performance of the three models can be elaborated and analyzed on the
basic of accuracy and loss value. Table 3 contains test accuracy and test loss of each
model. It can be inferred that, the 1D-CNN-GRU model architecture provide the best
test accuracy of 96.54%, 41.3% of test loss with an 80–20 train test split as compared
to the LSTM model architecture’s accuracy of 89.6%, test loss of 39.6% with an 80–20
train test split and the 1D-CNN model architecture’s accuracy of 90.65%, test loss of
42.2% with an 80–20 train test split. All these three experiments were designed to test
the overall performance of each model for classification of emotion.

Table 3. Test accuracy and test loss comparison

Classifier Test accuracy Test loss

1D - CNN 90.65% 42.2%
LSTM 89.6% 39.6%
1D CNN?+?GRU 96.54% 11.4%

Above discussed models performed exceptionally well at generalizing findings since

they classified each emotion with an accuracy rate of more than 80%. The loss function is
categorical cross entropy. Overfitting on training data was avoided by the use of dropout
layers, which also enhanced model results. Batch normalization layers also significantly
affect model accuracy. By altering the number of epochs, data units and input layers
the accuracy affects. When we implemented the GRU architecture we got a very less
accuracy, but after implementing the hybrid model of 1D-CNN and GRU, it crossed all
the levels.

Fig. 6. (a) Model accuracy and (b) Model loss of LSTM

12 S. K. Dash et al.

Fig. 7. (a) Model accuracy and (b) Model loss of 1D-CNN

Fig. 8. (a) Model accuracy and (b) Model loss of 1D-CNN-GRU

Figure 6, Fig. 7, Fig. 8 shows model accuracy and model loss of LSTM, 1D-CNN,
1D-CNN-LSTM respectively. In model accuracy curves the train and test curve goes
up award direction with increase of epoch, where as in model loss figure it goes down
with increase of epoch. Curves were found to have a minor variation, The 1D-CNN
model starts learning earlier than the LSTM model and where as it took the CNN model
about 50 epochs to reach a stable point, the LSTM model took about 130 epochs. No
overfitting took place, and the training process came to an end after 165 epochs. Each
training period takes between 16 and 330 ms. Figure 9 shows test model of LSTM,
1D-CNN and 1D-CNN-GRU, which makes a curve between test loss and test accuracy,
when the accuracy increases the value loss decreases by the increase of epoch.
 Ensemble Learning Model for EEG Based Emotion Classification 13

Fig. 9. Test model of (a) LSTM, (b) 1D-CNN, (c) 1D-CNN-GRU

Dropout and batch normalisation layers have a big impact on the model’s accuracy.
Additionally, we constructed a confusion matrix in Fig. 10 to explore the discrepancy
between the predicted and actual value. Each model’s F1-Score and recall value are close
to one, which shows that this model has good quality.
14 S. K. Dash et al.

Fig. 10. Confusion matrix of LSTM, 1D-CNN, 1D-CNN-GRU

 Ensemble Learning Model for EEG Based Emotion Classification 15

Fig. 10. (continued)

6 Conclusion

In this paper we have discussed the simplest feature extraction method with the best
classification model for emotion classification using DEAP dataset. In this paper, we
have described three deep learning models: 1D-CNN, LSTM, and 1D-CNN-GRU. As
compared to normal feature extraction strategies, FFT increased accuracy and extracted
crucial characteristics. The 1D-CNN design has proven to be best model to extract EEG
signal features. The 1D-CNN architecture has a classification accuracy of 90.8% which
is somewhat better than the 89.2% classification accuracy of the LSTM model. From all
the three models the 1D – CNN – GRU gives the best accuracy of 96.54%.
Bi-LSTMs were able to preserve data from the past as well as the future, which
contributed to increase the accuracy in the LSTM model. Even though the emotion
classification test showed that our model worked remarkably well, we still wish to assess
it further in additional datasets like DREAMER and AMIGOS and enhance our model.
More particular, while this version no longer undergoes testing on various people, it was
best trained using the DEAP dataset only.
In future we would try to work on quick output of the EEG data processing in real-
time online analysis systems, which limits calculation time. We will concentrate on a
multi-task cascaded-hybrid LSTM and CNN model in the future, which will combine
their features and improve the efficacy of the emotion classification model. The system
to detection of emotion can improve human experiences by minimizing the gap between
16 S. K. Dash et al.

computational technology and human emotions and allowing computers, BCI system and
robots to receive emotional feedback in real-time. Even with the help of this technology,
therapists can more thoroughly evaluate their patients and learn how to spot depression
early and prevent it before any outward situation occurs.

References
1. Alarcao, S.M., Fonseca, M.J.: Emotions recognition using EEG signals: a survey. IEEE Trans.
Affect. Comput. 10(3), 374–393 (2017)
2. Teplan, M.: Fundamentals of EEG measurement. Measur. Sci. Rev. 2(2), 1–11 (2002)
3. Li, X., Song, D., Zhang, P., Yu, G., Hou, Y., Hu, B.: Emotion recognition from multi-channel
EEG data through convolutional recurrent neural network. In: 2016 IEEE International
Conference on Bioinformatics and Biomedicine (BIBM), pp. 352–359. IEEE (2016)
4. Alhagry, S., Fahmy, A.A., El-Khoribi, R.A.: Emotion recognition based on EEG using LSTM
recurrent neural network. Int. J. Adv. Comput. Sci. Appl. 8(10), 355–358 (2017)
5. Lin, W., Li, C., Sun, S.: Deep convolutional neural network for emotion recognition using
EEG and peripheral physiological signal. In: Zhao, Y., Kong, X., Taubman, D. (eds.) ICIG
2017. LNCS, vol. 10667, pp. 385–394. Springer, Cham (2017). https://doi.org/10.1007/978-
3-319-71589-6_33
6. Li, Y., Huang, J., Zhou, H., Zhong, N.: Human emotion recognition with electroencephalo-
graphic multidimensional features by hybrid deep neural networks. Appl. Sci. 7(10), 1060
(2017)
7. Acharya, D., et al.: Multi-class emotion classification using EEG signals. In: Garg, D., Wong,
K., Sarangapani, J., Gupta, S.K. (eds.) IACC 2020. CCIS, vol. 1367, pp. 474–491. Springer,
Singapore (2021). https://doi.org/10.1007/978-981-16-0401-0_38
8. Zhang, Y., et al.: An investigation of deep learning models for EEG-based emotion recognition.
Front. Neurosci. 14, 622759 (2020)
9. Nath, D., Singh, M., Sethia, D., Kalra, D., Indu, S.: An efficient approach to EEG-based
emotion recognition using lstm network. In: 2020 16th IEEE International Colloquium on
Signal Processing and its Applications (CSPA), pp. 88–92. IEEE (2020)
10. Dar, M.N., Akram, M.U., Khawaja, S.G., Pujari, A.N.: CNN and LSTM-based emotion
charting using physiological signals. Sensors 20(16), 4551 (2020)
11. Chollet, F. (2017). Keras (2015)
12. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
13. https://www.eecs.qmul.ac.uk/mmv/datasets/deap/readme.html
14. Al-Qazzaz, N.K., Sabir, M.K., Ali, S., Ahmad, S.A., Grammer, K.: Effective EEG channels
for emotion identification over the brain regions using differential evolution algorithm. In:
2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC), pp. 4703–4706. IEEE (2019)
 Foundation for the Future of Higher Education
or ‘Misplaced Optimism’? Being Human
in the Age of Artificial Intelligence

Ashraf Alam(B) and Atasi Mohanty

Rekhi Centre of Excellence for the Science of Happiness, Indian Institute of Technology
Kharagpur, Kharagpur, West Bengal, India
[email protected]

Abstract. Several publications from across the globe have noted the rapid growth
of the field of Artificial Intelligence in Education (AIEd). Even though it has been
around for over 30 years, many professors are still confused about its implementa-
tion in their classrooms. This research surveys the existing literature on this issue
by means of a systematic review. Only 112 of the 2984 articles identified between
2006 and 2021 met the exclusion and inclusion criteria necessary for incorporation
in the final synthesis. Descriptive results show that STEM and subjects related to
computer science predominate in AIEd articles and that ‘quantitative methods’
predominate in empirical research. The findings are arranged into 4 categorical
groupings: (1) intelligent tutoring systems; (2) personalization and adaptive sys-
tems; (3) evaluation and assessment; and (4) prediction and profiling, all of which
can be applied to institutional and administrative services, academic support ser-
vices, and assessment and evaluation. The findings call for firm attention towards
the lack of critical thinking on the challenges and risks of AIEd and towards the
necessity for more comprehensive research on ethical and pedagogical techniques
in the deployment of AIEd in higher education.

Keywords: Smart tutoring systems · Machine learning · Pedagogy · Artificial

intelligence · Curriculum · Smart campuses · Teaching · Learning · Educational
technology

1 Introduction

In recent years, there has been a lot of buzz about how artificial intelligence (AI) may
be used in classrooms. The 2018 Horizon report cites artificial intelligence and adaptive
learning technologies as major advances in EdTech [1, 27]. Experts forecast a 43%
increase in the use of AI in education between 2018 and 2022. Netherlands’ Technical
University of Eindhoven has announced plans to hire 50 new faculty members to work
for a new ‘Institute for Artificial Intelligence Systems’ dedicated to AI education and
research [2, 3, 28]. Studies on the implementation of AI in classrooms date back around
30 years [4, 29].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 17–29, 2022.
https://doi.org/10.1007/978-3-031-23233-6_2
18 A. Alam and A. Mohanty

This year will be the 20th time the AIEd conference has been conducted by the
International AIEd Society [5, 6, 30]. The development of AI applications in higher
education raises new ethical challenges and risks [7, 31]. Managers, for instance, may
be tempted to replace training with more profitable automated AI solutions in times of
financial strain [8, 32]. Chatbots, expert systems, and intelligent tutors could cause con-
cern among educators, teaching assistants, counsellors, and administrators who work in
the education sector [9, 33]. While AI has the potential to enhance learning analytics,
the fact that these systems demand huge volumes of data, particularly sensitive informa-
tion about professors and students, creates major privacy and data protection problems.
Several groups, including the Analysis & Policy Observatory in Australia, have recently
published a framework for the ethical regulation of AI in educational settings [10, 34].
Every AI researcher has to be concerned about how their findings will be interpreted
morally. We are interested in learning about the novel ethical risks and repercussions that
have been envisioned by researchers and authors working in the field [11, 12, 35]. The
goal of this research article is to offer a synopsis of studies that examine AI’s potential
contributions to higher education.
Using a systematic review, this paper focuses on the following three areas of inquiry:

1. When looking back, how scholarly articles on AI have evolved?

2. Which ethical concerns, challenges, and potential hazards are taken into considera-
tion when imagining the role of AI in teaching?
3. In what ways might AI be used, and how broadly, in the context of higher education?

Though the origins of AI lie in computer science and engineering, other disciplines,
including philosophy, cognitive science, neurology, and economics, have had substantial
impacts on the area. Even though many academics are unaware of the breadth and,
more crucially, what is included, AI-based educational and pedagogical tools are being
introduced into higher education.

2 Education Using Artificial Intelligence (AIEd)

The first workshop on artificial intelligence (AI) was hosted in the 1950s by John
McCarthy at Dartmouth College in the United States. McCarthy first used the term arti-
ficial intelligence in the 1956 workshop proposal [13, 36]. Artificial intelligence studies
will continue on the premise that every aspect of learning or any other attribute of intel-
ligence may, in principle, be so perfectly characterized that a computer can be built to
reproduce it [14, 37]. The term artificial intelligence (AI) is often used to describe robots
that can do cognitive tasks, such as learning and problem-solving, that is traditionally
done by people. Machine learning, NLP, data mining, neural networks, and algorithms all
fall under the general category of artificial intelligence. There is no one technology that
is synonymous with AI [15, 38]. Machine learning (ML) is used to determine whether
students will pass a course or whether they will be accepted into a program [16, 39].
ML makes use of pattern recognition and prediction software to apply learned patterns
to problems [17, 40].
The concept of intelligent machines acting rationally is fundamental to AI. Every-
thing that can detect its environment and respond to it using some kind of actuator is
 Foundation for the Future of Higher Education 19

considered an agent [18, 41]. Specialists in the subject make a distinction between weak
and strong artificial intelligence as well as particular and wide artificial intelligence.
Whether computers can think for themselves, as opposed to just acting logically and
mimicking the human mind, is still a philosophical question open to debates and discus-
sions [19, 42]. An AI as advanced and comprehensive as this one appears unlikely to
appear very soon. In academia, intelligent agents and information systems are referred
to as GOFAI, which stands for “good old-fashioned AI”. John Haugeland, a philosopher,
came up with this term [20, 43].
There are now three types of AI software solutions utilized in education: intelli-
gent virtual reality, intelligent aid for group learning, and intelligent personal tutors.
Intelligent tutoring systems (ITS) have the potential to mimic in-person, one-on-one
tutoring [21, 44]. Based on learner models, algorithms, and neural networks, they could
choose the student’s learning path and the content to be taught. Cognitive scaffolding and
teacher-student communication are two further benefits [22, 45]. The vast majority of
research indicates that education is best accomplished in a group setting. Collaboration
and discussion are crucial to the educational process. It is essential, however, to facilitate
and govern digital teamwork [23, 46]. Assisting adaptive group construction based on
learner models, AIEd may also promote online group engagement or summarise talks
that a human teacher may use to guide students toward the course’s aims and objectives
[24, 47]. Collaborative learning may be aided by all these methods. Intelligent virtual
reality (IVR), which also makes use of ITS, is used to immerse students in realistic VR
and game-based learning environments [25, 48]. In remote or online labs, simulated
humans may play the roles of teachers, facilitators, or even other students [26, 49].
Artificial intelligence can provide immediate assessments and suggestions [21, 50].
Instead of relying only on periodic assessments, teachers may integrate AIEd into their
lesson plans for continuous monitoring of student progress [8, 51]. Algorithms have
been employed with a high degree of accuracy to predict whether a student would fail
an assignment or withdraw from a course [19, 52]. Learner-facing, teacher-facing, and
system-facing are the three primary perspectives from which to examine AI technologies
in education [13, 53].
Learner-facing artificial intelligence technologies include software used by students
to study a subject, such as adaptive or personalized learning management systems [18,
54]. Teacher-facing technology help reduce the workload of the educator by automating
tasks like course management, student assessment, instructor feedback, and plagiarism
checking [23, 55]. Teachers may monitor their students’ progress in their lessons with
the help of AIEd technologies and then respond accordingly with guidance and support.
System-facing helps managers and administrators gather data at the institutional level
[8, 56].
In this paper, we apply the concept of the student life cycle to the setting of higher edu-
cation as a theoretical model for describing the many services anchored in AI, available
both for administrative and institutional purposes and to support the academic teaching
and learning process.
20 A. Alam and A. Mohanty

3 Methods
A systematic review seeks to address specific issues by using a clear, consistent, and
repeatable search strategy to identify relevant studies for inclusion or exclusion. Next,
we summarise the findings, discuss how they may be used in practice, and call attention
to any gaps or inconsistencies by retrieving and codifying data from the included study.
This article provides a map of 112 papers dealing with AI in academic settings.

3.1 Search Strategy

Although problems exist in the peer-review process, our assessment restricted itself to
articles published in peer-reviewed journals due to their credibility and rigorous review
standards. The first search was performed in January of 2021 and 2984 records were
located. After weeding out the duplicates, it was agreed that only articles published in
2006 or later would be included. It was also decided that the corpus would only include
articles that dealt with the application of AI to academic settings.

3.2 Reliability of Agreement Amongst Raters

At this stage of screening, sensitivity rather than specificity was necessary, therefore
articles were included rather than rejected after being examined by a team of four coders.
Sessions were held often to discuss the first 80 entries and determine their inclusion or
exclusion. The coding decisions made by the four coders (A, B, C, and D) were evaluated
using Cohen’s kappa (κ), a coefficient for consistency across raters, to determine inter-
rater reliability. Random selection led to the examination of twenty papers. Overall, the
Kappa values for the four coders were quite high (between 0.81 and 0.89). Therefore, it
is reasonable to conclude that there is high inter-rater reliability.
After the first screening, 268 articles met the criteria to go on to the full-text phase.
Unfortunately, 81 publications could not be retrieved due to issues with the library’s order
system or with contacting the authors. A total of 189 articles were retrieved, screened,
and coded, however only 112 were included in the final synthesis.

3.3 Collection, Codification, and Analysis of Data

EPPI Reviewer, a tool used for systematic reviews, was loaded with all of the articles.
Articles were evaluated according to their research methodology (empirical vs. descrip-
tive, academic setting) and primary author’s area of study. AI applications were also
considered. The concept of AI and any references to its benefits and downsides in the
papers were also coded. Descriptive analysis was conducted in R, using the tidyr tool
for handling data.

3.4 Limitations

Although every effort was made to ensure the accuracy of the included studies, the
methods used in each research always result in some degree of error, and this systematic
 Foundation for the Future of Higher Education 21

review is no exception. Although the three databases of educational research that were
chosen are large and worldwide in reach, this analysis did not include research on AI that
had been published in other languages. This was accomplished by focusing on English-
language, peer-reviewed papers. Papers published in journals that were not indexed in any
of the three databases were not included in this review. Additional databases, publication
types, and language versions of publications might be used in future research to increase
the study’s breadth. This would need a thorough evaluation of the project’s available
resources and the feasibility of conducting the review.

4 Results

The term artificial intelligence (AI) is used to describe computer systems or intelligent
agents that exhibit human-like intelligence by mimicking human cognitive abilities such
as learning, memory, perception, and interaction with the environment, as well as lan-
guage [15, 57]. Artificial intelligence, or AI, refers to machines that can simulate human
intelligence. The purpose of this wide-ranging academic discipline is to learn about the
inner workings of the human brain so that we might apply those discoveries to the devel-
opment of better technological tools [8, 58]. Artificial intelligence (AI) may play the
role of either a teacher or a student in a language class.
The incorporation of agent-based, individualized instruction raises serious concerns
about personal data privacy [19, 59]. Agents may pick up on a wide variety of learner
attributes, like preferences and aptitude, without any human input. Data pertaining to an
individual’s identity is, in fact, confidential [4, 60].
Many learners are uncomfortable with the idea that their individual characteristics
may be made public. If they are struggling academically, students with special education
needs may feel their teachers will treat them differently. For this reason, the issue of
privacy must be resolved before employing agent-based systems [24, 60]. Many public
schools cannot afford the time and money needed to develop and execute AI-based
techniques, which is another barrier to its usage [13, 61].
We have utilized the concept of a student’s life-cycle as a framework to define the
many AI-based services provided at the institutional and administrative level, such as
admission, counselling, and library services, and at the academic support level, such
as evaluation, feedback, and tutoring [26, 62]. The following four domains of artificial
intelligence use were identified using a comprehensive literature study and iterative
coding procedure: affirmative action and personalization, evaluation and assessment,
profiling and prediction, and intelligent tutoring systems [5, 63].

4.1 Forecasting and Characterising

Many AI tools rely on learner models or profiles to make predictions, such as which
students are more likely to enroll in a program or to drop out, and to provide stu-
dents with timely assistance, feedback, and guidance on content-related concerns as
they go through their studies [12]. Dropout/retention, student models/performance, and
admission/scheduling are the three main subsets of profiling/prediction [24, 32].
22 A. Alam and A. Mohanty

Research in this area has made use of machine learning strategies for classifying
patterns, modelling student profiles, and making forecasts. Several machine learning
techniques, such as artificial neural networks (ANN), support vector machines (SVM),
random forests, and neural networks, are used to assess the overall prediction accuracy of
conventional logistic regression. In terms of percentage classification accuracy, machine
learning approaches exceed logistic regression [12, 33]. Classifier performance may also
be evaluated using the F1-score, which takes into consideration the proportion of cor-
rectly classified positive cases, erroneously classified negative instances, and incorrectly
classified positive instances [3, 34].
It is crucial to have a reliable evaluation of students’ academic performance for use
in admissions decisions and to boost educational services [5, 35]. The applications might
be sorted using the support vector machine (SVM) method, which has a 95% accuracy
rate. SVM may be used to find geographical patterns that provide prospective students
from specific places an edge in the college admissions process [19, 36].
Using a model trained with data from one state or province, an ANN may make
predictions about registration rates in other parts of the country [14, 37]. A student’s
course selection might be affected by several factors, including the quality of the course
and the instructor, the amount of work involved, the mode of delivery, and the scheduling
of the examination. Since admissions decisions can often be predicted with high accuracy,
an AI solution might free up administrative staff from routine tasks so that they can focus
on more complex cases [21, 38].
As a consequence of studies on student attrition and re-enrollment, early warning
systems have been developed to spot potentially disengaged freshmen. Predicting attri-
tion may be done using classification methods like logistic regression, decision trees
(DT), and artificial neural networks (ANN) [26, 39]. Students’ demographic, academic,
and economic characteristics should all be included in the data (e.g., age, sex, ethnicity,
GPA, etc.). The ANN model has the highest performance with an accuracy rating of over
80% which is checked after 10 rounds of cross-validation [15, 40].
Technology based on AI is helping with student profiling and modelling learning
behaviours for the purpose of predicting academic achievement [20, 41]. Several machine
learning algorithms are used to analyze student behavioural data from the virtual learning
environment to predict student involvement [6, 42]. Smart prescriptive algorithms enable
automated detection of disinterested learners and prompt intervention. Students’ progress
on projects is evaluated in workshops using face tracking and hand tracking [20, 43].
The outcomes supplied by multimodal data may help educators get insight into critical
features of project-based learning activities. How undergraduates learn to code may be
analyzed by looking at the code transcripts they write for software development projects
[11, 44]. Algorithms based on artificial intelligence may predict students’ academic
motivation based on their actions in a virtual classroom [21, 45]. Studying AI-based
models helps tremendously in creating intelligent tutoring systems and flexible learning
environments [5, 46].

4.2 Curriculum Technology that Uses Artificial Intelligence

Teaching-learning is the main priority for intelligent tutoring systems (ITS). In contrast,
there are some that have been institutionalized and given an administrative setting. The
 Foundation for the Future of Higher Education 23

first known ITS was the SCHOLAR system, which was established in 1970 and allowed
for question-and-answer sessions between professors and students but did not provide
continuous interaction [9, 47]. ITS is more efficient than other methods of instruction,
including lectures, textbooks, online reading, and homework [14, 48]. ITS provides stu-
dents with course materials and supports them via adaptive feedback, question-specific
strategies, and the capacity to identify when a student is struggling [7, 49]. The ITS
enables this by keeping tabs on where each student is at any given time.
Errors may be identified and corrected with the help of a discussion between the
learner and the machine using an instructional conversation toolkit-based tutoring system
or a pervasive interactive teaching robot with question recognition-based speech [9, 50].
The ITS acts as a peer mentor, collaborating with a student to find solutions to their
issues. They can talk, sign, and organize their messages. They do not get tutoring but
rather engage in cooperative problem-solving [24, 51]. A possible label for this scenario
is ‘cooperation amongst peers’. With the data gathered from each student’s participation
in the course, ITS can tailor their assistance to each individual’s needs while they are
enrolled in an online course [17, 52]. With this information, the system can recommend
the best books to read, exercises to do, and other individualized plans of action. A smart
assistant is stationed in a simulated statistical mechanics laboratory, where it presents
activities, assesses students’ understanding of the topic, and tailors its presentation of
the information to each individual learner [19, 53].
The basic purpose of ITS is to promote academically fruitful conversation movements
in online collaborative learning debates [23, 54]. It helps to encourage collaborative writ-
ing by automatically generating questions, providing automated feedback, and analyzing
the process. Teachers are able to act when necessary, thanks to intelligent assistants who
compile summaries of each student’s progress and engagement in group work, alert
notifications based on the identification of conflict situations, and data regarding each
student’s preferred method of learning [17, 55]. The ITS breaks up the tutoring obliga-
tions of the instructors by automating chores and offering timely feedback, leaving the
instructors responsible for supplying new clues and the correct responses. As a result, it
makes the lives of teachers easier [13, 56].

4.3 Constant Re-Evaluation

Artificial intelligence is used to determine how well students have retained the infor-
mation and applied what they have learned [5, 57]. Learners may benefit from using
tools like Latent Semantic Analysis and ePortfolios to map out unique courses of study.
Semantic web technology may be used to translate student credentials from different
universities to facilitate credit transfers [8, 58]. Course descriptions and syllabi may be
made available using these technologies as well. Algorithms may be used to pair up
prospective students with the skills and knowledge employers are looking for, allowing
instructors to tailor their lessons to the specific demands of the workforce. When it comes
to making judgments and evaluations, AI programs are highly precise and efficient. How-
ever, due to the calibration and training requirements of supervised machine learning
systems, they are best suited for large-enrollment classes or programs. Automated grad-
ing and feedback systems, assessing students’ knowledge, engagement, and academic
24 A. Alam and A. Mohanty

integrity, and evaluating teachers’ efficacy are the four main types of AI assessment and
evaluation used in the classroom [21, 59, 64].
Student essays may be graded automatically with the use of open-source Java soft-
ware used in automated grading, commonly known as Automated Essay Scoring (AES)
systems [18, 60]. Using AES would be challenging in small schools because of the large
number of pre-scored exams required for calibration, and AES may not be appropriate
for all forms of writing. One of the benefits of using algorithms to analyze text responses
is that it shifts the focus of evaluation away from students’ knowledge and abilities and
toward their ability to rewrite and improve their work [18, 61].
Intelligent agents provide reminders or help to pupils when they become lost or stuck
in their tasks. Aspiring pilots may now use software that will alert them if they start to
lose situational awareness in the air [23, 62]. The cognitive load on students is reduced
since lexical properties in machine learning systems give automatic feedback and help
students create better essays [8, 18]. For example, an adaptive testing-based automated
feedback system selects the most appropriate answers for each student based on Bloom’s
taxonomy of cognitive domains and then recommends related reading and exercises [9,
22]. Tools have been developed to help students evaluate their conceptual understanding
and get individualized instruction. These formulas factor in students’ performance on
tests and other assessments, as well as their activity in the virtual learning environment
[11, 16]. Academic integrity is assessed by using machine learning algorithms to identify
instances of possible plagiarism in student work [8, 17]. Data mining techniques are used
to analyze course evaluations and determine an instructor’s effectiveness based on several
different classification strategies [7, 18]. Comparatively speaking, using an algorithm to
evaluate instructional approaches yields more accurate results [19, 26].

4.4 A System that May Change to Fit the USER’S Needs

Biology, computer science, environmental education, and even animation design are just
some of the many areas that benefit from the usage of adaptive systems in the classroom
[8, 20]. Adaptive systems profile students’ behaviours and then provide them with indi-
vidualized assignments, resources, and content. E-learning recommendation systems
help students make informed course selections [21, 24]. Adaptive recommendation sys-
tems are useful in massive open online courses (MOOCs) because they provide students
with suggestions for further assignments, resources, and other users based on their indi-
vidual interests [22, 26]. Having access to students’ academic information helps teachers
with diagnostic tasks and personalize learning plans for each student [5, 23]. It also helps
with performance assessment and offers individualized help and criticism [11, 24].

5 Conclusion and Way Forward

This article looks at the field of AIEd. In this research, we provide a summary of the
many ways AI might be used to improve the educational experience for college stu-
dents, teachers, and administrators. Profiling and prediction, intelligent tutoring systems,
assessment and evaluation, and adaptive systems and customization were the four main
themes under which they were studied. In order to better conceptualize and comprehend
 Foundation for the Future of Higher Education 25

AIEd practice and research, a framework was developed via a systematic review. On
the other hand, there is still a substantial window of opportunity for educators to pursue
innovative and significant research and practice with AIEd that may have an effect on
teaching and learning in higher education due to the scarcity of longitudinal studies, the
predominance of descriptive and pilot studies from a technological standpoint, and the
predominance of quantitative methods, particularly quasi-experimental methods. There
is a dearth of research on the effects of policies and their actual implementation.
Throughout the various stages of a student’s academic career, AI-based tools and
services can be extremely helpful for all parties involved. This is important for uni-
versities with a large student body, like open and remote learning colleges. It could be
helpful to provide options for adaptable, interactive, and personalized education. This
may, for example, relieve teachers of the burden of manually grading hundreds, if not
thousands, of assignments, enabling them to focus on their core duty, i.e., providing
sympathetic human instruction. It is crucial to underline the need of considering not just
the technological elements of AIEd, but also the pedagogical, ethical, social, cultural,
and economic ones. The danger, of course, is in taking data and code at face value. Due
to education’s inherent complexity, it cannot be reduced to a set of purely quantitative
variables and methods. Digital data, like all digital technologies, may seem like a simple
technological solution to educational problems, but despite the promising results, they
do not offer such a solution.
Pedagogically sound goals, rather than technically feasible ones, should be pursued.
In China, teachers are already able to see data about their students’ participation and
emotions in class thanks to face recognition technology displayed on a dashboard. Some-
times, even the most sophisticated AI systems may make a mistake. An AI system cannot
acquire intelligence without the training data used to develop it. Every one of the many
issues discussed in the new UNESCO study on the prospects and obstacles of AIEd
for sustainable development is laden with important educational, societal, and ethical
implications. For example, it discusses how to assure inclusion and equality in AIEd,
how to educate teachers for AI-powered education, how to construct high-quality and
inclusive data systems, and how to be ethical and transparent while collecting, utilizing,
and distributing data.
The most surprising aspect of this assessment is the astounding lack of critical con-
sideration of the potential dangers and educational implications of using AI technologies
in higher education. Recently conducted empirical research seldom addressed privacy
problems in terms of ethical consequences. Educators and learning designers need to
perform more research on how to include AI applications throughout the student lifecy-
cle in order to take advantage of the immense potential that these technologies have for
constructing intelligent learning and teaching systems. Our systematic review’s absence
of authors with links to education departments demonstrates the need for a pedagogical
lens to be applied to these innovations in technology. Possible consequences of this the-
oretical vacuum for the whole area of educational technology are discussed. More than
forty percent of the studies in three premier journals for educational technology lacked
any practical application. Explicit educational views were missing from the papers that
were reviewed. Unfortunately, there is currently no evidence to back the development
26 A. Alam and A. Mohanty

of psychological and pedagogical theories of learning in connection with technologies

anchored in artificial intelligence.

References
1. Chatterjee, S., Bhattacharjee, K.K.: Adoption of artificial intelligence in higher education: a
quantitative analysis using structural equation modelling. Educ. Inf. Technol. 25(5), 3443–
3463 (2020)
2. Alam, A.: Challenges and possibilities in teaching and learning of calculus: a case study of
India. J. Educ. Gifted Young Sci. 8(1), 407–433 (2020)
3. Bates, T., Cobo, C., Mariño, O., Wheeler, S.: Can artificial intelligence transform higher
education? Int. J. Educ. Technol. High. Educ. 17(1), 1–12 (2020). https://doi.org/10.1186/s41
239-020-00218-x
4. Dilmurod, R., Fazliddin, A.: Prospects for the introduction of artificial intelligence technolo-
gies in higher education. Academicia: Int. Multi. Res. J. 11(2), 929–934 (2021)
5. Knox, J.: Artificial intelligence and education in China. Learn. Media Technol. 45(3), 298–311
(2020)
6. Alam, A.: Designing XR into higher education using immersive learning environments (ILEs)
and hybrid education for innovation in HEIs to attract UN’s education for sustainable devel-
opment (ESD) Initiative. In: 2021 International Conference on Advances in Computing,
Communication, and Control (ICAC3), pp. 1–9. IEEE (2021)
7. Guan, C., Mou, J., Jiang, Z.: Artificial intelligence innovation in education: a twenty-year
data-driven historical analysis. Int. J. Innovation Stud. 4(4), 134–147 (2020)
8. Chen, X., Xie, H., Zou, D., Hwang, G.J.: Application and theory gaps during the rise of
artificial intelligence in education. Comput. Educ.: Artif. Intell. 1, 100002 (2020)
9. Alam, A.: Educational robotics and computer programming in early childhood education:
a conceptual framework for assessing elementary school students’ computational thinking
for designing powerful educational scenarios. In: 2022 International Conference on Smart
Technologies and Systems for Next Generation Computing (ICSTSN), pp. 1–7. IEEE (2022)
10. Alyahyan, E., Düştegör, D.: Predicting academic success in higher education: literature review
and best practices. Int. J. Educ. Technol. High. Educ. 17(1), 1–21 (2020). https://doi.org/10.
1186/s41239-020-0177-7
11. Rampersad, G.: Robot will take your job: innovation for an era of artificial intelligence. J.
Bus. Res. 116, 68–74 (2020)
12. Chen, L., Chen, P., Lin, Z.: Artificial intelligence in education: a review. Ieee Access 8,
75264–75278 (2020)
13. Cope, B., Kalantzis, M., Searsmith, D.: Artificial intelligence for education: knowledge and
its assessment in AI-enabled learning ecologies. Educ. Philos. Theory 53(12), 1229–1245
(2021)
14. Alam, A.: Mapping a sustainable future through conceptualization of transformative learning
framework, education for sustainable development, critical reflection, and responsible citi-
zenship: an exploration of pedagogies for twenty-first century learning. ECS Trans. 107(1),
9827 (2022)
15. Yilmaz Ince, E., Kabul, A., Diler, İ: Distance education in higher education in the COVID-19
pandemic process: A case of Isparta Applied Sciences University. Int. J. Technol. Educ. Sci.
4(4), 345–351 (2020)
16. Alam, A. (2022). Employing Adaptive Learning and Intelligent Tutoring Robots for Virtual
Classrooms and Smart Campuses: Reforming Education in the Age of Artificial Intelligence.
In: Shaw, R.N., Das, S., Piuri, V., Bianchini, M. (eds) Advanced Computing and Intelligent
Technologies. Lecture Notes in Electrical Engineering, vol 914. Springer, Singapore
 Foundation for the Future of Higher Education 27

17. Holmes, W., Bialik, M., Fadel, C.: Artificial Intelligence in Education (2020)
18. Wang, T., Cheng, E.C.K.: Towards a tripartite research agenda: a scoping review of artificial
intelligence in education research. In: Cheng, E.C.K., Koul, R.B., Wang, T., Xinguo, Y.
(eds.) Artificial Intelligence in Education: Emerging Technologies, Models and Applications.
LNDECT, vol. 104, pp. 3–24. Springer, Singapore (2022). https://doi.org/10.1007/978-981-
16-7527-0_1
19. Nemorin, S., Vlachidis, A., Ayerakwa, H.M., Andriotis, P.: AI hyped? A horizon scan of
discourse on artificial intelligence in education (AIED) and development. Learn., Media
Technol. 1–14 (2022). https://doi.org/10.1080/17439884.2022.2095568
20. Alam, A.: Cloud-based e-learning: scaffolding the environment for adaptive e-learning
ecosystem based on cloud computing infrastructure. In: Computer Communication, Net-
working and IoT: Proceedings of 5th ICICC 2021, vol. 2, pp. 1–9. Springer Nature Singapore,
Singapore (2022)
21. Holmes, W., Porayska-Pomsta, K. (eds.): The Ethics of Artificial Intelligence in Education:
Practices, Challenges, and Debates. Taylor & Francis (2022)
22. Chen, X., Zou, D., Xie, H., Cheng, G., Liu, C.: Two decades of artificial intelligence in
education. Educ. Technol. Soc. 25(1), 28–47 (2022)
23. Paek, S., Kim, N.: Analysis of worldwide research trends on the impact of artificial intelligence
in education. Sustainability 13(14), 7941 (2021)
24. Akgun, S., Greenhow, C.: Artificial intelligence in education: addressing ethical challenges
in K-12 settings. AI and Ethics 2, 431–440 (2021). https://doi.org/10.1007/s43681-021-000
96-7
25. Bhimdiwala, A., Neri, R.C., Gomez, L.M.: Advancing the design and implementation of
artificial intelligence in education through continuous improvement. Int. J. Artif. Intell. Educ.
32(3), 756–782 (2022)
26. Alam, A.: Test of knowledge of elementary vectors concepts (TKEVC) among first-semester
bachelor of engineering and technology students. Periódico Tchê Química 17(35), 477–494
(2020)
27. Ouyang, F., Jiao, P.: Artificial intelligence in education: the three paradigms. Comput. Educ.:
Artif. Intell. 2, 100020 (2021)
28. Holmes, W., et al.: Ethics of AI in education: towards a community-wide framework. Int. J.
Artif. Intell. Educ. 32(3), 504–526 (2022)
29. Megahed, N.A., Abdel-Kader, R.F., Soliman, H.Y.: Post-pandemic education strategy: frame-
work for artificial intelligence-empowered education in engineering (AIEd-Eng) for lifelong
learning. In: International Conference on Advanced Machine Learning Technologies and
Applications, pp. 544–556. Springer, Cham (2022)
30. Dickler, R., Dudy, S., Mawasi, A., Whitehill, J., Benson, A., Corbitt, A.: Interdisciplinary
approaches to getting ai experts and education stakeholders talking. In: Rodrigo, M.M., Mat-
suda, N., Cristea, A.I., Dimitrova, V. (eds.) Artificial Intelligence in Education. Posters and
Late Breaking Results, Workshops and Tutorials, Industry and Innovation Tracks, Practition-
ers’ and Doctoral Consortium. AIED 2022. Lecture Notes in Computer Science, vol. 13356.
Springer, Cham (2022). https://doi.org/10.1007/978-3-031-11647-6_20
31. Alam, A.: Positive psychology goes to school: conceptualizing students’ happiness in 21st
century schools while ‘minding the mind!’are we there yet? evidence-backed, school-based
positive psychology interventions. ECS Trans. 107(1), 11199 (2022)
32. Wang, T., Cheng, E.C.K.: An investigation of barriers to Hong Kong K-12 schools
incorporating artificial intelligence in education. Comput. Educ.: Artif. Intell. 2, 100031
(2021)
33. Chan, L., Hogaboam, L., Cao, R.: Artificial intelligence in education. In: Applied Artificial
Intelligence in Business, pp. 265–278. Springer, Cham (2022)
28 A. Alam and A. Mohanty

34. Khosravi, H., et al.: Explainable artificial intelligence in education. Comput. Educ.: Artif.
Intell. 3, 100074 (2022)
35. Alam, A.: A digital game based learning approach for effective curriculum transaction for
teaching-learning of artificial intelligence and machine learning. In: 2022 International Con-
ference on Sustainable Computing and Data Communication Systems (ICSCDS), pp. 69–74.
IEEE (2022)
36. Schiff, D.: Out of the laboratory and into the classroom: the future of artificial intelligence in
education. AI & Soc. 36(1), 331–348 (2020). https://doi.org/10.1007/s00146-020-01033-8
37. Sadiku, M.N., Musa, S.M., Chukwu, U.C.: Artificial Intelligence in Education. iUniverse
(2022)
38. Lameras, P., Arnab, S.: Power to the teachers: an exploratory review on artificial intelligence
in education. Information 13(1), 14 (2021)
39. Ezzaim, A., Kharroubi, F., Dahbi, A., Aqqal, A., Haidine, A.: Artificial intelligence in
education-State of the art. Int. J. Comput. Eng. Data Sci. 2(2), 1–11 (2022)
40. Alam, A.: Possibilities and apprehensions in the landscape of artificial intelligence in edu-
cation. In: 2021 International Conference on Computational Intelligence and Computing
Applications (ICCICA), pp. 1–8. IEEE (2021)
41. Bozkurt, A., Karadeniz, A., Baneres, D., Guerrero-Roldán, A.E., Rodríguez, M.E.: Artificial
intelligence and reflections from educational landscape: a review of AI studies in half a
century. Sustainability 13(2), 800 (2021)
42. Sharples, M.: Automated essay writing: an AIED opinion. Int. J. Artif. Intell. Educ. 32(4),
1119–1126 (2022). https://doi.org/10.1007/s40593-022-00300-7
43. Biswas, G., Bull, S., Kay, J., Mitrovic, A. (eds.): AIED 2011. LNCS (LNAI), vol. 6738.
Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-21869-9
44. Ye, R., Sun, F., Li, J.: Artificial intelligence in education: origin, development and rise. In:
Liu, X.-J., Nie, Z., Jingjun, Y., Xie, F., Song, R. (eds.) ICIRA 2021. LNCS (LNAI), vol.
13016, pp. 545–553. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-89092-6_49
45. Alam, A.: Should robots replace teachers? Mobilisation of AI and learning analytics in edu-
cation. In: 2021 International Conference on Advances in Computing, Communication, and
Control (ICAC3), pp. 1–12. IEEE (2021)
46. Zhang, K., Aslan, A.B.: AI technologies for education: Recent research & future directions.
Comput. Educ.: Artif. Intell. 2, 100025 (2021)
47. Khazanchi, R., Khazanchi, P.: Artificial intelligence in education: a closer look into intelligent
tutoring systems. In: Handbook of research on critical issues in special education for school
rehabilitation practices, pp. 256–277. IGI Global (2021)
48. Nalbant, K.G.: The importance of artificial intelligence in education: a short review. J. Rev.
Sci. Eng. 2021, 1–15 (2021)
49. Zheng, Y., Zhou, Z., Blikstein, P.: Towards an inclusive and socially committed community
in artificial intelligence in education: a social network analysis of the evolution of authorship
and research topics over 8 years and 2509 papers. In: Rodrigo, M.M., Matsuda, N., Cristea,
A.I., Dimitrova, V. (eds.) Artificial Intelligence in Education. AIED 2022. Lecture Notes in
Computer Science, vol. 13355. Springer, Cham (2022). https://doi.org/10.1007/978-3-031-
11644-5_34
50. Lameras, P., Paraskakis, I., Konstantinidis, S.: A rudimentary progression model for artificial
intelligence in education competencies and skills. In: Auer, M.E., Tsiatsos, T. (eds.) IMCL
2021. LNNS, vol. 411, pp. 927–936. Springer, Cham (2022). https://doi.org/10.1007/978-3-
030-96296-8_84
51. Alam, A.: Social robots in education for long-term human-robot interaction: socially sup-
portive behaviour of robotic tutor for creating robo-tangible learning environment in a guided
discovery learning interaction. ECS Trans. 107(1), 12389 (2022)
 Foundation for the Future of Higher Education 29

52. Hamal, O., El Faddouli, N.E., Harouni, M.H.A., Lu, J.: Artificial intelligent in education.
Sustainability 14(5), 2862 (2022)
53. Campbell, C.: Artificial intelligence in education: contributors, collaborations, research
topics, and challenges. Educ. Technol. Soc. 25(1), 28–47 (2022)
54. Ungerer, L., Slade, S.: Ethical considerations of artificial intelligence in learning analytics in
distance education contexts. In: Prinsloo, P., Slade, S., Khalil, M. (eds.) Learning Analytics
in Open and Distributed Learning: Potential and Challenges, pp. 105–120. Springer Nature
Singapore, Singapore (2022). https://doi.org/10.1007/978-981-19-0786-9_8
55. Corbeil, M.E., Corbeil, J.R.: Establishing trust in artificial intelligence in education. In: Pal-
iszkiewicz, Joanna, Chen, Kuanchin (eds.) Trust, Organizations and the Digital Economy:
Theory and Practice, pp. 49–60. Routledge, New York (2021). https://doi.org/10.4324/978
1003165965-5
56. Xia, Q., Chiu, T.K., Lee, M., Sanusi, I.T., Dai, Y., Chai, C.S.: A self-determination theory
(SDT) design approach for inclusive and diverse artificial intelligence (AI) education. Comput.
Educ. 189, 104582 (2022)
57. Alam, A.: Possibilities and challenges of compounding artificial intelligence in India’s
educational landscape. Int. J. Adv. Sci. Technol. 29(5), 5077–5094 (2020)
58. Dai, C.P., Ke, F.: Educational applications of artificial intelligence in simulation-based
learning: a systematic mapping review. Comput. Educ.: Artif. Intell. 3, 100087 (2022)
59. Yang, S.J., Ogata, H., Matsui, T., Chen, N.S.: Human-centered artificial intelligence in edu-
cation: seeing the invisible through the visible. Comput. Educ.: Artif. Intell. 2, 100008
(2021)
60. Alam, A.: Investigating sustainable education and positive psychology interventions in
schools towards achievement of sustainable happiness and wellbeing for 21st century
pedagogy and curriculum. ECS Trans. 107(1), 19481 (2022)
61. Channa, F.R., Sarhandi, P.S.A., Bugti, F., Pathan, H.: Harnessing artificial intelligence in
education for preparing learners for the 21st century. Elementary Educ. Online 20(5), 3186
(2021)
62. Xu, W., Ouyang, F.: The application of AI technologies in STEM education: a systematic
review from 2011 to 2021. Int. J. STEM Educ. 9(1), 1–20 (2022)
63. Alam, A.: Pedagogy of calculus in India: an empirical investigation. Periódico Tchê Química
17(34), 164–180 (2020)
64. Chaudhry, M.A., Kazim, E.: Artificial intelligence in education (AIEd): a high-level academic
and industry note 2021. AI and Ethics 2(1), 157–165 (2022)
 AI Enabled Internet of Medical Things
Framework for Smart Healthcare

Jyoti Srivastava1,2 and Sidheswar Routray1(B)

1 Department of Computer Science and Engineering, School of Engineering, Indrashil
University, Rajpur, Mehsana, Gujarat, India
2 Department of Computer Engineering, Unitedworld School of Computational Intelligence,

Karnavati University, Gandhinagar, India

Abstract. Internet of Medical Things (IoMT) is basically the use of the Internet
of Things (IoT) in Smart Healthcare where automation in health monitoring is
provided using Artificial Intelligence (AI) so that without visiting the hospital for
every small problem an individual can get consultancy from doctors remotely. As
per the pandemic situation visiting the hospital an individual can result in con-
tact with harmful viruses, so using this Smart HealthCare System (SHS) without
visiting the hospital an individual can monitor their day-to-day health record and
can take preliminary precautions accordingly. Three major health domains are
defined in this work where AI-based SHS can help the individual depending on
the patient’s health situation they can belong to one of these domains. Now, SHS
deals with highly sensitive patient health data so few parameters must be satis-
fied by this system i.e., data security, data accuracy, system efficiency, Quality of
Service (QoS), System Reliability, etc. In this work, we will elaborate discussion
about these parameters and provide a comparison table to analyze the work done
by different researchers in this era to improve the performance of an Artificial
Intelligence (AI) based SHS using the IoMT framework.

Keywords: Internet of medical things (IoMT) · Smart healthcare system (SHS) ·

Artificial intelligence (AI) · Data accuracy · Energy efficiency · Quality of
service (QoS)

1 Introduction
The demand for advanced medical treatment is increasing everyday to ensure better
health and protectiveness from both tangible and non-tangible viruses thereby accelerat-
ing the need for automation in Healthcare System. On top of the new healthcare era, the
AI-based Internet of Medical Things (IoMT) framework is developing with various new
creative and smart solutions to various real-world problems [1]. Various organizations
and researchers are showing their interest in the Internet of Things (IoT) to enhance their
scope of work and throughput. IoT seems to be the new generation technology with a
universal applicability in almost every market field with increasing integration degree of
end products, systems, and services. IoT is basically a collection of nodes (computers,
laptops, Smartphones, or other smart devices) having their unique IDs connected via

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 30–46, 2022.
https://doi.org/10.1007/978-3-031-23233-6_3
 AI Enabled Internet of Medical Things 31

wired or wireless medium and able to communicate without any human intervention.
IoT can be used in various fields for providing automation in the implementation of smart
techniques for the betterment of society [2]. Figure 1 shows the use of IoT in different
eras including agriculture services [3], healthcare services [1], educational institutes [4],
traffic management [5], etc.

Fig. 1. Role of IoT in various market fields.

Artificial Intelligence gives the automation for providing essential features in IoMT
framework-based Smart Healthcare systems, System Security where AI can be used to
identify the intrusion within the system [6], detect immediate security attacks [7], Web-
based security assessment [8], etc. AI can also be used to provide an automatic alert to
all the respective medical staff for immediate actions in emergencies [9]. Along with the
patients, this technique will be beneficial for the doctors also as using AI doctors will
be able to easily maintain and monitor patient’s health records and can provide remote
medical advice.
Therefore, considering the susceptibility of patients’ health data we are motivated to
provide this comparative analysis of AI-based SHS using IoMT architecture using the
following parameters: data security, data accuracy, system efficiency, Quality of Service
(QoS), and System Reliability. The major contributions of the work are as follows.

• To discover the role of AI in different health domains of the IoMT framework along
with their application in SHS.
• To examine different IoMT architectures of AI-based SHS.
• To present a comparative analysis of various accurate data collection techniques
thereby creating a reliable AI-based SHS.
• To scrutinize different Data Security techniques to ensure protectiveness in AI-based
SHS.
32 J. Srivastava and S. Routray

• To present an extensive examination of different system efficiency algorithms for an

AI-based SHS.
• To present a comparative analysis of different QoS parameters for an AI-based SHS

This work consists of total 5 sections where Sect. 2 represents different AI-based
health domains with their application in SHS, Sect. 3 elaborates on different AI-enabled
IoMT architectures proposed by different authors for SHS, Sect. 4 gives an extensive
examination of different research challenges of AI-enabled SHS and the last section i.e.,
Sect. 5 is the concluding remarks for this work.

2 AI Based IoMT Health Domains

Mainly there are three domains in Smart Healthcare that are monitored remotely using
Artificial Intelligence i.e., Selfcare, Acute care, and Homecare [10]. The Selfcare system
deals with day-to-day medical problems where a person can easily monitor his regu-
lar health using AI-based smart devices and can take preliminary actions accordingly.
The next level is the Homecare system where patients can be monitored remotely by
healthcare providers and an alarm system is used to alert the medical staff in case of an
emergency. The acute care system is responsible only for critical situations that need
urgent responses from wearable/implanted AI-based smart devices that are specially
used in elderly care devices like fall detection. Figure 2 elaborates on all three AI-based
SHS domains.

Fig. 2. AI enabled smart healthcare domains

All three domains use different types of sensors or collections of different sensors
to fetch accurate medical data about the individual. For example, LM35 senses body
temperature usually used in self-care systems, DHT11 is also a humidity and temperature
 AI Enabled Internet of Medical Things 33

sensor used in Homecare Systems and AD8232 is an Electrocardiogram Sensor usually

used in Acute care Systems [11].

3 AI Enabled IoMT Architectures for Smart Healthcare Systems

In most AI-enabled Smart Healthcare Systems as shown in Fig. 3, the whole architecture
consists of three basic layers, where the bottommost one is primarily used for collecting
patients’ health records from wearable/implanted sensors on the human body, the middle
layer primarily analyses and does the processing on the collected data using some AI
algorithms and the topmost layer mostly deals with the AI results display to the end user
using some interface. But depending on different system requirements, names, protocols,
and features can be added or removed by different authors.

Fig. 3. AI enabled IoMT architectures for smart healthcare systems

Sun et al. [1] introduces a security protocol using AI where a three-tier architecture
is used namely the Sensor level, and Personal server level, Medical Server level where
medical server layer consists of various core algorithms and programs, and the sensor
level deals with sensors and medical devices and personal server level is a collection
of few personal servers for internal process and data storage [33]. Kumar et al. [12]
proposed a three-layered end-to-end architecture for connecting IoT sensors to smart
healthcare i.e., the Data storage layer, Data collection layer, and Data processing Layer.
The data collection layer consists of IoT devices for sensing medical data, then the data
storage layer is used to store this record on a wider range, and the last one i.e., the data
processing layer does an analysis of received data [34]. Sun et al. [13] proposed an AI-
based IoMT architecture mainly consisting of three layers namely, the Network Layer,
Perceptual Layer, and the Application Layer. The perceptual layer collects medical data
from the on-body sensors and transforms it into important information. The next layer
i.e., the network layer is responsible for platform and interface-related services required
for implementing particular data transmission techniques. The topmost layer i.e., the
application layer deals with different healthcare equipment and utilizes all collected
medical information for managing patient healthcare data remotely.

4 Research Challenges of AI Enabled Smart Healthcare Systems

AI-enabled IoMT framework for smart healthcare systems ensures automation and flex-
ibility for the end user but to create a such real-time framework, the developing team
34 J. Srivastava and S. Routray

needs to tackle various research challenges. There can be various research challenges that
are to be monitored during the development of AI-enabled SHS like amount of energy
consumed, packet delivery ratio, battery lifetime, quality of service, body movements,
temperature change, range of transmission, heterogeneous environment, power drain,
network throughput, delay, transmission rate, etc. In this research paper, we are focusing
on a few of these challenges and the rest can be explored in future work. The challenges
that occurred during IoMT system development differ from the wireless network as it
deals with very sensitive real-time medical data of the patients and a small delay, data
loss, or inconsistency in this healthcare system can result in severe health issues and can
be very risky for someone’s health. All these challenges need to be handled with care
during the development of AI-enabled SHS using the IoMT framework for a better user
experience. This system works upon very sensitive medical data of various patients using
small and ultra-low power wearable/implanted smart IoMT devices therefore most of the
real-time development challenges are handled in network and protocol designing tech-
niques considering effective topology, energy consumption, and effective channel. So,
this section provides a scrutinized analysis of different research challenges that occurred
during the development of the AI-based SHS system using the IoMT framework.

4.1 Data Accuracy

An AI-enabled SHS works with highly sensitive medical records and the AI algorithm
will predict efficient results only when it gets accurate medical data. If there is any
inconsistency in fetching the medical record from the patient’s body it can result in
severe physical and financial loss and can thereby end up with total system failure.
Hence in this section, we are elaborating Data accuracy parameter of an SHS. The data
accuracy comparison has been illustrated in Fig. 4. Table 1 shows a comparative analysis
of various work done regarding data accuracy in AI-enabled Smart Healthcare systems.
Tekieh MH et al. [14], 2015 proposed an electronic health record approach for the
collection of data using various techniques based on system requirements to accomplish
the accuracy of collected data. Shahin A et al. [15], 2014 also use an electronic health
record approach for data collection considering two parameters error rate and accuracy
of the collected data. Yang L et al. [16], 2016 use a Rule-Based approach to improve data
gathering speed maintaining the accuracy of collected data. Mdaghri Z A., et al. [17],
2016 proposed a support system for clinical decisions improving accuracy within the
collected data. Roy S et al. [18], 2016 proposed a new technique i.e., correlation-based
ratio analysis gathers medical information using parameters correlation and accuracy.
Rao AR et. al. [19], 2016 use an open dataset for prediction and accuracy having better
visualization because of GUI.
 AI Enabled Internet of Medical Things 35

Table 1. Data accuracy in AI enabled smart healthcare system

Sr. No. Authors and Parameters Data collection Advantages Disadvantages

publication technique
years
1 Tekieh MH Accuracy Electronic Few EHR based Future work not
et al. [14], health record techniques were suggested
2015 (EHR) discussed so
that based on
the problem
statement
optimal solution
can be choosed
2 Shahin A et. Accuracy and EHR Case study Can be applied
al. [15], 2014 error rate only to special
environmental
conditions
3 Yang L et al. Accuracy Rule – based Improvement in Still
[16], 2016 approach speed of data improvement in
collection accuracy
needed
4 MdaghriZA Accuracy Clinical Collected data Missing Values
et al. [17], decision is more accurate not handled
2016 support system properties
5 Roy S et al. Accuracy and Correlation Correlation Missing Values
[18], 2016 correlation based ratio helps in not considered
analysis collecting in this work
specific medical
record
6 Rao AR et. Accuracy and Used open Improved GUI Missing Values
al. [19], 2016 prediction dataset helps in not considered
improved in this work
visualization
36 J. Srivastava and S. Routray

Comparison for Accuracy

CR
GR
IG
% of Improvement
J48
in Accuracy
ID3
RF
0 50 100 150

Fig. 4. Accuracy comparison

4.2 Data Security

AI-based SHS mostly uses the cloud for storage of data so that we can get fast access
to medical data whenever required but it can lead to security threats of highly sensitive
medical data. Any security breach within an SHS can result in the tempering of sensitive
patient’s medical data and thereby causing treatment failure as well as users’ faith in
using smart IoMT devices. Table 2 shows a scrutinized analysis of different work done
regarding security in AI-enabled Smart Healthcare systems. Moosavi, S. R., et al. [20],
2015 proposed techniques to Encrypt system information and used biometric identifi-
cation and two-factor authentication for system security against unauthorized access.
Sun, Y., et al. [1], 2019 implement secure routing algorithms to avoid routing attacks
where the Attacker modifies the route of traffic to a new destination. This paper also
uses encryption and an intrusion detection system, builds redundancy in infrastructure,
and uses region mapping, authentication, and egress filtering to avoid Denial of Service
(DoS) attacks where the attacker generates so much network traffic that the SHS halts;
generally caused by a compromised node. Verma, G., et al. [21], 2021 uses advanced
techniques to Encrypt links and the network layer preventing message disclosure where
attacker targets sensitive information disclosure to access a patient’s log file. This paper
also uses hashing and digital signatures to avoid message modification by unautho-
rized users where an attacker can modify messages between a patient and a healthcare
provider. Also, the author discusses encryption, segmentation, and implementation of
network access control (NAC) techniques to avoid eavesdropping, and reply attacks
where the attacker listens to information through an open SHS communication channel
and can forward modified information. Verma, G., et al. [21], 2021 also uses symmet-
ric key security algorithms for compromised node attack where the attacker hacks into
openly deployed sensor nodes and injects false information into them. Sun, Y., et. al.
[1], 2019 use multi-path, multi-base data forwarding, identity verification protocols, and
cryptography to avoid Hello Flooding where the attacker uses a compromised node with
high transmission power to compromise all of its neighbors.
 AI Enabled Internet of Medical Things 37

Table 2. Data Security in AI enabled smart healthcare system

Sr.No Authors & Attack or threat Description Counter measures

publication years
1 Moosavi, S. R., Unauthorized Attacker gains access Encrypt system
et al. [20], 2015 access to a system by using information, use
another authorized biometric
user’s credentials identification or
two-factor
authentication
2 Sun, Y., et. al. Routing attacks Attacker modifies the Implement secure
[1], 2019 route of traffic to a routing algorithms
new destination
3 Verma, G., et al. Message Attacker targets Encrypt links and the
[21], 2021 disclosure sensitive information network layer
disclosure to access a
patient’s log file
4 Verma, G., et al. Message Attacker modifies Use hashing and
[21], 2021 modification messages between a digital signatures
patient and
healthcare provider
5 Verma, G., et al. Eavesdropping Attacker listens to Use encryption,
[21], 2021 information through segmentation, and
an open SHS implement network
communication access control (NAC)
channel
6 Verma, G., et al. Replying attack Attacker forwards Use encryption,
[21], 2021 modified information segmentation, and
after eavesdropping implement NAC
7 Verma, G., et al. Compromised Attacker hacks into Use symmetric key
[21], 2021 node attack openly deployed security algorithms
sensor nodes and
injects false
information
8 Sun, Y., et. al. Denial of service Attacker generates so Use encryption and
[1], 2019 (DoS) attack much network traffic an intrusion detection
that the SHS halts; system, build
generally caused by a redundancy in
compromised node infrastructure, and
use region mapping,
authentication, and
egress filtering
(continued)
38 J. Srivastava and S. Routray

Table 2. (continued)

Sr.No Authors & Attack or threat Description Counter measures

publication years
9 Sun, Y., et. al. Hello flooding Attacker uses a Use multi-path,
[1], 2019 compromised node multi-base data
with high forwarding, identity
transmission power verification protocols,
to compromise all of and cryptography
its neighbors
10 Algarni, A. [10], Black and gray Attacker inserts a Use a time-based
2019 hole attack malicious node into a threshold mechanism,
network that changes track pending packet
routing tables so that tables and node rating
neighboring nodes tables, and make sure
send the all nodes have
compromised node different IDs
all their data. Black
hole attacks don’t
reply to the
neighboring nodes;
gray hole attacks
reply with
non-critical data
11 Algarni, A. [10], Sybil attack Attacker uses a Validate sensors at a
2019 malicious sensor that central authority or
masquerades as by using sensor graph
multiple sensors to connectivity
modify the routing characteristics
table
12 Algarni, A. [10], Social engineering Attacker influences Raise awareness of
2019 users to reveal security concerns
information or through training,
perform an action auditing, and
that benefits the adequate security
attacker policies

Algarni, A. [10], 2019 uses a time-based threshold mechanism, to track pending

packet tables and node rating tables, and make sure all nodes have different IDs to avoid
Black and Gray hole attacks where the attacker inserts a malicious node into a network
that changes routing tables so that neighboring nodes send the compromised node all
their data. Black hole attacks don’t reply to the neighboring nodes; Gray hole attacks
reply with non-critical data. This paper uses various techniques to validate sensors at a
central authority or by using sensor graph connectivity characteristics to avoid a Sybil
attack where the attacker can use a malicious sensor that masquerades as multiple sensors
 AI Enabled Internet of Medical Things 39

to modify the routing table. It also Raises awareness of security concerns through train-
ing, auditing, and adequate security policies for social engineering where the attacker
influences users to reveal information or perform an action that benefits the attacker.

4.3 System Efficiency

The efficiency of an SHS will determine the size of the IoMT device, battery lifetime, and
usability of the smart devices which is usually very small hence the energy optimization
is handled by the routing protocol only. The energy consumption comparison has been
illustrated in Fig. 5. Therefore, a comparative analysis of various work done regarding
system efficiency is shown in Table 3 for an AI-enabled SHS. AI-enabled Smart Health-
care System collects data either from an implant or wearable smart healthcare devices.
Now the implant devices must maintain a long battery life of approx. 10–15 years to
prevent multiple surgery repetition [16]. Even if we consider wearable devices battery
lifetime is again very important because repeatedly changing the battery will reduce
the usability of smart devices. Rehman, A., et al. [22], 2021 proposed a new Energy
efficient IoT e-health model using AI with a homomorphic secret sharing technique for
better Packet delivery ratio, Delivery time, energy consumption, and Data leakage of
an IoMT device. Sodhro et. al. [23], 2021 proposed another energy-efficient technique
for better system efficiency to improve battery life while consuming minimal energy.
Lazarevska et al. [24], 2018 show a New Objective Function (NEWOF) for improved
energy consumption and total control of traffic overhead. Saba, T., et al. [25], 2020
projected another energy-efficient and secure e-healthcare technique to present a highly
secure healthcare system with increased medical data delivery with minimum energy
consumption. AbdulmohsinHammood D., et. al. [26], 2019 proposed an inter-WBAN
cooperation algorithm for an IoMT environment giving increased energy efficiency and
outage probability.

Table 3. System efficiency in AI enabled smart healthcare system

Sr.no Author and Parameters Efficiency Advantages Disadvantages

publication year measurement
technique
1 Rehman A., et. al. Packet Energy efficient 1. Disease 1. Increased
[22], 2021 Delivery IoT e-health diagnosis PDR under
Ratio (PDR), model using AI system very high load
Delivery with maintainability scenario
time, energy homomorphic increased 2. Different
consumption secret sharing 2. Secure IoT nodes
and data communication energy
leakage with medical consumptions
cloud are different
3. Low
intelligence to
avoid packet
collision
(continued)
40 J. Srivastava and S. Routray

Table 3. (continued)

Sr.no Author and Parameters Efficiency Advantages Disadvantages

publication year measurement
technique
2 Sodhro et. al. [23], Charge Energy efficient Increased Computational
2021 dissipation, algorithm for battery lifetime load is high
energy better system consuming
dissipation, efficiency minimal energy
battery
lifetime and
energy
discharge
3 Lazarevska et. al. PDR, New objective 1.Better energy 20% loss
[24], 2018 network function for efficiency during
lifetime, system energy 2. Improved mobility
energy TCTO plugin
consumption, 3. Better PDR
total control
traffic
overload
(TCTO)
4 Saba, T., et. al. [25], E2E delay, Energy-efficient 1. Reduced 1. Mobility
2020 throughput, and secure energy parameter not
packet loss, e-healthcare consumption considered
link technique 2. Increased 2.
breakages data delivery Inter-WBAN
and energy 3. Minimum data
consumption network delay transformation
with high requires
security improvement
in security and
energy
consumption
parameters
5 AbdulmohsinHamood Outage Inter-WBAN Highly efficient Low outage
D., et. al. [26], 2019 probability, cooperation with high probability
power algorithm for outage during
consumption, efficient IoMT probability for asymmetric
energy framework symmetric transmission
efficiency, transmission
transmission
rate
 AI Enabled Internet of Medical Things 41

Comparison based on Energy

Consumpon %

SEF-IoMT
% of Improvement in
NEWOF
Energy Consumpon
energy-efficient…
0 10 20 30 40

Fig. 5. Energy consumption comparison

4.4 Quality of Service

AI enables SHS applications that deal with highly sensitive real-time medical data like
ECG which is susceptible to execution timing and data loss. To fulfill such real-time
medical demand Quality of service (QoS) requirements must be justified in an AI-
enabled IoMT framework for SHS. To maintain system usability the sensing devices of
SHS have fixed memory and computational capabilities, so the QoS measures must be
adapted by the routing protocols. Table 4 shows a comparative analysis of various work
done regarding QoS in AI-enabled Smart Healthcare systems. Kumar, A., et al. [27], 2022
proposed a case study of various algorithms to improve the QoS of the IoMT framework.
Patan, R., et al. [28], 2020 proposed an AI-driven IoMT eHealth architecture implying
a grey filter Bayesian convolution neural network that will improve accuracy during
medical data analysis. Agnihotri, S., et al. [29], 2019 proposed the CARA-IoT algorithm
using ACO (AntColony Optimization) approach to choose the best consistent and smart
routing path depending firstly on the channel noise and secondly on communicating
content. Singh, P. D. et al. [30], 2021 proposed an ensemble-based classifier algorithm
that combines AI with Fog computing for smart health of early covid-19 detection
and also takes shared storage peer-to-peer, time stamping advantages of blockchain.
Khodkari, H., et al. [31] use multi-attribute decision-making techniques specifically the
simple weighting technique for better QoS, and due to the integration of cloud and IoT,
this SHS system gains the support of the new developers in this era. Sodhro, A. H., et.
al. [32]. 2021 proposed a new adaptive QoS computation algorithm (AQCA) to monitor
the performance indicators fairly and efficiently.
42 J. Srivastava and S. Routray

Table 4. Quality of service (QoS) in AI enabled smart healthcare system

Sr.No Authors & Parameters QoS Technique Advantages

publication years
1 Kumar, A., et al. QoS Case study Elaborated various
[27], 2022 algorithms so that as
per the project
requirement, the best
suitable technique
can be opted
2 Patan, R., et al. Time and overhead An AI driven IoMT 1. Design
[28], 2020 framework for suggestions
e-Health elaborated
architecture using considering
Grey filter Bayesian accuracy, overhead
convolution neural and execution time
network in comparison with
state-of-the-art
methods
2. Accurate medical
data analysis
3 Agnihotri, S., Packet loss and CARA-IoT This algorithm will
et al. [29], 2019 delay algorithm using choose the best
ACO (AntColony consistent and smart
Optimization) routing path
approach depending firstly on
channel noise and
then on the
communicating
content
4 Singh, P. D. et al. 1. Network delay, Ensemble-based 1. Combination of
[30], 2021 RAM and network classifier AI and fog
usages computing for smart
2. Aaccuracy, health enabling early
precision, recall, covid-19 detection
kappa static and root 2. Advantage of
mean square error blockchain is
utilized like shared
storage, peer-to-peer
connection and time
stamping
3. Better
classification
technique as per the
results
(continued)
 AI Enabled Internet of Medical Things 43

Table 4. (continued)

Sr.No Authors & Parameters QoS Technique Advantages

publication years
5 Khodkari, H., QoS measures, Multi attribute 1. Due to the
et al. [31], 2018 sensor data, big decision making integration of cloud
data, IoT protocols algorithm and IoT, most of the
and Cloud interface recent developers of
parameters this era shows their
interest in this
system
2. The IoT
paradigms improves
QoS of the system
6 Sodhro, A. H., et. 1. Transmission Adaptive QoS High visualization,
al. [32]. 2021 power, route Computation battery lifetime and
selection and duty algorithm (AQCA) power optimization
cycle
2. AQCA and
Quality of
experience (QoE)
used for QoS
measurement

5 Conclusion
Artificial Intelligence (AI) gives automation to the system to increase ease and simplicity
within a system. Al can be used in the Internet of Medical Things (IoMT) framework to
incorporate automation in the healthcare sector which is also known as Smart Healthcare.
As per the current scenario, every day new harmful viruses are coming leading to severe
health degradation. A person visiting hospitals or other healthcare for every small health
problem may get infected by these harmful viruses and therefore Smart Healthcare
System (SHS) is introduced. In SHS without visiting the hospital a person can measure his
regular health records using wearable/implanted smart devices and can take preliminary
precautions on their own. A doctor can also monitor his patient’s health from remote
areas using SHS applications. Based on the patient’s health scenario SHS can be divided
into three domains as discussed in this work. AI-based SHS using the IoMT framework
usually have three layers in their architecture namely the bottommost sensor level for
collection of medical data by means of wearable/ implanted smart devices, the middle
layer used for processing of these data with a set of algorithms, and the topmost layer used
for the end-user interface. This work gives an explanative analysis of different AI-based
SHS architectures proposed by various authors based on their system requirements.
SHS deals with highly sensitive medical records of various patients and also requires
quick responses from the system therefore few research challenges need to be defined
and tackled effectively. A comparative analysis of these research challenges namely
44 J. Srivastava and S. Routray

Data security, Data accuracy, System efficiency, Quality of Service (QoS), and System
reliability has been discussed in this paper.

References
1. Sun, Y., Lo, F.P.W., Lo, B.: Security and privacy for the internet of medical things enabled
healthcare systems: a survey. IEEE Access 7, 183339–183355 (2019)
2. Chen, S., Xu, H., Liu, D., Hu, B., Wang, H.: A vision of IoT: applications, challenges, and
opportunities with china perspective. IEEE Internet Things J. 1(4), 349–359 (2014)
3. Sinha, B.B., Dhanalakshmi, R.: Recent advancements and challenges of Internet of Things
in smart agriculture: a survey. Future Gener. Comput. Syst. 126, 169–184 (2022)
4. Al-Emran, M., Malik, S.I., Al-Kabi, M.N.: A survey of Internet of Things (IoT) in education:
opportunities and challenges. In: Hassanien, A.E., Bhatnagar, R., Nour Eldeen, M., Khal-
ifa, M.H., Taha, N. (eds.) Toward Social Internet of Things (SIoT): Enabling Technologies,
Architectures and Applications. SCI, vol. 846, pp. 197–209. Springer, Cham (2020). https://
doi.org/10.1007/978-3-030-24513-9_12
5. Elkin, D., Vyatkin, V.: IoT in traffic management: review of existing methods of road traffic
regulation. In: Silhavy, R. (ed.) Applied Informatics and Cybernetics in Intelligent Systems:
Proceedings of the 9th Computer Science On-line Conference 2020, vol. 3, pp. 536–551.
Springer International Publishing, Cham (2020). https://doi.org/10.1007/978-3-030-51974-
2_50
6. Routray, S., Mao, Q.: A context aware-based deep neural network approach for simultaneous
speech denoising and dereverberation. Neural Comput. Appl. 34(12), 9831–9845 (2022)
7. Mohamed Shakeel, P., Baskar, S., Sarma Dhulipala, V.R., Mishra, S., Jaber, M.M.: Maintain-
ing security and privacy in health care system using learning based deep-Q-networks. J. Med.
Syst. 42(10), 1–10 (2018)
8. Alsubaei, F., Abuhussein, A., Shandilya, V., Shiva, S.: IoMT-SAF: internet of medical things
security assessment framework. Internet of Things 8, 100123 (2019)
9. Routray, S., Mao, Q.: Phase sensitive masking-based single channel speech enhancement
using conditional generative adversarial network. Comput. Speech Lang. 71, 101270 (2022)
10. Algarni, A.: A survey and classification of security and privacy research in smart healthcare
systems. IEEE Access 7, 101879–101894 (2019)
11. Padhy, S., Dash, S., Routray, S., Ahmad, S., Nazeer, J., Alam, A.: IoT-based hybrid ensemble
machine learning model for efficient diabetes mellitus prediction. Compu. Intell. Neurosci.
2022, 1–11 (2022). https://doi.org/10.1155/2022/2389636
12. Kumar, N.: IoT architecture and system design for healthcare systems. In: 2017 International
Conference on Smart Technologies for Smart Nation (SmartTechCon), pp. 1118–1123. IEEE
(2017)
13. Sun, L., Jiang, X., Ren, H., Guo, Y.: Edge-cloud computing and artificial intelligence in
internet of medical things: architecture, technology and application. IEEE Access 8, 101079–
101092 (2020)
14. Tekieh, M.H., Rashemi, B.: Importance of data mining in healthcare. In: Proceedings of
2015 IEEE/ACM international conference advances in social networks analysis and mining
2015—ASONAM’15, pp 1057–1062 (2015)
15. Shahin, A., Moudani, W., Chakik, F., Khalil, M.: Data mining in healthcare informa-
tion systems case studies in Northern Lebanon. In: 2014 Third International Conference
e-Technologies Networks Devices, pp 151–155 (2014)
16. Yang, L., Li, Z., Luo, G.: MH-Arm: a multimode and high-value association rule mining
technique for healthcare data analysis. In: 2016 International Conference on Computational
Science and Computational Intelligence, no. 71432002, pp. 122–127 (2016)
 AI Enabled Internet of Medical Things 45

17. Mdaghri, Z.A., El Yadari, M., Benyoussef, A., El Kenz, A.: Study and analysis of data mining
for healthcare. In: 2016 4th IEEE International Colloquium on Information Science and
Technology (CiSt), pp. 77–82 (2016)
18. Roy, S., Mondal, S., Ekbal, A., Desarkar, M.S.: CRDT: correlation ratio based decision
tree model for healthcare data mining. In: 2016 IEEE 16th International Conference on
Bioinformatics, Bioengineering, pp. 36–43 (2016)
19. Rao, A.R., Clarke, D.: A fully Integrated open-source toolkit for mining healthcare bigdata:
architecture and applications. In: Proceedings on 2016 IEEE international conference on
healthcare informatics, ICHI 2016, pp. 255–261 (2016)
20. Moosavi, S.R., et al.: SEA: A secure and efficient authentication and authorization architecture
for IoT-based healthcare using smart gateways. Procedia Comput. Sci. 52(1), 452–459 (2015)
21. Verma, G., Prakash, S.: Internet of Things for healthcare: research challenges and future
prospects. In: Hura, G.S., Singh, A.K., Hoe, L.S. (eds.) Advances in Communication and
Computational Technology. LNEE, vol. 668, pp. 1055–1067. Springer, Singapore (2021).
https://doi.org/10.1007/978-981-15-5341-7_80
22. Rehman, A., Saba, T., Haseeb, K., Larabi Marie-Sainte, S., Lloret, J.: Energy-efficient IoT e-
health using artificial intelligence model with homomorphic secret sharing. Energies 14(19),
6414 (2021)
23. Sodhro, A.H., et al.: Decentralized energy efficient model for data transmission in IoT-based
healthcare system. In: 2021 IEEE 93rd Vehicular Technology Conference (VTC2021-Spring),
pp. 1–5. IEEE (2021)
24. Lazarevska, M., Farahbakhsh, R., Shakya, N.M., Crespi, N.: Mobility supported energy effi-
cient routing protocol for IoT based healthcare applications. In: 2018 IEEE Conference on
Standards for Communications and Networking (CSCN), pp. 1–5. IEEE (2018)
25. Saba, T., Haseeb, K., Ahmed, I., Rehman, A.: Secure and energy-efficient framework using
Internet of Medical Things for e-healthcare. J. Infect. Public Health 13(10), 1567–1575 (2020)
26. AbdulmohsinHammood, D., Rahim, H.A., Alkhayyat, A., Ahmad, R.B.: Body-to-body coop-
eration in internet of medical things: toward energy efficiency improvement. Future Internet
11(11), 239 (2019)
27. Kumar, A., Joshi, S.: Applications of AI in healthcare sector for enhancement of medical
decision making and quality of service. In: 2022 International Conference on Decision Aid
Sciences and Applications (DASA), pp. 37–41. IEEE (2022)
28. Patan, R., Ghantasala, G.P., Sekaran, R., Gupta, D., Ramachandran, M.: Smart healthcare and
quality of service in IoT using grey filter convolutional based cyber physical system. Sustain.
Cities Soc. 59, 102141 (2020)
29. Agnihotri, S., Ramkumar, K.R.: Content based routing algorithm to improve QoS in IoMT
networks. In: Saha, A., Kar, N., Deb, S. (eds.) Advances in Computational Intelligence,
Security and Internet of Things: Second International Conference, ICCISIoT 2019, Agartala,
India, December 13–14, 2019, Proceedings, pp. 195–206. Springer Singapore, Singapore
(2020). https://doi.org/10.1007/978-981-15-3666-3_17
30. Singh, P.D., Kaur, R., Dhiman, G., Bojja, G.R.: BOSS: A new QoS aware blockchain assisted
framework for secure and smart healthcare as a service. Expert Syst. e12838 (2021)
31. Khodkari, H., Maghrebi, S.G., Asosheh, A., Hosseinzadeh, M.: Smart healthcare and quality
of service challenges. In: 2018 9th International Symposium on Telecommunications (IST),
pp. 253–257. IEEE (2018)
32. Sodhro, A.H., Malokani, A.S., Sodhro, G.H., Muzammal, M., Zongwei, L.: An adaptive
QoS computation for medical data processing in intelligent healthcare applications. Neural
Comput. Appl. 32(3), 723–734 (2019). https://doi.org/10.1007/s00521-018-3931-1
46 J. Srivastava and S. Routray

33. Srivastava, J., Routray, S., Ahmad, S., Waris, M.M.: Internet of Medical Things (IoMT)-based
smart healthcare system: Trends and progress. Comput. Intell. Neurosci. 2022, 1–17 (2022).
https://doi.org/10.1155/2022/7218113
34. Malla, P.P., Sahu, S., Routray, S.: Investigation of breast tumor detection using microwave
imaging technique. In 2020 International Conference on Computer Communication and
Informatics (ICCCI), pp. 1–4. IEEE (2020)
 Metaverse and Posthuman Animated Avatars
for Teaching-Learning Process: Interperception
in Virtual Universe for Educational
Transformation

Ashraf Alam(B) and Atasi Mohanty

Rekhi Centre of Excellence for the Science of Happiness, Indian Institute of Technology
Kharagpur, Kharagpur, West Bengal, India
[email protected]

Abstract. It has been quite some time that the attention of teachers has been
focused on the incorporation of Metaverse in education. Since Facebook’s
announcement that it will be rebranding and promoting itself as Meta, there has
been much interest in this subject. To the best of our knowledge, no research
as of date has systematically synthesised the results associated with the imple-
mentation of Metaverse in education, even though several studies have conducted
literature reviews on Metaverse in general. This research does a detailed litera-
ture assessment on how the Metaverse might be used in the classroom in order
to fill this knowledge gap. Research trends, focal points, and limitations of this
‘study subject’ are revealed using both bibliometric analysis and text analysis.
It is evident from the collected data that lifelogging for educational metaverse
merits additional research. The results indicate that research into mobile-, hybrid-
, and micro-learning environments are still in its infancy. There is no study on
the effectiveness of utilising metaverse to educate children with special needs.
In order to widen the attraction of metaverse to educators throughout the globe
and to improve their capacity to assist successful learning and teaching in the vir-
tual environment, more in-depth research shall be carried out along the proposed
roadmap as suggested by the results of the current investigation.

Keywords: Metaverse · Virtual avatars · Review article · Education ·

Curriculum · Pedagogy · Teaching-learning · E-learning · Educational
technology · Schools · Universities

1 Introduction
The inkling of Metaverse is not very new; it has been explored in works of science fiction
such as Snow Crash and has recently received increased attention with the release of
the film which is an adaptation of the book, Ready Player One [1, 31]. Second Life and
World of Warcraft were already out there as well-known and well-liked examples [2, 7].
Mark Zuckerberg’s public unveiling of Metaverse in October 2021, however, sparked a
surge in the platform’s popularity [3, 46]. Many instructors and researchers have begun

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 47–61, 2022.
https://doi.org/10.1007/978-3-031-23233-6_4
48 A. Alam and A. Mohanty

outlining potential future agendas and implementation scenarios in their teaching and
research methods [4, 19]. The virtual environment, which depicts people as realistically
as possible, may enhance the interpersonal aspects of education, contributing to the
rising popularity of online courses [5]. Since ‘Metaverse’ is still a relatively new term,
the current investigation addresses the need to evaluate the current state of research on
the topic [6].
Avatars, which are digital representations of real-world humans, administer the econ-
omy and conduct daily activities in the 3D virtual environment known as the Metaverse
[7]. Instead of being a platform built by a single corporation, which would inevitably
lead to monopolisation, the word ‘Metaverse’ describes a whole ecosystem. It denotes
a whole new dimension of life that is not just beyond the purview of any nation’s or
company’s governing bodies, but of all governments altogether [8]. Metaverse offers
opportunities for immersion, collaboration, and involvement, all of which contribute to
the expansion of social experiences and the birth of ‘parallel worlds’ [9]. The growth of
the metaverse requires the following three stages: (1) digital twins, which permit simu-
lating and modelling the physical environment digitally [10]. Numerous academics and
thinkers have already explored the educational potential of the Metaverse [11]. The inte-
gration of Metaverse with Second Life as a learning management system was previously
explored in the context of enhancing educational outcomes [12]. The Metaverse, which
emphasises the virtuality component, might be the next place where people congregate
and form social connections; as such, institutions of higher learning should take the
initiative to use it for educational purposes [13].
Mirror worlds, lifelogging, AR, and virtual worlds are the four branches of the
metaverse technology tree [14]. Through the use of digital data superimposed on top
of the user’s perception of the actual world, augmentation technology enhances the
environment with a new visual element [15]. In contrast, simulation technology creates
and alters models of the actual world in order to give virtual interactions and experiences
[16]. The other group focuses on the interplay between the internal and exterior spheres
[17]. Information regarding the user’s external environment and methods for influencing
it are presented to emphasise the importance of this aspect of the technology [18]. When
these two dimensions are combined, four distinct variations of the Metaverse emerge
[19]. Thanks to the technologies present in the Augmented Reality Metaverse, we now
design smart environments dependent on location networks, like Pokémon Go [20].
Lifelogging Metaverse collects daily details about people and things using augmented
reality applications like Facebook and Instagram [21]. The technique generates virtual
maps and models in the Metaverse of Mirror Worlds using GPS data from programmes
like Google Earth and Google Maps [22]. Digitally interactive avatars representing a
wide variety of identities form the basis of the technology powering the Metaverse of
Virtual Worlds [23].
González Crespo and colleagues utilised OpenSim to research the potential of virtual
worlds as a platform for knowledge exchange and education [4, 9, 13, 24]. After seeing
the potential of augmented reality and mobile education for the classroom, Reyes and
colleagues developed the Metaverse for the purpose of teaching mathematics [24]. The
findings suggested that incorporating Metaverse into mathematics education has the
potential to boost students’ achievement [25]. Park and Kim also categorised the many
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 49

sorts of instructional Metaverse worlds with different genres such as survival, labyrinth,
multi-choice, racing/jump, and escape room [8, 12, 17, 22].

2 Objectives of the Research and Knowledge Gap

What we know about the Metaverse now is based on the widespread perception among
members of Generation Z that their digital persona is identical to their ideal one. In other
words, people believe that their online digital identities are reflected and represented in
their offline physical selves. The researchers argue that the development and influence
of Generation Z need a new definition of the term ‘Metaverse’, as it has evolved from
its previous iteration. Furthermore, the speedy growth of deep learning and mobile tech-
nology has made the Metaverse more accessible from any place at any time, improved
the precision of visual and linguistic recognition, and created more immersive environ-
ments than ever before. Thus, it is crucial to examine the evolution of the Metaverse in
teaching, the structure of the Metaverse, and the current trends in historical research.
Many questions, such as what sort of Metaverse is utilised in education, what kinds of
learning circumstances are used, and what kinds of assessment methodologies are used,
remain unanswered. This study employs bibliometric and content analysis to provide a
thorough assessment of the literature about the Metaverse’s use in the classroom, filling a
need that has so far gone unfilled. Content analysis was used to conduct a comprehensive
review of the papers that were analysed in order to determine the research subjects that
researchers choose to emphasise while discussing the educational applications of the
Metaverse. The following research questions are addressed in this investigation:

1. What is the educational trend in the Metaverse?

2. What sorts of metaverses are employed in classrooms?
3. Where and at what grade level can these learning scenarios be used?
4. What technologies have been used, and how does the Metaverse represent the pupils’
digital identities?
5. In what ways Metaverse is used to effectively educate students?
6. What concerns arise from the Metaverse’s potential influence on schooling?

3 Methods and Methodology

The current investigation uses qualitative and quantitative synthesis techniques to inves-
tigate the literature on teaching using the Metaverse. A human evaluation is more likely
to introduce outcome reporting bias, and results that are obtained from interpretation
are likely to be subjective, than an automated review. Therefore, a mixed-method review
including content analysis and bibliometric analysis is necessary to systematically rec-
ognize and classify the evolution and knowledge base of a subject matter. PRISMA
guidelines were followed in this systematic literature review.
The databases Web of Science (WOS) and Scopus were combed using the search
phrase “education and metaverse” to locate scholarly articles. All scholarly articles
published as of September 25, 2022, are included in this analysis. The research articles
were not considered if they did not meet all three of the following criteria: (1) be written
50 A. Alam and A. Mohanty

in English, (2) deal with Metaverse in general, and (3) be available online. Consequently,
21 articles were in the Scopus database, and 36 publications were located in the WOS
database.
The research used both bibliometric analysis and content analysis. The evaluation
and interpretation of the data employed data triangulation techniques to provide a more
holistic perspective and strengthen the reliability of the study’s findings. The VOSviewer
software was used to do the bibliometric analysis and synthesis by categorising and
mapping the phrases taken from the keywords, titles, and abstracts into a similarity
matrix based on their level of relatedness.

4 Results and Discussion

The results are discussed in light of each of the aforementioned research questions.

4.1 Educational Metaverse: A Categorical Analysis

The first academic investigations towards Metaverse implementation in teaching started

in 2007. Research in the WOS database increased after 2008, with the number of research
reaching a peak of five per year in 2009, 2010, and 2013. There is a precipitous fall
beginning in 2013, with no new studies being uncovered in 2014. In 2015, the highest
number of studies ever recorded occurred, with six in total. From 2015 to 2019, there
was a precipitous drop in the number of published study findings, and in 2019, no new
studies were conducted at all. In 2020 and 2021, researchers conducted an average of
four investigations each year. Using regression analysis, we found that the Metaverse’s
popularity in the WOS database fluctuated from 2007 to 2021, but has generally been
on the increase in recent years. The highest number of studies that were published in
the Scopus database occurred in the year 2009, with a total of four articles published
in 2001 and 2013. After 2013, the number of research published annually decreased.
There was a notable increase in the number of scientific publications published from
four in 2019 to six in 2020. The year 2021 saw a decline in the number of studies, with
just three being uncovered. Recent years, especially after Covid-19 pandemic, have seen
increased metaverse research activity in both databases, perhaps as a consequence of the
widespread use of simulated worlds. The first wave of Metaverse study, which lasted
from 2007 to 2013, has been compared to both Web 2.0 and to earlier examples like
Second Life. The investments made in Metaverse technology explain the sudden surge
in 2021, which marks the beginning of the third wave (2021 onwards). Web 3.0 and
emerging technologies like augmented and virtual reality (AR/VR), as well as the rise in
processing data and the ability to display virtual worlds, are responsible for the second
wave (2014–2020) (e.g., Facebook).
When it comes to publication type, conference proceedings constitute the bulk of
publications discussing the spread of Metaverse in educational research. There are twice
as many studies of the article type in the WOS database than there are in the Sco-
pus database. The search rates for books and scholarly journals are equivalent in both
databases. Both databases show a preponderance of conference papers as the primary
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 51

source for Metaverse research. Shorter publication schedules of conference proceedings

explain the popularity of conference papers.
There was an assumption that the author’s nationality revealed where the concept of
the Metaverse first emerged inside the realm of educational research. The United States,
Brazil, Japan, Spain, and South Korea led the world in research activity. It is worth noting
that there has not been a research done on the efficacy of the Metaverse as a teaching
tool in either the Arab world or Africa. This may be because these countries’ inadequate
infrastructure hinders them from introducing such technologically advanced educational
environments. This raises the question of whether or not this kind of classroom really
deepens the digital divide rather than helping to bridge it via offering an accessible
education for all students. The potential of openness and open educational resources
must be continuously used to keep the Metaverse of education welcoming and accessible
to everyone.
Since the study of the metaverse is still in its infancy, almost half (41.7%) of the
studies just summarised previous works and expounded on theories without actually con-
ducting any tests. Twenty-eight percent of the studies used a hybrid design, followed by
eighteen percent each that used quantitative and qualitative methodologies. The research
found that the two most often cited types of data collecting were interviews and surveys.
This is largely attributable to the speed with which information may be gathered via
participant interviews and surveys. Log data may also document how students engaged
with the Metaverse settings and progressed, offering better insights throughout the whole
learning process compared to surveys or interviews alone.

4.2 A Wide Variety of Virtual Worlds for Use in Education

Most studies included sophisticated modelling and close interaction tools, suggesting
a preference for the Virtual Worlds (VW) Metaverse [26]. VW employs a system that
makes use of 3D technology to virtually portray elaborate computer graphics works
[27]. VW articles used language and translation grids as the basis for a virtual reality
debate and communication platform, among other educational elements from the VW
Metaverse category [28]. Because they include virtual simulations of environments that
are either expensive to create (like aircraft simulators for student training) or challenging
(like learning about nuclear energy and safety) for students (because of high risk), VW
Metaverse types have proven to have useful educational implications [29].
Augmented reality (AR) is the process of superimposing computer-generated
imagery (CGI) onto a user’s view of the real world to create a more lifelike, three-
dimensional effect [30]. One such use is the transformation of a real-world book page
into something like a 3D movie 31]. Examined studies included those that used Microsoft
HoloLens to instruct students in aircraft maintenance or those that used augmented real-
ity projects including the scanning of QR codes to get students interested in the material
[31]. The articles as a whole did not make effective use of AR Metaverse’s educational
potential [32]. As an example, none of the publications we read advocated for the use
of simulated, digitally presented three dimensions to educate students about intangible
concepts [33].
According to the literature we reviewed, Lifelogging and Mirror Worlds are the
two Metaverse categories that are used least often in the classroom [34]. Lifelogging
52 A. Alam and A. Mohanty

Metaverse is the practise of using social media for learning purposes in order to record,
collate, and analyse one’s day-to-day activities, thoughts, and relationships with others
[35]. Lifelogging metaverse research is a kind of study that combines augmented reality
with personal communication via voice interaction and recognition [36]. Because of its
potential instructional value in enhancing one’s ability to represent and use knowledge
appropriately in response to criticism from other users in the network, this fits the criteria
for a Lifelogging Metaverse [37]. Lifelogging students can use their critical thinking and
creative imagination to explore the site’s diverse data and reconstruct knowledge through
the power of the group [36].
The final type of Metaverse is called Mirror Worlds (MW), and it is used to over-
come geographical and physical barriers to education by extending real-world settings
through networking technologies and GPS [38]. Only one study was classified as a MW
Metaverse, and that was because it used a game-based approach to immersive learning
by gathering students in a traditional classroom and mirroring the physical space onto
an online platform [39]. The research did not make full use of the MW Metaverse, while
reflecting what Kye termed as ‘efficient expansion’ approach for modelling the physical
world [3, 7, 16, 21]. For example, MW users may work together on significant projects
and even play games with others who are thousands of miles away from them [40].
The bulk of studies in the field of education have focused on Virtual Worlds Meta-
verse, while just a small number have made use of Augmented Reality Metaverse, and
an even smaller number have made use of Lifelogging and Mirror Worlds Metaverses
[41]. The outcomes are discussed in this article [42]. Nonetheless, 3D technologies were
employed in the papers to either use or explain virtual worlds [43]. However, as the
Metaverse road map indicates, the technology was not completely used or explained
in the reviewed publications [44]. Lifelogging may give novel forms of data that may
be analysed to probe hitherto unexplored areas of blending psychology and educational
technology [45]. Those who have developed advanced digital skills may also benefit
from using MW technology, making them worthy of special attention to the ways in
which they learn in the MW Metaverse [46].

4.3 Situations for Learning, Tiers of Education, and VR Learning Environments

Fifty-three percent of Metaverse researches were in the fields of natural science, math-
ematics, and engineering, while fifteen percent were in the realms of general education
and eleven percent were in the arts and humanities [47]. The Metaverse may provide
technical assistance to the STEM disciplines by way of 3D modelling software for use
in the classroom, by helping students draw parallels between real-world experiments
and virtual things, and by offering data-mining-driven, autonomous tutoring systems
[48]. For these and other reasons, several disciplines might benefit from integrating
Metaverse into their work [49]. For this reason, the arts and humanities have made sub-
stantial use of Metaverse for language learning, since it may allow users to communicate
with speakers of various languages [17]. Metaverse has the potential to merge the virtual
world with traditional classrooms, therefore fostering novel avenues for interdisciplinary
and problem-based learning [23]. Only 6% of social scientists use Metaverse, which is
notable [29]. In reality, the Metaverse may also be quite useful in these areas [36]. For
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 53

instance, students may form virtual networks in the subject of archaeology via the use
of the Metaverse, which can be accessed during online e-learning sessions [38].
It was found that 62.9% of Metaverse research was conducted involving students
who were at the university level [41]. The findings prove that Metaverse can be used
in higher education to give students and faculty access to interactive and immersive
experiences that open the door to exploring novel pedagogical practises, ICTs, and
emerging technologies [46]. From this vantage point, it appears that the Metaverse may
help alleviate the limitations and inefficiencies of traditional 2D online education [49].
However, there has only been a modest amount of research into the Metaverse in primary,
secondary, and tertiary institutions [7]. No studies were found that specifically addressed
disabled students, so more studies are needed to determine how to create inclusive
and accessible Metaverse in educational environments [12]. The emergence of virtual
freedom in space and time may allow students with impairments and special needs to
engage more fully [18].
Metaverse provides several opportunities for the development of educational situa-
tions [25]. The findings showcase nine different pedagogical settings where the Meta-
verse is used for learning [28]. Online education received the most study (31.3%), fol-
lowed by problem-based education, gaming-based education, collaborative education,
and project-based learning (PBL) [31]. Metaverse provide pupils with a more engaging
and interactive setting, as well as the opportunity for more collaborative study [35]. By
analysing the Metaverse’s applications in a variety of contexts, it may help both edu-
cators and students improve their use of the platform for instruction [38]. This means
that proposals for innovative methods of instruction and evaluation will be made, and
that plans for their use in virtual reality settings will be explored [41]. How can inno-
vative tools like eye-tracking and voice-recognition software be incorporated into the
instructional design process?
When it comes to online education, the Metaverse is where it is at. This is largely due
to the fact that students can engage with various digital resources through virtual worlds,
made possible by the Metaverse and high-performance servers [45]. Meanwhile, collab-
orative learning is frequently combined with virtual learning scenarios [48]. Through the
Metaverse’s built-in social networks, students can communicate with one another and
exchange course-related knowledge to promote cooperative learning [3]. Metaverse’s
simulated environment is ubiquitous in modern blended-learning setups [7]. The results
demonstrate the usefulness of integrating real-world experiments involving the Meta-
verse and virtual systems with online lectures and tutorials [8]. According to Kanematsu’s
study, using Metaverse (Second Life) for virtual course lectures is a viable option for
students enrolled in science, technology, engineering, and mathematics (STEM) pro-
grammes [5, 7, 11, 19]. In this way, the teacher can guide the students through the
STEM curriculum experiment in the physical classroom while also providing them with
support in the Metaverse [13].
Using the same principles as game-based education, the Metaverse offers a virtual,
engaging environment for learning [11]. Getchell and colleagues showed how Metaverse
pioneers new-fangled possibilities for game-based teaching-learning by letting teachers
develop flexible game-based learning environments and providing students more control
over their education at a cheaper price [2, 6, 18, 25]. Estudante and Dietrich, in their study,
54 A. Alam and A. Mohanty

propose utilising the free software Metaverse to create a VR game for mobile devices
[3, 8, 15, 19]. The students are taught to think like physicists while solving issues [19].
This game may help students get a deeper understanding of chemistry concepts like the
periodic table, chemical equilibrium, and molar mass while having fun at the same time
[22]. As a result, the Metaverse’s platform might be utilised to improve students’ drive
to study and their capacity for collaboration via the use of game-based instruction [28].
Students are given brief lectures and faced with topics connected to nuclear power,
and they actively engage in problem-solving via Metaverse chat sessions led by their
lecturers [33]. According to the findings, including Metaverse in PBL courses has the
potential to raise interest, stimulate debate, and improve students’ level of understanding
[36]. As a matter of fact, by using the Metaverse, these learning scenarios may pique
the interest of both students and instructors while providing the optimal environment
for their teaching and learning activities [41]. The Metaverse’s built virtual world is
able to change the static conventional teaching paradigm into a dynamic one in these
different learning settings since it provides learning materials and timely evaluations
[46]. Collaboration among students is made possible by this [49].

4.4 Students’ Avatars (Digital Personas) in the Metaverse

It is believed that as students build their digital identities in metaverse settings, they will
have a greater social presence, or be seen as more authentic by their peers [1]. ‘Digi-
tal Identity’ is characterised by student’s choice or building of avatars and interaction
patterns as the self-image or inner goals [6]. In the Metaverse, avatars may be either
immersive, present, or representative [8]. Avatars were a way for students to share their
digital selves in the Metaverse and were allowed in the evaluated papers [12]. Partic-
ularly, a few studies allowed participants to choose an avatar from a pool of pre-made
ones [16]. Students may create and customise their own avatars in several other classes
[19]. From this, we might infer that pupils tailor their avatars to reflect their own qualities
[21]. According to the findings from González Crespo’s research, students may express
their individuality and taste via the creation of customised digital avatars [6, 14, 19, 23].
In order to engage with the virtual environment, students may draw upon a wide range of
skills, such as walking, flying, purchasing products, and personalising their appearance
[25]. In addition, the Metaverse’s social chatting feature allows students to successfully
convey their ideas to one another via the use of both overt and covert references to real-
world objects, locations, and emotions [29]. Thanks to the digital avatar’s integration of
real-world objects, characters, and settings, the user is immersed in a highly realistic,
three-dimensional simulation [34].
Metaverse characters often have the ability to blink, which encourages rigorous
coursework and more student participation [36]. Video games are another popular
medium where digital avatars appear [41]. Most research shows that students use their
digital avatars to participate in class discussions and form relationships with their peers
in Second Life [45]. Students who spend significant time in the virtual world of Sec-
ond Life will be able to create digital avatars that take on a variety of jobs, including
socializers, and so expand their social circles [49]. In addition, future research might
investigate how students of different cultures use avatars [3]. Improving the educational
Metaverse’s gaming options is a top priority [8].
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 55

The Metaverse provides a wealth of pedagogical and technical resources for educa-
tors, allowing students to participate in immersive learning that has been shown to boost
engagement and motivation [11]. The seven types of technological resources are as
follows: wearable, immersive, educational, modelling and simulation, gaming applica-
tions, artificial intelligence, mobiles, and sensors [15]. The Metaverse is used to provide
students with hands-on experience, which not only encourages collaboration and skill
development but also keeps them interested and involved in what they are learning [12].
Virtual reality, augmented reality, mixed reality, and other similar virtual technologies
must be combined to provide a truly immersive experience [27]. The benefit of multi-
modal immersion is further shown by the emergence of technologies that open portals to
and transport us inside Metaverse realms [36]. The aforementioned four technologies are
the most widely used immersive interfaces in the Metaverse, and they have the potential
to enhance classroom instruction by immersing pupils psychologically and so promoting
transferable skills [46]. In their study, Siyaev and Jo discuss how MR may be used in the
Metaverse to enhance learning via the use of deep learning voice interaction modules,
bringing together the real and virtual worlds [5, 9, 11, 18]. To help students form more
meaningful relationships with the digital world, MR may focus largely on the voice
interaction that happens during the learning process [11]. To further facilitate students’
access to immersive learning, VR allows for the administration of virtual worlds and the
development of shareable avatars [21]. Students in the Metaverse may control their own
avatars in line with the shown environment, make social connections with other students,
and create custom avatars utilising virtual reality to provide an immersive experience
[34].
Games are another popular kind of application utilised in the Metaverse to pro-
vide meaningful learning opportunities [38]. In terms of gaming apps, Pokémon Go has
been the most popular recently [27]. As a result of advancements in real-time virtual
reality and augmented reality, fictitious interactive 3D avatars can now be created and
used to lure Pokémon inside games [19]. When the Metaverse is employed in the class-
room, the immersive surroundings may play a crucial role in teaching various subjects
and make it easier to draw connections between different areas of knowledge [24]. If
blended with learning management systems [30]. Rapanotti and Hall have integrated
the Metaverse with the Second Life platform to provide a more immersive virtual world
platform for higher education [2, 9, 17, 24]. Using the resources supplied by Second
Life, students may construct a 3D virtual avatar, resulting in an immersive learning
experience [41]. Students fashion their own digital avatars in Second Life and engage in
virtual interactions with other students and can spend virtual cash to buy or manufacture
the resources [44]. Together, the institution’s LMS, HotPotatoes, Massive Open Online
Course (MOOC), Moodle, Teleduc, Eduquito, and Sloodle form a hybrid learning plat-
form in the Metaverse for teaching and learning [48]. The Metaverse and MOOCs have
made it possible for numerous students to have free and simultaneous access to a wealth
of topical material to expand their understanding [43]. The use of virtual labs is common
in the teaching of natural sciences, mathematics, and engineering [4]. By providing a
dynamic, collaborative, and interactive learning environment, VLL increases students’
56 A. Alam and A. Mohanty

motivation to learn and value their education [8]. Moodle, a modern learning manage-
ment system, may improve upon traditional methods of distributing course materials
and encourage student interaction [45].
Estudante and Dietrich developed software for Apple and Google smartphones that
may be used to construct an augmented reality version of the Metaverse [21]. The
OpenSim platform may be used with geographic mobility to provide content tailored to
each organization’s needs and methods of operation [24]. Connectivity to the internet
and data sharing in public virtual worlds are also made feasible by geospatial mobility
[5]. Metaverse on mobile devices might help students study if they use their avatars [47].
Another popular resource in the Metaverse is the Blinking system, which keeps track of
students’ blink rates using specialised software [13]. When a student’s emotions are up
in the air, the blinking mechanism causes them to blink more often, which helps teachers
decipher their responses [19].
AI provides a foundation for the Metaverse’s central concept—the analysis of its
complex data for interpretation, supervision, control, and planning [13]. Neuro-symbolic
AI might fill the role of subject-matter experts in aviation maintenance courses, provid-
ing guidance on technical matters and providing access to all the materials needed for
effective training and teaching [19]. To improve learning efficiency, convolutional neural
networks are increasingly being used to process not just visual and textual data, but also
audio data, such as commands and language recognition [31]. To help users navigate
virtual spaces, Web 3.0 often uses a combination of machine learning and semantic
database modelling [41]. The AI possibilities of the Metaverse allow for the creation of
new roles for intelligent NPCs to play as mentors, peers, and mentees [37].
Many different types of technology have been used in the creation of a thriving
ecosystem in Metaverse, as has previously been mentioned [11]. There are, however,
a number of state-of-the-art technologies that are not being used [13]. As one possible
solution to the problems of cheating and insecure user data, a blockchain-based educa-
tional system is being created [25]. It is also feasible to wonder whether the ICT-based
competencies mentioned in the literature are sufficient to prepare students and teachers
for this new educational environment (‘Metaverse in Education’), or if additional abilities
are necessary for enhanced teaching and learning [42]. Along with the new learning pos-
sibilities presented by technological advancements, users may be vulnerable to threats
[10]. Privacy may be compromised by sensors designed to read pupils’ emotions and
activities [26]. The potential for privacy breaches also increases in a digital classroom
where physical items may be used to track user activity [34]. These risks should be taken
into account by educators and researchers when they design Metaverse applications for
use in the classroom [39].

4.5 Alterations in Educational Multiverse

One of the primary challenges with using Metaverse in the classroom is the technology
and tactics used for immersion [12]. There are likely many classroom settings in which
the Metaverse may be helpful [8]. For instance, the Metaverse’s escape game built on
the VR platform might be a useful tool for facilitating mobile learning in the context
of a game-based classroom [9]. The findings show that students are more invested in
their education when using games in the Metaverse as a means of instruction, and that
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 57

they are eager to use smart gadgets for scientific practise [23]. Teaching in a metaverse
environment, such as Second Life, is also effective [37]. It has the potential to enhance
student learning by encouraging cross-lingual interaction and discussion [28].
The Metaverse is helpful for students in a wide variety of disciplines. The stud-
ies show that the Metaverse may help students connect, become more motivated and
engaged, and broaden the possibilities of learning [14]. When combined, these features
expand the educational potential of the Metaverse considerably [8]. Students majoring
in aviation maintenance now have an affordable online alternative to flying, thanks to
metaverse, which not only allows them to engage with one another but also to undertake
fake aircraft repair [3]. The translation system will also allow for more direct interactions
between students, adding to the advantages of the metaverse for language instruction,
and the language grid system can be integrated with Second Life [15]. Finally, students
in a Metaverse-powered classroom may have trouble distinguishing fiction from fact
[13]. Therefore, it is reasonable to assume that the students in the class will form a new
set of social networks for learning, resulting in novel scholastic opportunities [29].
Various pedagogical, technological, and other types of challenges have been associ-
ated with using the Metaverse in the classroom [22]. Technically speaking, the biggest
factor was network congestion (21.1%), followed by smartphone interface design prob-
lems (8.2%) and blink capture problems (5.9%). Getchell emphasised the significance
of punctuality in network communications and the increased demands they place on the
host server and the overall network infrastructure [8, 11, 17, 26]. Because of the current
state of the network connection size, the results of student evaluations may be skewed.
There is further evidence from studies conducted on smartphones that the Metaverse has
an interface problem that is independent of the Metaverse app [5, 9, 16, 23]. It has been
suggested that if too few children use smartphones at once, this might have a negative
impact on their ability to work together and communicate [32]. This is due to the small
size of smartphone screens [40].
Dáz argues that the structure of the Metaverse could give students access to interesting
digital resources, encourage them to interact with educational material, and inspire them
to invent exciting new activities [1, 7, 19, 25]. Educators are responsible for developing,
refining, and supplying the Metaverse server administrators with digital resources for
students to utilise [33]. It is worth noting that both students’ time management and
the Metaverse’s implementation in the classroom present challenges [38]. Students have
trouble accessing the Metaverse due to the increased difficulty in managing their time and
the many technical barriers that exist [4]. Students often lack the technical proficiency and
expertise necessary to effectively apply what they have learned [27]. Finally, the amount
of work, planning, and experimentation necessitated to successfully deploy Metaverse
in the classroom are limitations on its growth [38].

5 Conclusion and Way Forward

This study demonstrates the fundamental limitations of the Metaverse, despite the solid
groundwork it provides for its implementation in the classroom. For instance, this system-
atic review has certain restrictions due to the databases and keywords used. Non-English
studies on the use of the Metaverse in education were excluded from this analysis. In
58 A. Alam and A. Mohanty

addition, there were not many publications covering this issue in the leading journals for
educational technology, which may be an indication of how new this field is. Because
of this, the current study contributes to the body of knowledge among academics and
practitioners on potential avenues for future research on this topic, specifically the use
of the Metaverse in teaching and learning.
This study provides an in-depth analysis of how the Metaverse may be used in the
classroom. The findings suggest that the Metaverse might be used to solve real-world
problems in a virtual setting, opening up new educational opportunities that were pre-
viously unattainable due to constraints of time, place, and resources. Furthermore, they
reveal the gap in knowledge about the use of lifelogging in metaverse education. More
research is needed to determine how using Metaverse affects students with disabilities. It
must be delineated that providing teachers with technical support, encouraging teacher
training in both asynchronous and synchronous ways, and providing pupils with a col-
laborative, engrossing, and dynamic computer-simulated platform is extremely essential
for the successful integration of Metaverse in education.
However, although Metaverse technology is not new, it has evolved considerably
during the last two decades. With the advancement of technology comes the return of
both its advantages and its drawbacks. Many of the papers we looked at, for example,
focused more on the benefits of the Metaverse than the threats it brought. With the backing
of large technology companies, it is becoming more popular, but caution is warranted
since it poses potential threats to schools. In spite of its short existence, this technology
is extremely vulnerable due to its developmental stage. For instance, how can we ensure
the safety and privacy of our users? How does one make money in a virtual setting that
generates copious amounts of data? In a world dominated by algorithms and AI, how
do we define right and wrong? What kind of social and physiological implications may
we foresee from the Metaverse, a realm where physical and digital realities merge? We
need to scientifically explore the benefits of the Moreover, the analysed research makes
it clear that Metaverse in education is based on cutting-edge technology, which may be a
gift for schools or universities with cutting-edge infrastructure but a curse for people who
are affected by it, especially in poor nations. Metaverse may be made more accessible
and inclusive for every learner in order to contribute to the Sustainable Development
Goals (SDGs), especially SDG-4 addressing quality education.
These findings underscore the fact that application of Metaverse in classroom teach-
ing is still in its early stages, with all the associated benefits and drawbacks that it entails.
Further, vital concerns remain unanswered. For instance, in light of the impact of EdTech
firms, how will we ensure that students are given the tools they need to succeed? How do
we ensure their safety in a world where computers make all the important decisions? Is
this a brave new world where everything goes, or are we all bound by digital restraints?
Before we rush into the Metaverse, perhaps we should take a moment to reflect on the
costs we have already incurred. Is Metaverse a user that mines and profits from user-
produced data, or we, as its products, are the users? Will there be a swarm of metabots
out to fool people, or will only humans have access? When we are cut off from the real
world like this, do we have the resources to deal with cyber pathologies in these games?
If we wish to use the artificial Metaverse for educational reasons, do we have a plan to
humanise these processes? Although the novelty of the Metaverse may tempt us to go in
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 59

headfirst, there are still many serious issues to think about before making a permanent
transition there.

References
1. Alam, A.: Challenges and possibilities in teaching and learning of calculus: A case study of
India. Journal for the Education of Gifted Young Scientists 8(1), 407–433 (2020)
2. Mughal, M.Y., Andleeb, N., Khurram, A.F.A., Ali, M.Y., Aslam, M.S., Saleem, M.N.: Per-
ceptions of teaching-learning force about metaverse for education: a qualitative study. Journal
of Positive School Psychology 6(9), 1738–1745 (2022)
3. Mystakidis, S.: Metaverse. Encyclopedia 2(1), 486–497 (2022)
4. Mustafa, B.: Analyzing education based on metaverse technology. Technium Social Sciences
Journal 32, 278–295 (2022)
5. Alam, A.: Pedagogy of Calculus in India: An Empirical Investigation. Periódico Tchê Química
17(34), 164–180 (2020)
6. Zhong, J., Zheng, Y.: Empowering future education: learning in the Edu-METAVERSE. In:
2022 International Symposium on Educational Technology (ISET), pp. 292–295. IEEE (2022
July)
7. Fitria, T.N., Simbolon, N.E.: Possibility of metaverse in education: opportunity and threat.
SOSMANIORA: Jurnal Ilmu Sosial dan Humaniora 1(3), 366–376 (2022)
8. Barráez-Herrera, D.P.: Metaverse in the context of virtual education. Metaverse 3(1), 9 (2022)
9. Alam, A.: Possibilities and challenges of compounding artificial intelligence in India’s
educational landscape. Int. J. Adv. Sci. Technol. 29(5), 5077–5094 (2020)
10. Wang, H., Chen, D., Deng, Q.: The formation, development and research prospect of
educational metaverse. Educ. J. 11(5), 260–266 (2022)
11. Phakamach, P., Senarith, P., Wachirawongpaisarn, S.: The metaverse in education: the future
of immersive teaching & learning. RICE J. Creative Entrepreneu. Manage. 3(2), 75–88 (2022)
12. Yu, J.E.: Exploration of educational possibilities by four metaverse types in physical
education. Technologies 10(5), 104 (2022)
13. Alam, A.: Test of knowledge of elementary vectors concepts (TKEVC) among first-semester
bachelor of engineering and technology students. Periódico Tchê Química 17(35), 477–494
(2020)
14. Lee, H., Hwang, Y.: Technology-enhanced education through VR-making and metaverse-
linking to foster teacher readiness and sustainable learning. Sustainability 14(8), 4786 (2022)
15. Areepong, T., Nilsook, P., Wannapiroon, P.: A study of a metaverse interdisciplinary learning
community. In: 2022 Research, Invention, and Innovation Congress: Innovative Electricals
and Electronics (RI2C), pp. 290–296. IEEE (2022 August)
16. Alam, A.: Designing XR into Higher Education using Immersive Learning Environments
(ILEs) and Hybrid Education for Innovation in HEIs to attract UN’s Education for Sustainable
Development (ESD) Initiative. In: 2021 International Conference on Advances in Computing,
Communication, and Control (ICAC3), pp. 1–9. IEEE (2021)
17. Zhai, X., Chu, X., Wang, M.: Education metaverse: innovations and challenges of the new
generation of internet education formats. Metaverse 3(1), 13 (2022)
18. Singh, J., Malhotra, M., Sharma, N.: Metaverse in Education: An Overview. Applying
Metalytics to Measure Customer Experience in the Metaverse, 135–142 (2022)
19. Alam, A.: Possibilities and Apprehensions in the Landscape of Artificial Intelligence in Edu-
cation. In: 2021 International Conference on Computational Intelligence and Computing
Applications (ICCICA), pp. 1–8. IEEE (2021)
60 A. Alam and A. Mohanty

20. Burnett, G.E., Harvey, C., Kay, R.: Bringing the Metaverse to Higher Education: Engaging
University Students in Virtual Worlds. In: Methodologies and Use Cases on Extended Reality
for Training and Education, pp. 48–72. IGI Global (2022)
21. Fernandes, F.I.L.I.P.E., Werner, C.L.Á.U.D.I.A.: A Systematic Literature Review of the Meta-
verse for Software Engineering Education: Overview, Challenges and Opportunities. Preprint,
Sep. (2022)
22. Alam, A.: Should Robots Replace Teachers? Mobilisation of AI and Learning Analytics in
Education. In: 2021 International Conference on Advances in Computing, Communication,
and Control (ICAC3), pp. 1–12. IEEE (2021)
23. Contreras, G.S., González, A.H., Fernández, M.I.S., Martínez, C.B.: The importance of the
application of the metaverse in education. Mod. Appl. Sci. 16(3), 1–34 (2022)
24. Rospigliosi, P.A.: Metaverse or simulacra? roblox, minecraft, meta and the turn to virtual
reality for education, socialisation and work. Interact. Learn. Environ. 30(1), 1–3 (2022)
25. Alam, A.: A digital game based learning approach for effective curriculum transaction for
teaching-learning of artificial intelligence and machine learning. In: 2022 International Con-
ference on Sustainable Computing and Data Communication Systems (ICSCDS), pp. 69–74.
IEEE (2022)
26. Mistretta, S.: The Metaverse—An Alternative Education Space. AI, Computer Science and
(2022)
27. Dahan, N.A., Al-Razgan, M., Al-Laith, A., Alsoufi, M.A., Al-Asaly, M.S., Alfakih, T.: Meta-
verse framework: a case study on E-learning environment (ELEM). Electronics 11(10), 1616
(2022)
28. Alam, A.: Educational robotics and computer programming in early childhood education:
a conceptual framework for assessing elementary school students’ computational thinking
for designing powerful educational scenarios. In: 2022 International Conference on Smart
Technologies and Systems for Next Generation Computing (ICSTSN), pp. 1–7. IEEE (2022)
29. Wang, M., Yu, H., Bell, Z., Chu, X.: Constructing an Edu-metaverse ecosystem: a new and
innovative framework. IEEE Transactions on Learning Technologies (2022)
30. Batnasan, G., Gochoo, M., Otgonbold, M.E., Alnajjar, F., Shih, T.K.: ArSL21L: arabic sign
language letter dataset benchmarking and an educational avatar for metaverse applications.
In: 2022 IEEE Global Engineering Education Conference (EDUCON), pp. 1814–1821. IEEE
(2022 March)
31. Alam, A.: Employing adaptive learning and intelligent tutoring robots for virtual classrooms
and smart campuses: reforming education in the age of artificial intelligence. In: Shaw, R.N.,
Das, S., Piuri, V., Bianchini, M. (eds) Advanced Computing and Intelligent Technologies.
Lecture Notes in Electrical Engineering, vol 914. Springer, Singapore (2022)
32. Jovanović, A., Milosavljević, A.: VoRtex metaverse platform for gamified collaborative
learning. Electronics 11(3), 317 (2022)
33. Hwang, G.J., Chien, S.Y.: Definition, roles, and potential research issues of the metaverse
in education: An artificial intelligence perspective. Computers and Education: Artificial
Intelligence, 100082 (2022)
34. Alam, A.: Investigating Sustainable Education and Positive Psychology Interventions in
Schools Towards Achievement of Sustainable Happiness and Wellbeing for 21st Century
Pedagogy and Curriculum. ECS Trans. 107(1), 19481 (2022)
35. Arpaci, I., Karatas, K., Kusci, I., Al-Emran, M.: Understanding the social sustainability of
the Metaverse by integrating UTAUT2 and big five personality traits: A hybrid SEM-ANN
approach. Technology in Society, 102120 (2022)
36. Lee, H., Woo, D., Yu, S.: Virtual reality metaverse system supplementing remote education
methods: based on aircraft maintenance simulation. Appl. Sci. 12(5), 2667 (2022)
 Metaverse and Posthuman Animated Avatars for Teaching-Learning Process 61

37. Alam, A.: Mapping a sustainable future through conceptualization of transformative learning
framework, education for sustainable development, critical reflection, and responsible citi-
zenship: an exploration of pedagogies for twenty-first century learning. ECS Trans. 107(1),
9827 (2022)
38. Wang, Y., Lee, L.H., Braud, T., Hui, P.: Re-shaping Post-COVID-19 Teaching and Learning:
A Blueprint of Virtual-Physical Blended Classrooms in the Metaverse Era. arXiv preprint
arXiv:2203.09228 (2022)
39. Alam, A.: Positive Psychology Goes to School: Conceptualizing Students’ Happiness in 21st
Century Schools While ‘Minding the Mind!’Are We There Yet? Evidence-Backed. School-
Based Positive Psychology Interventions. ECS Transactions 107(1), 11199 (2022)
40. Yang, J., Zhou, Y., Huang, H., Zou, H., Xie, L.: Metafi: device-free pose estimation via
commodity wifi for metaverse avatar simulation. arXiv preprint arXiv:2208.10414 (2022)
41. Sutopo, A.H.: Developing Teaching Materials Based on Metaverse. Topazart (2022)
42. Alam, A.: Social robots in education for long-term human-robot interaction: socially sup-
portive behaviour of robotic tutor for creating robo-tangible learning environment in a guided
discovery learning interaction. ECS Trans. 107(1), 12389 (2022)
43. Teng, Z., Cai, Y., Gao, Y., Zhang, X., Li, X.: Factors Affecting Learners’ Adoption of an
Educational Metaverse Platform: An Empirical Study Based on an Extended UTAUT Model.
Mobile Information Systems (2022)
44. Gupta, Y.P., Chawla, A., Pal, T., Reddy, M.P., Yadav, D.S.: 3D networking and collaborative
environment for online education. In: 2022 10th International Conference on Emerging Trends
in Engineering and Technology-Signal and Information Processing (ICETET-SIP-22), pp. 1–
5. IEEE (2022 April)
45. Alam, A.: Cloud-Based E-learning: Scaffolding the Environment for Adaptive E-learning
Ecosystem Based on Cloud Computing Infrastructure. In: Satapathy, S.C., Lin, J.CW., Wee,
L.K., Bhateja, V., Rajesh, T.M. (eds) Computer Communication, Networking and IoT. Lecture
Notes in Networks and Systems, vol 459. Springer, Singapore (2023). https://doi.org/10.1007/
978-981-19-1976-3_1
46. Kim, K., Jeong, Y., Ryu, J.: Does the real face provision improve the attention and social
presence in metaverse as learning environments?. In: Society for Information Technol-
ogy & Teacher Education International Conference, pp. 1760–1765. Association for the
Advancement of Computing in Education (AACE) (2022 April)
47. Lim, S., Byun, H.: Development of Learning Analysis Framework on Metaverse
48. Hines, P., Netland, T.H.: Teaching a Lean masterclass in the metaverse. International Journal
of Lean Six Sigma, (ahead-of-print) (2022)
49. Jagatheesaperumal, S.K., Ahmad, K., Al-Fuqaha, A., Qadir, J.: Advancing Education Through
Extended Reality and Internet of Everything Enabled Metaverses: Applications, Challenges,
and Open Issues. arXiv preprint arXiv:2207.01512 (2022)
 Tuning Functional Link Artificial Neural
Network for Software Development Effort
Estimation

Tirimula Rao Benala1(B) and Satchidananda Dehuri2

1 Department of Information Technology, JNTU-GV College of Engineering, Vizianagaram,
Jawaharlal Nehru Technological University, Gurajada-Vizianagaram 535003, Andhra Pradesh,
India
[email protected]
2 Department of Computer Science, Fakir Mohan University, Vyasa Vihar, Janugnai,

Balasore 756019, Odisha, India

Abstract. Software development effort estimation (SDEE) is a critical task in

project management for accurate planning, staffing, resource allocation, schedul-
ing, and cost estimation. Detailing nonlinear correlations between cost drivers and
project costs using conventional parametric methodologies is difficult. In this con-
text, we evaluated methods based on foundation-centered swarm intelligence and
functional link neural networks for SDEE. The primary objective of this study was
to investigate the use of functional link artificial neural networks for improving the
predictability of SDEE. The findings are presented as computational intelligence
methods for SDEE.

Keywords: Functional link artificial neural network · Software development

effort estimation (SDEE) · Chebyshev polynomial · Particle swarm
optimization · Genetic algorithm · Back-propagation

1 Introduction
The use of modern software procedures has helped software development companies
provide high-quality software on time and minimize cost. Therefore, precise cost/effort
estimation is crucial in the early stage of the software development life cycle. Accurate
effort estimation influences several fundamental project management tasks, including
budgeting, personnel, and resource allocation. Various cost estimation methods have
been proposed. Conventionally, these methods are separated into algorithmic and non-
algorithmic software cost estimation methodologies. SLIM [9] and COCOMO are two
well-known algorithmic approaches [8, 10]. The most prevalent nonalgorithmic strate-
gies include expert judgment, estimation by analogy [5, 11, 12], and machine learn-
ing techniques [13–15]. Machine-learning-based techniques, such as artificial neural
networks, analogy-based estimates, support vector regression, and classification and
regression trees have been used as an alternative to software effort prediction models.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 62–81, 2022.
https://doi.org/10.1007/978-3-031-23233-6_5
 Tuning Functional Link Artificial Neural Network for SDEE 63

Artificial neural networks (ANNs) are used to solve complex, mathematically ill-
defined problems. The higher-order neural network (HON) is a unique neural network
that has attracted considerable research attention because it helps overcome the limita-
tions of many non-HONs. The functional link neural networks (FLANN) is an extension
of the HON. Therefore, examining HON characteristics is critical, even though this study
focuses mainly on the CFLANN. A feed-forward neural network is inadequate for use in
several applications. Therefore, a two-layer neural network with threshold activation is
used. Minsky and Papert [16] revealed that the network always converges to linearly sep-
arable functions such as AND, OR, and NOT but cannot represent or learn XOR. Thus,
feed-forward neural networks exhibit limited expressiveness. A hidden layer with an
arbitrary activation function was introduced to address this bottleneck. This architecture
was characterized by a feed-forward neural network with many layers. Multilayer per-
ceptron with BP learning is the most widely used method. Under finite norms, the error
in approximating any bounded continuous function can be arbitrarily small if sufficient
hidden units are accessible in an MLP [17, 18]. MLP can express numerous nonlinear
functions, and except for gradient descent, it does have any other limitations. When
calculating partial derivatives, the function is assumed to be continuous in BP learning.
Consequently, the approach is seldom helpful. Second, BP learning increases the
computation cost of MLP, which slows convergence. Several local minimums exist
along the cost-function surface in the weight space for complex computer workloads.
According to a mathematical study, the gradient descent technique of BP converges to
a local minimum [19–22]. Lyapunov’s stability theory states that the output tracking
error cannot converge to zero [23]. The determination of many design factors of neural
network architecture, such as the number of hidden units and neurons in a hidden unit,
is challenging. Designing the architecture of a neural network is complex [24].
Numerous alternative neural network topologies have been proposed for nonlinear
systems based on the Lyapunov stability theory to solve the aforementioned instability
and convergence problems [25]. The explicit, hidden layer of the normal feed-forward
neural network is eliminated, and an implicitly hidden layer is established by equipping
the input layer with higher-order units known as functional expansions [26, 27]. Thus,
an entirely novel neural network paradigm known as HONs, which includes FLANNs
[28–31] and ridge polynomial neural networks (RPNNs) [32], has been proposed.
We investigated software development effort estimation (SDEE) methods based on
ANNs with functional links. FLANN’s learning algorithms incorporate genetic algo-
rithm (GA), particle swarm optimization (PSO), adaptive PSO (APSO), and improved
PSO (ISO), in addition to BP, to reduce computation load and speed convergence rate
relative to MLP. The orthogonal basis function is the polynomial functional expansion of
the Chebyshev polynomial. The FLANN is a neural network with a single layer and no
hidden layers. Chebyshev’s polynomial orthogonal basis function (functional link unit)
introduces nonlinearity and increases the dimension of the input vector. Therefore, the
FLANN requires lower processing power compared with MLP. A set of linearly inde-
pendent functional expansion units (nonlinear units) are used to generate hyperplanes
to improve discrimination in the input pattern space. A notable aspect of the FLANN
architecture for predicting software development effort is the generation of the output
(effort) by expanding the inputs (cost drivers) using orthogonal basis functions. Based
64 T. R. Benala and S. Dehuri

on the information in the network, the FLANN hidden layer can be used to predict the
effort required to construct software.
The rest of the paper is organized as follows: Sect. 2 covers the FLANN. Section 3
presents swarm-intelligence-based learning algorithms for the FLANN. Section 4 con-
tains the performance metrics that were used to evaluate SDEE accuracy. Section 5
details the COCOMO’81 test suite from the promise repository. The results of the pro-
posed SDEE models performance studies using the COCOMO’81 test suite are discussed
in Sect. 6. Finally, Sect. 7 presents the conclusion of the study.

2 Functional Link ANN-based SDEE

The FLANN is a HON introduced by Klassen and Pao in 1988 [33]. It is a single-layer
feed-forward neural network for creating arbitrary complex decision regions with an
input layer and an output layer. By analyzing the final output layer to forecast output
and increasing the input vector (cost drivers), the FLANN model incorporates non-
linear input–output interactions (effort in Person-Months). Furthermore, an implicitly
concealed unit is produced after the input vector (cost driver) has been expanded using
the Chebyshev polynomial. Therefore, the weighted summation closely represents the
software development effort. Swarm-intelligence-based learning techniques train the
network and improve the output.
The block diagram of an n-dimensional input FLANN architecture is displayed in
Fig. 1. A nonlinear network with only one visible layer is used. In this method, the input
space is expanded into a high-dimensional feature space using a functional expansion
block or functional link (Chebyshev orthogonal polynomial). To evaluate the input pat-
tern, as opposed to the linear weighting of the linear links in an MLP, the functional links
produce a set of linearly independent orthogonal basis functions (nonlinear functions). To
accommodate the nonlinear characteristics of the problem, this phenomenon increases
the dimension of the input pattern. Therefore, prediction accuracy may increase in the
expanded feature space [24, 34]. FLANN learning can be considered an approximation
function that is approximating or interpolating a continuous, multivariate function. The
FLANN is a collection of orthogonal basis functions with a fixed number of weight
parameters. The functional expansion unit improves the discrimination capability of
the FLANN by increasing the dimension of an n-dimensional input pattern in the m-
dimensional feature space. The problem is obtaining the weight parameters that extend
the best possible approximation of the set of input–output examples after completing the
measure design issue of selecting the basis function [35]. Using the Stone–Weierstrass
theorem, Chen et al. [1] in 2008 revealed that the FLANN could be used as a universal
approximator.
Let k be the number of input–output pattern pairs that the FLANN should learn. The
input–output relationship of the FLANN can be described as follows. Let us consider a set
of basic functions Υ = {Φ(A)}i∈N , N = {1, 2, . . . } be the set of basis functions, where
A is a subset of n-dimensional Euclidean space Rn , „ with the following properties:
j
(1) φ1 = 1 and the subset Υj = {φi ∈ Υ }i=1 is a linearly independent set, that is, if
N  1/2
j
i=1 (θi φi ) = 0, then θi = 0 for all i = 1, 2, . . . , j, and supj φ  < ∞.
2
i=1 i A
 Tuning Functional Link Artificial Neural Network for SDEE 65

Thus, the FLANN consists of N basis functions {φ1 , φ2 , . . . .., φN } ∈ ΥN , with the
following input–output relationship for the jth output:
 


y j = ρ sj ; (1)
N
where sj = i=1 (θji φi (I )), I ∈ AR , i.e., I = [i1 , i2 , . . . .., in ] is the input
n T
     
T
pattern vector, y ∈ Rm , that is, y = y1 , y2 , . . . .., yn is the output vector, and
θj = [θj1, θj2, . . . . . . θjN ] is the weight vector associated with the jth output of the
FLANN. Here, ρ(.) = tanh(.) is a nonlinear function (.).
Consider the m-dimensional output vector (1), which can be expressed as

O = θ Φ, (2)

where θ is a (m × N ) weight matrix of the FLANN depicted by θ =

[θ1 , θ2 , . . . , θm ]T ,  = [φ1 , φ2 , . . . .., φN ]T is the basis function vector, and O =
[O1 , O2 , . . . . . . , ON ]T is a matrix of the linear outputs of the FLANN. The m-


dimensional output vector y can be expressed as follows:



y = ρ(O) = hθ (x), (3)

Assume the input pattern vector I k be of dimension n and the output ok be a scalar.
The training patterns are denoted by (Ik , ok ) and the network weight is θ (k), where k
is the iteration number. According to (4), the jth output of the FLANN at iteration k is
expressed as follows:
 N   
yj (k) = ρ θji (k)φi (Xk ) = ρ θj (k)φ T (Xk ) , (4)
i=1

Fig. 1. Functional link neural networks (FLANN) architecture

66 T. R. Benala and S. Dehuri

For all I ∈ A and j = 1, 2, . . . ., m, where φ(Xk ) = [φ1 (Xk ), φ2 (Xk ), . . . .., φN (Xk )].


Let ej (k) = yj (k) − yj (k) denote the corresponding error

When using the BP algorithm for a single layer, the update rule for all FLANN
weights is represented as follows:
θ (k + 1) = θ (k) + μδ(k)∅(Xk ), (5)
where (θ (k))m×N is the FLANN weight matrix, and δ and μ are the error and learning
rates, respectively.

2.1 Justification of the Use of Chebyshev Polynomial as the Orthogonal Basis

Function
The nonlinear approximation capacity of the Chebyshev orthogonal polynomial is potent
according to the approximation theory [6]. Chebyshev polynomial was combined with
the FLANN to obtain the CFLANN [7]. The Weierstrass theorem states that a polynomial
is always arbitrarily close to any continuous function. Notably, polynomial power series
interpolation can be used to obtain an excellent estimate of a function with only a
few data points. A good fit with little error requires a polynomial of extremely high
degree, which is computationally not feasible and restricts the application of power
series. Unlike power series with slow convergence problems, the Chebyshev polynomials
exhibit a high computational economy with fast convergence. Thus, the Chebyshev
series is considerably more competent than other power series of the same degree. The
efficacy of Chebyshev polynomial orthogonal basis functions concerning convergence
criteria is superior to that of other orthogonal polynomials [6]. Moreover, Chebyshev
polynomials involve simple computation compared with trigonometric polynomials.
These characteristics motivated us to use the CFLANN for approximating the target
function for SDEE [24, 36, 37].
The polynomials are described in the section. Additionally, higher-order Chebyshev
polynomials can be created using recursive formulas in Table 1.

Table 1. Chebyshev polynomials (−1 ≤ x ≤ 1).

Consider the following two-dimensional input pattern I = [i1 , i2 ]T obtained by using

Chebyshev
  functions: τ = [1, A0 (x1 ), A1 (x1 ), . . . ..; 1, A0 (x2 ), A1 (x2 ), . . . ..] where
Ai xj is a Chebyshev polynomial, i is the order of the polynomials, and j = 1, 2.
The following theorem emphasizes that the CFLANN is homomorphic with a feed-
forward MLP.
 Tuning Functional Link Artificial Neural Network for SDEE 67

Theorem. Assume a feed-forward MLP neural network with only one hidden layer and
a linear activation function for the output layer. If all activation functions of the hidden
layer satisfy the Riemann integrable condition, the feed-forward neural network can be
represented as a Chebyshev neural network. Lee et al. (1998) provided a detailed proof
of the theorem.

3 Swarm Intelligence-Based Learning Algorithms for the CFLANN

This section provides an overview of four learning algorithms, namely classical PSO,
adaptive PSO, improved PSO, and GA.

3.1 Classical PSO

PSO is a stochastic population-based swarm intelligence algorithm that was developed
for function optimization by James Kennedy and Russell Eberhart in 1995 [47]. Each
solution in PSO can be modeled as a particle flying through the hyper-space problem.
In the uniform random search space, a population of particles is initialized with random
positions −→x k and velocity −
→v k . The PSO algorithm traverses the search space by using
the fitness function. The root mean square error is used as a fitness function. In every
iteration, the fitness function for each particle is evaluated with its current position,
and this information is used to adjust the trajectory of the particle by determining the
best particle in the sub-swarms and the entire swarm population [38]. The position and
velocity of a particle are updated using the following equations in the classical PSO
algorithm at iteration t.
−
→ → 
vk (t + 1) = w ⊗ −
→
vk (t) + −
→
c1 ⊗ −
→r1 (t) ⊗ −pk (t) − −
→
xk (t)
 
+−→
c ⊗−
2
→r (t) ⊗ −
2
→
p (t) − −
g
→
x (t) ,
k (6)

−
→
xk (t + 1) = −
→
xk (t) + −
→
vk (t + 1), (7)

In Eq. (6), symbol ⊗ represents point-by-point vector multiplication. A value less

than one “w” is referred to as inertia. Here, c1 and c2 are real constants that are non-
negative. Venter and Sobieski (Shi Y and Eberhart RC 1999) labeled c1 as “self-
confidence” and c2 as “swarm confidence.” Furthermore, − →r1 and −→
r2 are two uni-
formly distributed random numbers in the range [−1,1]. In the original PSO algorithm,
w = 1, c1 = 2, and c2 = 2 were used. Here, − →
pk (t) represents the best position found
−
→
by the given particle, pg (t) refers to the best position found by any particle, and the
corresponding fitness values are referred to as the best of the particle and global best,
respectively. The process of evaluating fitness function, velocity update, and updating
position is repeated until a stopping criterion is reached. The objective of the algorithm
is to change the particle position (weight vectors) in each iteration until the maximum
change in the best fitness value is less than the specified threshold value for a given
number of iterations as follows:
 −  
f →
pg (t) − f (−
→
pg (t − 1)) ≤∈, t = 2, 3, . . . .., I (8)
68 T. R. Benala and S. Dehuri

3.2 Improved PSO Technique

Based on classical PSO, an ISO improves search efficiency and increases the likelihood
of obtaining the global optimum without considerably reducing the speed of convergence
or the simplicity of the PSO structure. The mathematical model of ISO is as follows:
−
→ → 
vk (t + 1) = λ ⊗ −
→ c1 ⊗ −
vk (t) + −
→ →
r1 (t) ⊗ −pk (t) − −
→
xk (t)
→ 
+−→
c ⊗−
2
→
r (t) ⊗ −
2 p (t) − −
g
→x (t) ,
k (12)

−
→
xk (t + 1) = −
→
xk (t) + −
→
vk (t + 1), (13)

where λ is the newly defined adaptive inertia weight whose value decreases with each
iteration. The inertia weight decreases linearly until Gen1 and nonlinearly from Gen1 +
1 to Gen2, as presented in eqns. (14) and (15). The inertia weight adaptation mechanism
 Tuning Functional Link Artificial Neural Network for SDEE 69

enables the IPSO algorithm to achieve the best results by balancing the trade-off between
exploration and exploitation [24].

Adaptive Inertia Weight (λ):

The following equations describe the adaptive inertia weight proposed by Dehuri et al.
[28]:

λ1
λ1 = λ0 − ×i ∀i = 1..Gen1, (14)
Gen1
(Gen1 + 1) − i
λ1 = (λ0 − λ1 ) × exp ∀i = Gen1 + 1..Gen2, (15)
i

where λ0 represents the initial weight, λ1 represents the end point of linear selection,
Gen1 represents the number of generations during which inertia weight is decreased
linearly, and Gen2 represents the maximum generation. The values of λ0 and λ1 are
based on empirical observations.
The more the particle improves, the smaller the area it should explore. The exploration
capability of PSOs is greater than the exploitation capability. Therefore, the model out-
performs local search in terms of global search. With a high probability, the self-adaptive
evolutionary strategy generates small Gaussian and Cauchy perturbations suitable for
local search optimization. The model improves the particle by calibrating PSO solutions.

Self-adaptive Cauchy Mutation: A random variable has Cauchy distribution (C(t)),

if its density function is as follows:
t
C(t) = , −∞ < x < +∞ (16)
π(t 2 + x2 )
The self-adaptive Cauchy mutation is represented as follows:

vki (t + 1) = vki × exp τ × Cgi (0, 1) + τ × Cki (0, 1) , (17)

xki (t + 1) = xki (t) + vki (t + 1), (18)

70 T. R. Benala and S. Dehuri
 Tuning Functional Link Artificial Neural Network for SDEE 71

3.3 Adaptive PSO

APSO is an advancement over PSO that improves search efficiency and increases the
likelihood of reaching the global optimum by dynamically varying the inertia weight
based on population fitness variance. The inertia weight determines how the previous
velocity of the particle influences its velocity at the current time step [39, 40]. The inertia
weight wi is updated by calculating the population fitness variance as follows:

M fi − favg 2
τ= , (22)
i=1 f

where favg = average fitness of the population, fi = fitness of the ith particle in the
population, and M = total number of particles.
   
f = −max fi − favg  i = 1, 2, . . . ., M , if max fi − favg  > 1 (23)

 
= −1 if max fi − favg  < 1 (24)

If τ is significant, then the population is in exploration aptitude (global search),

whereas small τ enhances the exploitation capability (local search) of the swarm. To
circumvent this problem, the inertia weight is adjusted as follows:

w(k) = λw(k − 1) + (1 − λ)τ (25)

A forgetting factor λ is selected as 0.95 for faster convergence.

72 T. R. Benala and S. Dehuri

3.4 GA

Inspired by biological evolution, Holland created the GA in 1975 [2]. The GA is a prob-
abilistic search algorithm based on Darwin’s principle of biological evolution through
reproduction and “survival of the fittest” [3]. The algorithm begins with a randomly
generated population of individuals (chromosomes). The search for a global optimum is
managed by shifting from an initial population of individuals to a new population using
genetics-like operators such as selection, crossover, and mutation. The explanation of
the proposed genetic-CFLANN algorithm pseudo-code is as follows:
 Tuning Functional Link Artificial Neural Network for SDEE 73

3.5 BP
BP is a fundamental neural network platform that was first used in the 1960s and pop-
ularized in 1989 by Rumelhart, Hinton, and Williams. BP is used to train feed-forward
neural networks and is applicable to other ANNs and functions. When fitting a neural
network, BP is used to efficiently compute the gradient of the loss function for the net-
work weights for a single input–output sample. To train multilayer networks, gradient
methods can be used to update weights to minimize loss. Gradient descent and stochas-
tic gradient descent are popular methods. In BP, the chain rule is used to calculate the
74 T. R. Benala and S. Dehuri

gradient of the loss function for each weight one layer at a time, iterating backward from
the last layer to avoid unnecessary calculations of intermediate terms [41]. The proposed
BP-CFLANN algorithm pseudo-code is as follows:

4 Performance Evaluation Metrics

This section presents a few metrics for the performance evaluation of the proposed mod-
els widely used in existing SDEE literature. To evaluate the efficacy of proposed swarm
intelligence-based SDEE models, the following performance metrics are considered
to provide a powerful performance evaluation: standardized accuracy (SA) [43], mean
magnitude of relative error (MMRE), median magnitude of relative error (MdMRE),
and PRED (0.25) [42]. When the MMRE and MdMRE values are at most 0.25, and
the PRED value is at least 0.75, the SDEE model performs well. MREs are defined as
MMRE. The MRE of the ith project is defined as follows:
 
 yi − yi 


MRE i =   , (30)
yi 

Therefore, MMRE = ni=1 MRE i /n
MdMRE, a global error measure, is defined as the median of all MREs and is less
sensitive to outliers.
MdMRE = median(BRE i ), (31)
PRED(x) is described as the percentage of predictions falling within the actual known
value x, specified as follows:
100 N
PRED(x) = × Di , (32)
N i=1
 Tuning Functional Link Artificial Neural Network for SDEE 75


1 if MBRE < 100
x
Di = , (33)
0 otherwise

when x = 25, the PRED metric is defined as PRED (0.25).

MAE, a non-symmetric evaluation criterion, is used as an error indicator. The MAE
is determined using Eqs. (34) and (35). Because the residuals are not standardized, the
model with MAE as a performance measure is difficult to understand. To avoid this
difficulty, Sheppred and McDonell created a new metric called SA, as expressed in
Eq. (36). To estimate the effect size, Eq. (37) is used. SA measures whether a prediction
model is relevant and outperforms random guessing. It is a percentage (0–100 scale)
or number (0–1 scale). Higher SA values indicate the superiority of a model to random
guessing, which explains how well the strategy works. The effect size DELTA ( ) checks
if the prediction models are by accident or by choice by comparing them to a random
guess model. A value larger than 0.5 is better than that less than 0.2 (Shepperd and
MacDonell 2012; Azzeh, M. et al. 2015).
 
AE i = yi − yi ,


(34)
N
i=1 AE i
MAE = , (35)
N
MAE
SA = 1 − , (36)
MAE P0
MAE − MAEP0
= , (37)
s P0


• Actual and projected project efforts are yi and yi . The mean absolute error of the pre-


diction model is MAE. MAE P0 is the mean of many random guesses. a yi for the target
case t is predicted by random sampling (with equal probability) over all the  remaining
n − 1 case and considering yt = yr , where r is taken from 1, . . . . . . ., n r = t.

5 Description of the Dataset

In this part, the COCOMO’81 Dataset from the prospective repository test suite was
used to benchmark the prospective model. Boehm conceived the idea for COCOMO’81
(1981) [8]. COCOMO is an open-source, well-documented algorithmic cost model.
COCOMO’81 is the most practical, well-known, and often-cited conventional model.
Boehm illustrated the “fundamental,” “intermediate,” and “detailed” levels of project
estimation. Initially, a preliminary estimate is made. The first estimate is refined in two
steps [44]. The most frequently used model is the intermediate COCOMO. The interme-
diate COCOMO model considers the following parameters, in addition to the product
size: (a) software product size, (b) project development mode B (scaling factor), and (c)
15 cost drivers (effort multipliers). Organic, semi-detached, and embedded development
are three problematic modes, with the values 1.05, 1.12, and 1.20, respectively. Next,
76 T. R. Benala and S. Dehuri

15 effort multipliers are used to modify the software development effort based on its
qualities. The cost drivers are the multiplying factor, that is, very low, low, nominal,
high, very high, and extra high. Based on the factor’s effect on productivity, an actual
number (effort multiplier) is assigned to each rating.

6 Experiments and Results

We tested four CFLANN models. First, PSO-CFLANN trains CFLANN with PSO. Sec-
ond, ISO-CFLANN tweaks CFLANN’s weights using ISO. Third, the APSO-CFLANN
model fine-tunes weights. Fourth, genetic-CFLANN optimizes CFLANN weights using
a GA. The proposed techniques are compared with well-known techniques such as func-
tional link artificial neural networks with BP learning and ANNs with back-propagation
learning. Method validity depends on selecting and creating experimental circumstances.
This section describes the circumstances of the experiment. The proposed solutions are
implemented on a PC with an Intel Core i5- 2410M, 2.30 GHz CPU, 4 MB RAM, and
MATLAB 8.3 programming environment. Using min-max normalization in the interval
[0, 1] eliminates the unequal feature effect [45]. The dataset is randomly partitioned
into training and testing subsets. Leave-one-out cross-validation is used to assess model
accuracy. The algorithm is trained on N−1 random occurrences and validated on one.
Cross-validation is repeated N times [46]. The training set constructs the model, the
validation set tweaks to control parameters, and the test set predicts model performance.
The outcomes of the strategies outlined in Sect. 3 are reported. The outcomes of
simulations using the COCOMO dataset for PSO-CFLANN, APSO-CFLANN, ISO-
CFLANN, Genetic-CFLANN, BP-CFLANN, and BP-ANN are presented in Tables 2 and
3. According to Table 2, the MMRE and MdMRE results revealed that PSO-CFLANN
is superior to all other approaches, with values of 0.000259 and 8.49E−05, respectively.
PRED (0.25) delivers identical results across all approaches. Because the research on
SDEE indicates that MMRE and PRED are biased performance indicators, SA and
DELTA are used to determine the optimal technique, and results are depicted in Table 3.
At PSO-CFLANN, the SA and DELTA values are greater, with SA = 99.94 (in percent-
age) and DELTA = 3.4895, thereby confirming the superiority of the PSO-CFLANN to
other approaches.
The comparison of the prediction models used for the COCOMO’81 Dataset is
displayed in Figs. 2, 3, and 4. The efficacy of each model is determined using three
distinct types of graphs. The first graph is a 2Dplot of MAE values for a line graph,
whereas the second is a histogram. On the x-axis of each MAE graph, the number
of COCOMO dataset simulation projects is indicated. The MAE graphs reveal that for
PSO-CFLANN and ISO-CFLANN, the generalization error assessed over the validation
examples increases until project number 20, and subsequently steadily decreases as the
number of projects increases, overcoming the overfitting issue (low training error and
high testing error). By contrast, numerous strategies exhibit overfitting concerns (low
training error and high testing error). The MAE graphs reveals that as the number of
training examples increases, PSO-CFLANN and ISO-CFLANN become better suited to
handle overfitting concerns. Figure 4 depicts the boxplot of all approaches, indicating
that ISO, GA, and BP CFLANNs have more error dispersion than other approaches.
 Tuning Functional Link Artificial Neural Network for SDEE 77

Compared with those of APSO and BP CFLANNs, PSO-CFLANNs’s boxplots have

the shortest interquartile distance and are normally distributed. Therefore, alternative
methods are preferred.

Table 2. Outcomes and comparisons of all techniques on the COCOMO81 dataset

MMRE MdMRE PRED (0.25)

Training Validation Testing Training Validation Testing Training Validation Testing
PSO-CFLANN 6.96E−05 0.000355 0.000259 0.000903 0.000141 8.49E−05 1 1 1
APSO-CFLANN 0.003111 0.00467 0.006586 0.00236 0.001302 0.001905 1 1 1
ISO-CFLANN 0.004959 0.000786 0.010225 0.004335 0.004416 0.010225 1 1 1
Genetic-CFLANN 0.008699 0.01085 0.008252 0.007867 0.001392 0.001929 1 1 1
BP-CFLANN 0.00463 0.004696 0.011901 0.000469 0.000286 0.000655 1 1 1
BP-ANN 0.009619 0.004001 0.024008 0.009502 0.001354 0.002902 1 1 1

Table 3. Outcomes and comparisons of all techniques on the COCOMO81 dataset

SA DELTA
Training Validation Testing Training Validation Testing
PSO-CFLANN 0.99986 0.99928 0.99947 0 2.7627 3.4895
APSO-CFLANN 0.99986 0.99982 0.9999 0 3.4424 2.7082
ISO-CFLANN 0.99981 0.99997 0.99985 0 0.933 1.0638
Genetic-CFLANN 0.99962 0.99959 0.99988 0 0.26906 0.5378
BP-CFLANN 0.99982 0.99982 0.99982 0 0.17986 0.202
BP-ANN 0.99965 0.99985 0.99964 0 0.59613 0.1208

Fig. 2. MAE outcomes and comparison to all techniques on the COCOMO dataset
78 T. R. Benala and S. Dehuri

Fig. 3. MAE outcomes and comparison to all techniques on the COCOMO dataset

Fig. 4. Boxplots of all techniques on the COCOMO dataset

7 Conclusion and Future Work

We detailed the approach for using swarm intelligence techniques, such as PSO, APSO,
ISO, and GA, to optimize the weight parameters of the CFLANN. The approaches antic-
ipate the software development effort by using the optimum weight value acquired by
swarm intelligence techniques and the set of Chebyshev polynomials orthogonal basis
functions selected for the functional expansion of feature vectors. The results of the
 Tuning Functional Link Artificial Neural Network for SDEE 79

empirical study revealed that the strategies of swarm intelligence enhance the perfor-
mance of CFLANN. In most instances, the PSO-CFLANN model provided outcomes
that were comparable or superior to the best results obtained by the APSO-CFLANN,
ISO-CFLANN, genetic-CFLANN, BP-CFLANN, and BP-ANN. The PSO-CFLANN
model’s architectural complexity is considerably less than that of MLP and is the same
as or less than that of the CFLANN with BP. This feature of PSO-CFLANN may encour-
age SDEE researchers to perform additional studies in this domain. Future studies should
concentrate on the symbiotic interaction between swarm intelligence and other soft com-
puting techniques, as well as the concurrent growth of architecture and weights with a
Pareto set of solutions for the next generation of SDEE. Calibration of input patterns
from low to high dimension utilizing clones of HONs such as ridge polynomial neural
networks and pi-sigma neural networks is another research direction.

Acknowledgment. Dr. Satchidananda Dehuri, Professor of Computer Science (Erstwhile P. G.

Department of Information and Communication Technology), Fakir Mohan University would like
to thank SERB, Govt. of India for financial support under Teachers’ Associateship for Research
Excellence (TARE) fellowship vide file no. TAR/2021/000065 for the period 2021–2024.

References
1. Chen, C.H., Lin, C.J., Lin, C.T.: A functional-link-based neurofuzzy network for nonlinear
system control. Fuzzy Systems, IEEE Transactions on 16(5), 1362–1378 (2008)
2. Holland, J.H.: Adaption in Natural and Artificial Systems. The University of Michigan Press,
Ann Arbor (1975)
3. Michalewicz, Z.: Genetic Algorithms + Data Structures = Evolution Programs. Springer,
Heidelberg (2013). https://doi.org/10.1007/978-3-662-03315-9
4. Chiu, N.H., Huang, S.J.: The adjusted analogy-based software effort estimation based on
similarity distances. J. Syst. Softw. 80(4), 628–640 (2007)
5. Huang, S.J., Chiu, N.H.: Optimization of analogy weights by genetic algorithm for software
effort estimation. Inf. Softw. Technol. 48(11), 1034–1045 (2006)
6. Lee, T.T., Jeng, J.T.: The Chebyshev-polynomials-based unified model neural networks for
function approximation. IEEE Trans. Syst. Man Cybern. Part B Cybern. 28(6), 925–935
(1998)
7. Pao, Y.H., Takefuji, Y.: Functional link net computing: theory, system, architecture and
functionalities. IEEE Comput. 76–79 (1992)
8. Boehm, B.W.: Software Engineering Economics, vol. 197. Prentice-Hall, Englewood Cliffs
(1981)
9. Putnam, L.H., Myers, W.: Measures for Excellence: Reliable Software on Time, Within
Budget. Prentice Hall Professional Technical Reference (1991)
10. Huang, X., Ho, D., Ren, J., Capretz, L.F.: Improving the COCOMO model using a neuro-fuzzy
approach. Appl. Soft Comput. 7(1), 29–40 (2007)
11. Shepperd, M., Schofield, C.: Estimating software project effort using analogies. IEEE Trans.
Softw. Eng. 23(11), 736–743 (1997)
12. Auer, M., Trendowicz, A., Graser, B., Haunschmid, E., Biffl, S.: Optimal project feature
weights in analogy-based cost estimation: improvement and limitations. IEEE Trans. Softw.
Eng. 32(2), 83–92 (2006)
13. Heiat, A.: Comparison of artificial neural network and regression models for estimating
software development effort. Inf. Softw. Technol. 44(15), 911–922 (2002)
80 T. R. Benala and S. Dehuri

14. Shin, M., Goel, A.L.: Empirical data modeling in software engineering using radial basis
functions. IEEE Trans. Softw. Eng. 26(6), 567–576 (2000)
15. Oliveira, A.L.: Estimation of software project effort with support vector regression. Neuro-
computing 69(13), 1749–1753 (2006)
16. Minsky, M., Papert, S.: Perceptrons. MIT Press, Cambridge (1969)
17. Hornik, K., Stinchcombe, M., White, H.: Multilayer feed-forward networks are universal
approximators. Neural Netw. 2(5), 359–366 (1989)
18. Cybenko, G.: Approximations by superpositions of a sigmoid function. Math. Controls Sig.
Syst. 2, 303–314 (1989)
19. Guillermo, V.: A distributed approach to neural network simulation program. Master thesis,
The University of Texas at E1 Paso, TX (1998)
20. Zurada, J.M.: Introduction to Artificial Neural System. West Publishing Company, St. Paul
(1992)
21. Beale, R., Jackson, T.: Neural Computing: An Introduction. Hilger, Philadelphia (1991)
22. AlBataineh, A., Kaur, D., Jalali, S.M.J.: Multi-layer perceptron training optimization using
nature inspired computing. IEEE Access 10, 36963–36977 (2022)
23. Man, Z., Wu, H.R., Liu, S., Yu, X.: A new adaptive back-propagation algorithm based on
Lyapunov stability theory for neural networks. IEEE Trans. Neural Netw. 17(6), 1580–1591
(2006)
24. Dehuri, S., Cho, S.B.: A comprehensive survey on functional link neural networks and an
adaptive PSO–BP learning for CFLNN. Neural Comput. Appl. 19(2), 187–205 (2010)
25. Kosmatopoulos, E.B., Polycarpou, M.M., Christodoulou, M., Ioannou, P.: High-order neural
network structures for identification of dynamical systems. IEEE Trans. Neural Netw. 6(2),
422–431 (1995)
26. Giles, C.L., Maxwell, T.: Learning, invariance and generalization in higher-order neural
networks. Appl. Opt. 26(23), 4972–4978 (1987)
27. Pao, Y.H.: Adaptive Pattern Recognition and Neural Network. Addison-Wesley, Reading
(1989)
28. Dehuri, S., Roy, R., Cho, S.B., Ghosh, A.: An improved swarm optimized functional link
artificial neural network (ISO-FLANN) for classification. J. Syst. Softw. 85(6), 1333–1345
(2012)
29. Mirea, L., Marcu, T.: System identification using functional link neural networks with dynamic
structure. In: 15th Triennial World Congress, Barcelona, Spain (2002)
30. Cass, R., Radl, B.: Adaptive process optimization using functional link networks and
evolutionary algorithms. Control EngPract. 4(11), 1579–1584 (1996)
31. Pao, Y.-H., Philips, S.M.: The functional link net learning optimal control. Neurocomputing
9, 149–164 (1995)
32. Shin, Y., Ghosh, J.: Ridge polynomial networks. IEEE Trans. Neural Netw. 6(2), 610–622
(1995)
33. Klasser, M.S., Pao, Y.H.: Characteristics of the functional linknet: a higher order delta rule
net. In: IEEE proceedings of 2nd Annual International Conference on Neural Networks, San
Diago, CA (1988)
34. Elyounsi, A., Tlijani, H., Bouhlel, M.S.: ISAR-image recognition using optimized HONN
by a Metaheuristic algorithm. In: 2022 IEEE 9th International Conference on Sciences of
Electronics, Technologies of Information and Telecommunications (SETIT), pp. 97–103.
IEEE, May 2022
35. Patra, J.C., Van den Bos, A.: Modeling of an intelligent pressure sensor using functional link
artificial neural networks. ISA Trans. 39(1), 15–27 (2000)
36. Qi, Y., Pan, L., Liu, S.: A Lyapunov optimization-based online scheduling algorithm for
service provisioning in cloud computing. Futur. Gener. Comput. Syst. 134, 40–52 (2022)
 Tuning Functional Link Artificial Neural Network for SDEE 81

37. Vilsen, S.B., Stroe, D.I.: Transfer learning for adapting battery state-of-health estimation from
laboratory to field operation. IEEE Access 10, 26514–26528 (2022)
38. Barrera, J., Coello, C.A.C.: A review of particle swarm optimization methods used for mul-
timodal optimization. In: Lim, C.P., Jain, L.C., Dehuri, S. (eds.) Innovations in Swarm Intel-
ligence, vol. 248, pp. 9–37. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-642-
04225-6_2
39. Gad, A.G.: Particle swarm optimization algorithm and its applications: a systematic review.
Arch. Comput. Methods Eng. 1–31 (2022)
40. Nickabadi, A., Ebadzadeh, M.M., Safabakhsh, R.: A novel particle swarm optimization
algorithm with adaptive inertia weight. Appl. Soft Comput. 11(4), 3658–3670 (2011)
41. Backpropagation – Wikipedia, 1 August 2022. https://en.wikipedia.org/wiki/Backpropa
gation
42. Cohen, J.: A power primer. Psychol. Bull. 112, 155–159 (1992)
43. Shepperd, M., MacDonell, S.: Evaluating prediction systems in software project estimation.
Inf. Softw. Technol. 54(8), 820–827 (2012)
44. Mall, R.: Fundamentals of Software Engineering. PHI Learning Pvt. Ltd., New Delhi (2018)
45. Kocaguneli, E., Menzies, T.: Software effort models should be assessed via leave-one-out
validation. J. Syst. Softw. 86(7), 1879–1890 (2013)
46. Kohavi, R., John, G.H.: Wrappers for feature subset selection. Artif. Intell. 97(1–2), 273–324
(1997)
47. Kennedy, J., Eberhart, R.C.: Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks, Perth, Australia, pp. 1942–1948 (1995)
 METBAG – A Web Based Business Application

Nannapaneni Akshaj and B. K. Tripathy(B)

School of Information Technology Engineering, VIT, Vellore, India

[email protected]

Abstract. Digitization has made the online business in the domain of core sector
of metals, minerals and ores more competitive. A lot of web and mobile appli-
cations are available to make the process smooth. However, a user has to follow
the price trend and analyse it in order to make the right choice. This and possible
repetition of the process leads to its slowing down. Also, from the company per-
spective the operations and transactions involved makes the process complex. So,
automation in generating the Irrevocable cooperate purchase order (ICPO) and
Sales and Purchase Agreement (SPA), invoices and Non Circumvention and Non-
Disclosure Agreement (NCNDA) forms is highly recommended. For increment in
business accurate data analytics plays an important role. It is our aim in this paper
to focus on developing an automated web application to frame a transaction inte-
grated with all necessary features which is user-friendly from the letter of intent
till the generation of invoice for the user and an auto generated dashboard for the
company. The process includes the price prediction for the products and also takes
care of the security of user credentials in view using hashing and multi-lingual
jumbled salting.

Keywords: Predictive analysis · LSTM · Regression · F1-Score · SHA-512 ·

Salting · Descriptive analysis

1 Introduction
The competition in the online business in the domain of core sector of metals, minerals
and ores has increased extensively due to the businesses moving digital [1]. Many indus-
trial companies do not have e-commerce websites, they have struggled to fully embrace
this strategy because of the complexities involved, particularly those related to distribu-
tor management, and thus they remain stuck at the pilot stage. Direct e-commerce sales
can help industrial companies, as it gives them more of a connection to end customers
and their needs. The analysis framework has proved to be useful for assessing the power
of every considered methodology to affect the planning of business processes in web
applications. Requirements for methodology to design business processes are represent-
ing component activities, describing possible workflows, defining and managing state
of web transaction, specifying which activities can be suspended or resumed, describe
how two or more users are related in a transaction, specify how content navigation and
operation affect each other, which contents will be provided to user to support execution
of an activity, defining which information objects are affected by executing the activity

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 82–94, 2022.
https://doi.org/10.1007/978-3-031-23233-6_6
 METBAG – A Web Based Business Application 83

and describing how the activity will be customized. Three dimensions of analysis frame-
work are the business requirements, user requirements and system requirements. The
user requirements include a simple user interface, all necessary features and tools for
completing the transaction, security of credentials, etc. Business requirements mainly
include the track of user activity on the platform, business performance using numbers
and statistics and reduced work load for repetitive work. But the web applications in
this sector lack a lot of user requirements like lack of proper UI, integrated analytics
system [9] for the admin because of which they go to third-party tools, credential secu-
rity and other user utilities. For the security of the credentials, SHA-512 hash algorithm
has become popular because of complex hash generated by it [10]. But hashes can be
cracked using guess and check technique [7]. Therefore, salting along with it is used
to enhance security. Also, the price prediction systems are of great use for the business
transactions and can be largely seen in a number of applications. But for this accurate
price based on recent market dynamics are needed to be predicted. Also, there are num-
ber of other utilities which are to be included in order to provide a good experience to
the user [2]. Also, providing a data analytical system [6] and dashboard to the business
admin helps in taking data-driven decisions and prevents them from going to third-party
for the services. So, this application focuses on an automated web application to frame
a transaction integrated with all necessary features which is user-friendly from the letter
of intent till the generation of invoice for the user and an auto generated integrated dash-
board for the company [8]. The process includes the price prediction for the products
based on recent market dynamics and also takes care of the security of user credentials
in view using hashing and multi-lingual jumbled salting.

2 Literature Review
Many industrial companies do not have an e-commerce web application, they find it
difficult to fully implement this strategy due to the complexity involved, particularly
related to the management of distributors, and are therefore stuck in the pilot phase.
Many of them do not meet all user requirements, such as getting real-time prices of
products, user-friendly interface, lack of complete information, ease of activities to be
performed like submission of LOI & ICPO, reordering etc. Often these things become the
reason for loss of business. Direct selling by e-commerce can help industrial companies
because it allows them to be closer to the end customer and his needs.

2.1 Gaps and Solutions

Gaps in the current system are:

• Lack of user utilities

• Lack of integrated analytics systems because of which businesses go for third-party
tool, thus increasing their expenditure.
• Price prediction for the products being sold.
• Simple user interface.
• Security of the user credentials.
84 N. Akshaj and B. K. Tripathy

So, this system takes care of the above-mentioned gaps and develops a system with
the objective of:

• Develop metal price prediction model using the market price values to get relevant
predicted price values.
• Using techniques like SHA-512 hashing [3] and multi-lingual jumbled salting for
security of the user credentials [4]
• Use of technologies like bootstrap, chart.js, HTML, CSS and jQuery for designing
and developing a good user interface.
• Applying the SQL techniques and plotting techniques to extract the valuable infor-
mation from the collected user data and putting it in the form of dashboards and
tables.

If provided with right computational power, the system can be implemented to a

number of different sectors and the number of domains can also be increased. Also, the
prediction model can be scaled to any number of domains, even then producing highly
accurate results.

2.2 Deep Neural Networks (DNN) and LSTM

Neural networks are uncertainty based models, which have been developed to mimic
the functionalities of the human brain. Of late Deep Neural Network which are neural
networks with deep learning concept embedded into them (In fact the number of hidden
layers have been increased to a good number) ([12, 21]. There are several types of
DNNs like Convolutional Neural Networks (CNN) [13], recurrent Neural Networks
(RNN) [12], generative Adversarial Neural Networks (GAN) [12] and so on. There are a
number of applications of these nets, mainly for image classification ([14–19]). LSTM is
a special kind of recurrent neural network capable of handling long-term dependencies
([20, 21]). In this work we use LSTM in the price prediction component.

3 Architecture of the System

The architecture of the proposed system comprises of two main components; from user
side and from admin side.

3.1 Architecture of the User Side of the System

The user side system architecture is provided in Fig. 1. The below figure explains the
flow user follows in order to complete a transaction. This also dives the flow of data in
and out of the database. Thus giving a brief about the user side of the application.

Description of Modules

• About: This module describes about the company, their operations and about the CEO
of the company.
 METBAG – A Web Based Business Application 85

Fig. 1. Architecture of the user side of the system

• Pricing: This module displays the price trends. The module also gives the option of
price conversions to desired currencies along with the option of price prediction of
the product chosen by the user and the price prediction using the different domains
of provided based on the user input. The domains are the different products and each
domain has a separate data of prices which is a time-series data, which would be the
input for the model after processing.
• Metals: This module displays all the products with the analysis of each product and
the query tab.
• Query: This module is for the user to clear his queries regarding a particular product
that he wishes to know about.
• Registration: This module is for the user signup. The entered password is hashed and
also a random multi-lingual salt generated is also hashed and then both the hashes
are concatenated in alternative jumbled manner to generate a 256 characters long
password string. Other details like the username and email are also collected.
• Login: The username and password are required for authentication. Password entered
is hashed and then the password stored in database is extracted and sliced to match
with the entered password hash.
• LOI: After selecting product the user fills the LOI form with the LOI PDF. After
submitting, the user will receive Sales and Purchase Agreement and the ICPO and
Sales and purchase Agreement link after reviewing LOI.
• ICPO: This module is for the order submission. The user enters all the details of the
order and submits the order. After placing the Order, the user will receive NCNDA
form and the invoice.
• Reorder: The user can click on the reorder tab against each order displayed in the user
login. After clicking, the user gets the auto filled ICPO form and the user can place
the order.
86 N. Akshaj and B. K. Tripathy

3.2 Architecture of the Admin Side of the System

The admin side system architecture is provided in Fig. 2 The below figure explains the
flow user follows in order to complete a transaction. This also dives the flow of data in
and out of the database. Thus giving a brief about the admin side of the application.

Fig. 2. Architecture of the system- Admin side

Description of the Modules

• Registration: This module is for the admin signup. The entered password is hashed
and also a random multi-lingual salt generated is also hashed and then both the hashes
are concatenated in alternative jumbled manner to generate a 256 characters long
password string. Other details like the username and email are also collected.
• Login: The username and password are required for authentication. Password entered
is hashed and then the password stored in database is extracted and sliced to match
with the entered password hash.
• Total Business: This module shows the amount of each product sold and amount of
revenue total generated by each product
• LOI Track: This module shows the list of all the LOI pending to be converted with
the download option of each LOI file.
• Dashboard: This is the visualization module displaying the business analytical
dashboard

4 Workflow Diagrams

We provide the workflow diagrams of the system in two components in this section
(Fig. 3 and Fig. 4).
 METBAG – A Web Based Business Application 87

Fig. 3. Workflow diagram- Users’ side

Fig. 4. Workflow diagram admin side

5 Procedures

The entire procedure is divided into 3 parts; price prediction, password security and
dashboard. We provide the three procedures below.
88 N. Akshaj and B. K. Tripathy

5.1 Procedure: Price Prediction

STEP 1: Takes the input as the product name from the user.
STEP 2: Then based on the input, the data about the specific product is extracted
from the dataset.
STEP 3: Data in the dataset is extracted from open-source web data.
STEP 4: All the data is pre-processed and store in the desired format by removing
all the null values.
STEP 5: Data is split as 80% for training data and 20% for testing.
STEP 6: The step value is set as 40, i.e., for predicting a value, previous 40 values
are considered.
STEP 7: The LSTM model is [11] created with 2 hidden layer and one output node.
Dropout for each layer is set as 0.3.
STEP 8: The model is then compiled using adam optimizer and loss used is mean
squared error.
STEP 9: number of epochs are set as 50.
STEP 10: The testing data is prepared by using the last 20% of the data concatenated
with the last 40 values of the training dataset.
STEP 11: Then the testing data is fed into the LSTM model and the future price is
displayed as output.

5.2 Procedure: Password Security

STEP 1: The password entered by the user is hashed using the SHA-512 hash algorithm.
STEP 2: From a string containing multilingual characters, integers and special
symbols, a random string is generated.
STEP 3: The random generated string is hashed using SHA-512 hash algorithm and
then is used to salt the password hash by appending it in between the password hash in
alternate manner.
STEP 4: This results in a hash string of 256 characters long which is stored in the
database.

5.3 Procedure: Dashboard

STEP 1: The data stored in the database generated by the user transaction on the web
application is used for generating the analytical dashboard.
STEP 2: The data present in the database is computed using the SQL queries and
then that data is transferred to chart where those data points are plotted/displayed in
appropriate manner.

6 Result Analysis
6.1 Price Prediction
The performance metrics used are Confusion matrix and the F1-Score. Confusion matrix
consists of:
 METBAG – A Web Based Business Application 89

• True Positive: a true positive is an outcome where the model correctly predicts the
positive class.
• True Negative: a true negative is an outcome where the model correctly predicts the
negative class.
• False Positive: A false positive is an outcome where the model incorrectly predicts
the positive class.
• False Negative: False Negatives (FN) are negative outcomes that the model predicted
incorrectly.

The Technique used was regression, but for testing on the basis of the confusion
matrix, the output has to be of the classification type. So, to convert the output and
original array of prices to binary form, the elements of the arrays were normalized. Then
the normalized values were rounded to the nearest integers, therefore all the values were
either 0 or 1 using the formula
 
Element of the Prediction Price Array - min(Prediction Price Array)
round (1)
max(Prediction Price Array) - min(Prediction Price Array)
The F-score, also called the F1-score, may be a measure of a model’s accuracy on
a dataset. It is used to evaluate binary classification systems, which classify examples
into ‘positive’ or ‘negative’.
TP
F1 − score = (2)
TP + 2 (FP
1
+ FN )
Value of the F1-Score is between 0 and 1. Higher the value better is the model.
We use python language for the implementation of the system [5]. The dataset used is
presented in Fig. 5.

Fig. 5. Dataset

The screen shots of user input is as presented in Fig. 6.

The building of the model is as in Fig. 7.
Figure 8 shows the screen output and Fig. 9 shows the performance matrix.
90 N. Akshaj and B. K. Tripathy

Fig. 6. User Input

Fig. 7. Model Building

 METBAG – A Web Based Business Application 91

Fig. 8. Output on the Screen

Fig.9. Performance Metrics

6.2 Password Security

The resultant password after salting and hashing is 256 characters long and padded with
the noisy salt characters. Although Hash is irreversible, the attackers try “guess and
Check” method to crack the password. They hash a random generated string and match
with the hashed password. So as this password will now have the noisy characters, this
will become more difficult to guess the password string.
The user registration is shown in Fig. 10, The password generated salt and password
using hashing is shown in Fig. 11. A resultant password is shown in Fig. 12.
92 N. Akshaj and B. K. Tripathy

Fig. 10. Registration

Fig. 11. The generated salt and password hashes

Fig. 12. The resultant password

7 Conclusions and Future Work

We developed a system for a complete business transaction with all the necessary mod-
ules available to the user at a single place. It reduces the tedious task for the user to go for
different websites in search of market price research and analysis, which many a times
leads to losing the client. Also, using regression model and LSTM, which is the best when
it comes to dealing time series data, the prices prediction of the metals helps the user to
get the insight of the next price, whether it would increase or decrease. It is seen that
the model performance evaluated using F1-score is coming to be 0.963855421686747
for iron ore, 0.997289972899729 for COMEX gold, 0.9642857142857143 for plat-
inum, 0.90032153408306 for palladium and 0.9939393939393939 for Silver 5000oz.
The security of the credentials is also a major concern. providing security using SHA-
512 hashing which itself is a highly preferred security practice and in addition to that
 METBAG – A Web Based Business Application 93

multi-lingual jumbled salting provides a higher security to the credentials. An integrated

analytical dashboard provides ease to the administration to take data-driven decisions by
sharing the analysis with the stakeholders. The visualization dashboard giving a holistic
view of the business makes it easier for them to under-stand. In future, the data collected
from the user transactions to be used for the prediction of future business growth using
the machine learning models. Larger dataset can be created for further enhancing the
accuracy and performance of the models. Text extraction script for content of document
submitted by the user to a get the summary of the document.

References
1. Distante, D., Rossi, G., Canfora, G.: Modelling business processes in web applications: an
analysis framework. In: Proceedings of the 2007 ACM symposium on applied computing,
pp. 1677–1682 (2007)
2. Verma, J., Shahrukh, M., Krishna, M., Goel, R.: A critical review on cryptography and hashing
algorithm SHA-512. Int. Res. J. Modernization Eng. Technol. Sci. 03(12), 1760–1764 (2021)
3. Sumagita, M., Riadi, I., Sh, J.P.D.S., Warungboto, U.: Analysis of secure hash algorithm
(SHA) 512 for encryption process on web-based application. Int. J. Cyber-Security and Digital
Forensics (IJCSDF) 7(4), 373–381 (2018)
4. Kharod, S., Sharma, N., Sharma, A.: An improved hashing-based password security scheme
using salting and differential masking. In: 2015 4th International Conference on Reliability,
Infocom Technologies and Optimization (ICRITO), (Trends and Future Directions), pp. 1–5.
IEEE (2015)
5. Nagpal, A., Gabrani, G.: Python for data analytics, scientific and technical applications. In:
2019 Amity international conference on artificial intelligence (AICAI), pp. 140–145. IEEE
(2019)
6. Thomas, D.M., Mathur, S.: Data analysis by web scraping using python. In: 2019 3rd Inter-
national conference on Electronics, Communication and Aerospace Technology (ICECA),
pp. 450–454. IEEE (2019)
7. De Guzman, F.E., Gerardo, B.D., Medina, R.P.: Implementation of enhanced secure hash
algorithm towards a secured web portal. In: 2019 IEEE 4th International Conference on
Computer and Communication Systems (ICCCS), pp. 189–192. IEEE (2019)
8. Manjushree, B.S., Sharvani, G.S.: Survey on Web scraping technology. Wutan Huatan Jisuan
Jishu 16(6), 1–8 (2020)
9. Appelbaum, D., Kogan, A., Vasarhelyi, M., Yan, Z.: Impact of business analytics and
enterprise systems on managerial accounting. Int. J. Account. Inf. Syst. 25, 29–44 (2017)
10. Gupta, P., Kumar, S.: A comparative analysis of SHA and MD5 algorithm. Int. J. Comp. Sci.
Info. Technol. 5(3), 4492–4495 (2014)
11. Liu, S., Liao, G., Ding, Y.: Stock transaction prediction modelling and analysis based on
LSTM. In: 2018 13th IEEE Conference on Industrial Electronics and Applications (ICIEA),
pp. 2787–2790. IEEE (2018)
12. Bhattacharyya, S., Snasel, V., Hassanian, A.E., Saha, S., Tripathy, B.K.: Deep learning
research with engineering applications. De Gruyter Publications (2020). https://doi.org/10.
1515/9783110670905
13. Maheswari, K., Shaha, A., Arya, D., Tripathy, B.K., Rajkumar, R.: Convolutional neural
networks: a bottom-up approach. In: Bhattacharyya, S., Hassanian, A.E., Saha, S., Tripathy,
B.K. (eds.) Deep Learning: Research and Applications, pp. 21–50. De Gruyter Publications
(2020). https://doi.org/10.1515/9783110670905-002
94 N. Akshaj and B. K. Tripathy

14. Bose, A., Tripathy, B.K.: Deep learning for audio signal classification. In: Bhattacharyya, S.,
Hassanian, A. E., Saha, S., Tripathy, B. K. (eds.) Deep Learning: Research and Applications,
pp. 105–136. De Gruyter Publications (2020). https://doi.org/10.1515/9783110670905-00660
15. Adate, A., Tripathy, B.K.: A Survey on Deep Learning Methodologies of Recent Applications.
In: Acharjya, D.P., Mitra, A., Zaman, N. (eds.) Deep Learning in Data Analytics. SBD, vol.
91, pp. 145–170. Springer, Cham (2022). https://doi.org/10.1007/978-3-030-75855-4_9
16. Kaul, D., Raju, H., Tripathy, B.K.: Deep Learning in Healthcare. In: Acharjya, D.P., Mitra,
A., Zaman, N. (eds.) Deep Learning in Data Analytics. SBD, vol. 91, pp. 97–115. Springer,
Cham (2022). https://doi.org/10.1007/978-3-030-75855-4_6
17. Tripathy, B.K., Parikh, S., Ajay, P., Magapu, C.: Brain MRI segmentation techniques based
on CNN and its variants. In: Chaki, J. (ed.) Brain Tumor MRI Image Segmentation Using
Deep Learning Techniques, Chapter-10, pp. 161–182. Elsevier publications (2022). https://
doi.org/10.1016/B978-0-323-91171-9.00001-6
18. Bhardwaj, P., Guhan, T., Tripathy, B.K.: Computational biology in the lens of CNN, studies
in big data. In: Roy, S.S., Taguchi, Y.-H. (eds.) Handbook of Machine Learning Applications
for Genomics, (Chapter 5), vol. 103 (2021). ISBN: 978-981-16-9157-7 496166_1_En
19. Prabhavathy, P., Tripathy, B.K., Venkatesan, M.: Analysis of diabetic retinopathy detection
techniques using CNN models. In: Mishra, S., Tripathy, H.K., Mallick, P., Shaalan, K. (eds.)
Augmented Intelligence in Healthcare: A Pragmatic and Integrated Analysis. Studies in Com-
putational Intelligence, vol. 1024, pp. 87–102. Springer, Singapore (2022). https://doi.org/10.
1007/978-981-19-1076-0_6
20. Sandhu, S.S., Tripathy, B.K., Jagga, S.: KMST+: A K-Means++-Based Minimum Spanning
Tree Algorithm. In: Panigrahi, B.K., Trivedi, M.C., Mishra, K.K., Tiwari, S., Singh, P.K.
(eds.) Smart Innovations in Communication and Computational Sciences. AISC, vol. 669,
pp. 113–127. Springer, Singapore (2019). https://doi.org/10.1007/978-981-10-8968-8_10
21. Adate, A., Tripathy, B.K.: S-LSTM-GAN: Shared Recurrent Neural Networks with Adver-
sarial Training. In: Kulkarni, A.J., Satapathy, S.C., Kang, T., Kashan, A.H. (eds.) Proceedings
of the 2nd International Conference on Data Engineering and Communication Technology.
AISC, vol. 828, pp. 107–115. Springer, Singapore (2019). https://doi.org/10.1007/978-981-
13-1610-4_11
 Designing Smart Voice Command Interface
for Geographic Information System

Shivam Pant1 and Narayan Panigrahi2(B)

1 Intern, Centre for Artificial Intelligence and Robotics, Post Graduate Student,
Gautam Buddha University, Noida, India
[email protected]
2 Scientist-G, Centre for Artificial Intelligence and Robotics, Bangalore, India

[email protected]

Abstract. Voice based command activation has gained momentum as a feature

of user interface replacing the mouse or pointing device-based menu selection.
Among various software driven systems such as internet of things, search engines
and robotic systems use the voice command interface as the primary interac-
tion between the user and the system keeping the conventional user interface as
secondary. Voice command interface has much pay offs and advantages over con-
ventional menu driven user interface or mouse-driven user interface. This paper
presents the research landscape on voice command interface and proposes a design
of voice-based interface for a GIS (Geographic Information System). Further we
compare the proposed voice command interface with open source and COTS
(Commercial Off-The-Shelf) voice command interface models. The proposed
voice interface model has significant reduction in the word error rate by 3.49%
compared to existing models.

Keywords: Virtual assistant · VOS viewer · ASR · Computer science · IOT · GIS

1 Introduction

The voice-based command is integrated in most of the IOT based device these days for
the ease of the user. The purpose to work in design and implementation of such a device
for defence is because the normal graphic user interface (GUI) and mouse that is used,
usually needs the attention of the user. There are many times when the user might be
stuck in a situation where the user may not be capable enough to focus on all the tasks
at the same. Some of the scenarios to name are: - while navigating and driving, while
handling different kind of machineries at the same time. In this moment it’ll be better if
the user can get the assistance side by side so that the cognitive abilities of user can be
divided fairly towards the more important things that needs to be focused upon at that
moment of emergency.
Some of the key applications of voice assistant are as follows: -
Unmanned aerial vehicle operations: - The voice command can play a very effec-
tive role in Unmanned aerial vehicle operations by simply making the Unmanned aerial

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 95–112, 2022.
https://doi.org/10.1007/978-3-031-23233-6_7
96 S. Pant and N. Panigrahi

vehicle operate several functions just by uttering the respective query attached to that
function instead of manually maneuvering over the physical console.
Pilot Communication & autopilot with the voice-enabled console: - There are
times when the Pilot might be switching to autopilot mode and would require accessing
some emergency buttons even in the autopilot mode. So, in such a tough situation, the
voice command interface can come handy so that the pilot need not have to think much
on dealing with the physical panel rather he would just utter the command to initiate the
whole process smoothly.
Specially abled person: - This voice command interface can work as a boon for
almost every specially abled individual so that they can make use of the system to
perform several functions where in no physical assistance would be required from the
user’s end.
Health check-in and communication: - This voice command interface can also
be integrated with a health-based application which can easily maneuver the cardinal
state of a person in a situation where they might be collapsing or on the verge of getting
collapsed more over the interface can detect and directly communicate the same with
their emergency contact.
The voice assistant comes with the ability to render through the queries of user within
seconds even when there is a lot of disturbance and noise. The organization for which
even one second can cost a life and pride of the country, this becomes a crucial role to
leave the minor tasks in an automated mode and focus on the tasks that really needs a
cognitive attention to work with. The assistant can really work it out at times when the
human needs to open a particular application say “maps” or needs to see how much time
it will take to hop on from not 1 but 5 different locations at the same time by rendering
over the maps and providing the user an output that is backed by facts. There is no limit
to what a voice assistant can be capable of doing for the user.

1.1 Review of Literature on Voice Command Interface

A search of research literature on “dimensions.com” was carried out. It has resulted in

1500 + numbers of papers developed in Java programming language; VOS viewer is
simply a tool which is used to create maps that are based on network data. It can create
three types of map-based visualizations, which are as follows:

• Network Visualization
• Overlay Visualization
• Density Visualization

The 3 visualizations that was created by using CSV file of related topic of the research
will be shown and then a theoretical inference would be generated out of it.
In Fig. 1, it is shown that out of all the keywords present in form of text in our
publications, only these are the relevant keywords that have occurred more than ten times
in every paper. The relevance score has been generated with the help of a machine learning
algorithm that is working in the backend of the model [1]. The Network map diagram of
the review literature on voice command interface gives the collaborative picture of area
 Designing Smart Voice Command Interface for Geographic 97

Fig. 1. Network Map-based Visualization of most co-occurring keywords

of research on voice command interface versus other technologies. Prominent amongst

the other area of study are tabulated below (Table 1).
The outcome which came after data exploration of the top cited paper in the field of
research proves that the all the five keywords present in the Fig. 2 are very much related
and relevant to the voice assistant and is in constant integration with the smart voice
based assistant. [2].

2 Design and Implementation

After describing the architecture, the model’s implementation will be thoroughly dis-
cussed under this heading, along with the scenarios that have been created to check the
efficiency of the same.

2.1 Review of Literature on Voice Command Interface

The dataset for the voice-based command model NAKSHA includes every query-based
command utilized by the model. The model has been evaluated to see how it performs
using examples of instructions that relate to real-world situations and can be used to
generate real-world scenarios.
Following are the commands for the smart voice-based system:

• Wake-up command will help the system and will listen and greet with a given set of
commands, in this case, which is “Hello Naksha”.
• For online search for a particular website the query named “Website” is used wherein
the user can speak “open the website” along with the respective website that user wants
the assistant to open. e.g., “Naksha Open Website Facebook” as a result assistant will
open.
98 S. Pant and N. Panigrahi

Table 1. Representation of Area of research with the number of publications affiliated

Area Of Research Publication Count in Common with % of 267

Voice Assistant
Computer Science Theory Methods 92 34.457
Computer Science Information 60 22.472
Systems
Engineering Electrical Electronic 59 22.097
Computer Science Artificial 47 17.603
Intelligence
Computer Science Cybernetics 38 14.232
Computer Science Interdisciplinary 33 12.36
Applications
Telecommunications 32 11.985
Computer Science Software 27 10.112
Engineering
Acoustics 14 5.243
Business 14 5.243
Imaging Science Photographic 12 4.494
Technology
Medical Informatics 12 4.494
Health Care Sciences Services 11 4.12
Psychology Multidisciplinary 9 3.371
Computer Science Hardware 8 2.996
Architecture
Engineering Multidisciplinary 8 2.996
Geriatrics Gerontology 6 2.247
Gerontology 6 2.247
Psychology Applied 6 2.247
Public Environmental Occupational 6 2.247
Health
Education Educational Research 5 1.873
Engineering Biomedical 5 1.873
Engineering Industrial 5 1.873
Linguistics 5 1.873

• Similarly, there’s an “Open maps” query which once spoken will redirect the user
to the map location of the Centre for Artificial Intelligence and Robotics, DRDO,
Bangalore that is where this model was developed.
 Designing Smart Voice Command Interface for Geographic 99

Fig. 2. Relationship among keywords

• If the user wants to shut down the system, then the user can simply speak “Bye” and
the voice assistant will greet the user goodbye and the system will stop.
• For additional activities a query named “music” is also included, wherein a user can
request to play a track of their choice and the assistant will play the track for the user
which for now is targeted in a location inside a directory of the local machine.
• There’s another feature where if a user is interested in closing or opening a new tab
while the user is working online, can do so by just speaking “open new tab” and “close
this tab”.
• “Introduction” is another query wherein the assistant will brief the user about its origin
and the purpose for which it is built (Xie et al., 2019) [5].

This was the brief theoretical knowledge about the dataset i.e., the query-based voice
commands that is being used in this proposed model.
Designing a voice-based command interface for a system has many dependencies
and factor. The prime factor being: -

(a) Response time: - The response time is an efficient parameter in the designing of the
voice command interface, which decides that how quick the interface is responding
to the given query. The Naksha AI has a response time of 5 s in which it listens to
the query and responds with an answer within 5 s time.
(b) Error rate: - The error rate is simply defined as the number of errors divided by the
total number of words. This parameter defines the efficiency of the voice assistant.
Lower the word error rate more structured would be the input, hence, the output of
100 S. Pant and N. Panigrahi

the voice assistant would be more reliable. The Naksha AI has a word error rate of
16.5% respectively.

Below is provided the Table 2 which comprises of the dataset for smart voice-based
command interface for a menu-driven system describing all the functions in brief for the
given query-based common system pointwise.

Table 2. Dataset for voice-based command system

S No Query Function
1 Hello Wake word & Assistant will greet
2 Introduce The Assistant will tell the user about it’s origin & purpose
3 Bye The assistant will shutdown the process
4 Open Maps Assistant will open the base map location for CAIR, DRDO
5 Youtube Search The assistant will directly youtube search the given phase
6 Open New Tab Assistant will open a new tab in google.com
7 Close this tab Assistant will close the existing tab on google.com
8 The time The assistant will tell us the time according to IST
9 Website The assistant will open a website
10 Music The assistant will play the music from a local directory inside the
system
11 Google Search It will google search the phrase that user said

2.2 Modules Used

Datetime Module

• In python programming language the python datetime module can be used to provide
a specific date and time in the python code.
• The classes that are supplied by the datetime module provide functions that are used
to deal with dates and times.
• The Date Time module is an inbuilt module of python.

Web Browser Module

• This module allows user to access content from web easily.

• This module with the help of open() function allows user to open a certain application
which is accessible online.
 Designing Smart Voice Command Interface for Geographic 101

• With the help of this module we can easily open and close a particular application
through open() and close() function.

Pyttsx3 Module

• It is a python-based module.
• It is used to convert text into speech.
• Pyttsx3 works offline unlike other libraries which require intern connectivity.
• This module is compatible with both versions, python2 and python3.
• This particular module is supported by two voices i.e., male and female for windows.

2.3 Methodology

The explanation of the proposed voice command-based model for a menu driven system
that has been constructed end to end on visual studio code has been explained through
a flow chart.

Fig. 3. Block diagram of the proposed model

Therefore, a reader can comprehend how to construct a comparable type of model

by analysing Fig. 3. The flowchart demonstrates each step of the model development
process, from the input of the given set of commands as depicted from the Table 2, to
importing the libraries that have been explained in detail in Subsect. 2.2, implementation
of the model explained in Subsect. 2.4 and further analyzing and testing the model in
different scenarios explained in detail in the Subsect. 3.1. For a comparative study please
102 S. Pant and N. Panigrahi

refer the Subsect. 3.3. Therefore, from comprehending the issue statement to creating a
dataset, to implementing and coding the model to improve an existing automatic speech
recognition system [6].
The above methods will now be illustrated step by step, along with snippets from
the programs.

2.4 Implementation

Step 1: First, understand the problem statement, which was to create a voice-based
command system that could function without the need to define each command line by
line accurately, instead providing a query for a function to be executed.
Step 2: Flowchart of the proposed model has been created to better understand the logic
behind creating it and the complexities that needs to get resolved with it.
Step 3: Installed the respective libraries on Visual Studio Code where the model has
been coded (Fig. 4).

Fig. 4. Importing Libraries

Step 4: After importing the libraries, now the voice module for the assistant is imported
with which it will respond to the user, there are several voice-based application pro-
gramming interface that can be used for the voice command-based model. Some of the
most popular models that are used are sapi5, kaladi, deep speech and Alexa skills.
In Fig. 5, the voice architecture of the server application programming interface has
been displayed for the better understanding of the working of it inside the voice-based
command system. Here the managed app and native app both are interacting with the
application programming interface and system.speech* which further interacts with a
speech engine that recognizes and synthesizes the voice signal programming interface
and system.speech which further interacts with a speech engine which recognizes and
synthesize the voice signal (Fig. 6).
Step 5: The function has been defined which will help the assistant understand how to
speak after understanding the voice command it can speak and execute the task according
to the given rules.
 Designing Smart Voice Command Interface for Geographic 103

Fig. 5. Diagram of voice architecture

Fig. 6. Importing voice

Step 6: Wishme() function has been created to greet the user on the basis of IST i.e.
Indian standard time.
Wherein time-wise greeting is assigned to the user:

• If the time is greater than 0:00 h and less than 12:00 then greet with “Good Morning”.
• If the time is greater than or equal to 12:00 h and less than 18:00 then greet with
“Good Afternoon”.
• Else greet with “Good Evening”

These are the following program that has been coded inside the wish me function.
Therefore, whenever the code will run this would be the initial voice commands that
the assistant will use to greet the user after which it’ll take the further queries of the user
and execute them. This is one of the essential functions that need to get included inside
104 S. Pant and N. Panigrahi

the assistant for a better human interaction as the first thing all humans do is greet each
other. In order to give it a touch this function has been created.
Step 7: In this step, the execution of the given tasks will take place as shown in Fig. 7.
With the help of the task execution function. This function will let the assistant know
which task needs to get executed when the user calls up that query. Inside this function
user can add a different kind of task according to the required use case and then can
provide the respective commands that need to be programmed for the task to get executed.
As soon as the task gets executed the assistant awaits another task again.

Fig. 7. Task execution function

Step 9: In this step we will simply define different kind of smart voice-based
command/query- based command that we will be using as an input to train and test
the model. The model will be provided with several queries and will be tested in a noisy
environment to check how quick it understands and interprets the voice signal to generate
the given query in the output.
Some of the given voice-based command on this model are (Fig. 9):

• “Hello”,“Music”,“Open Maps”,“YouTube search”,“Google search”,“Wikipedia

search”,“Open GIS”
 Designing Smart Voice Command Interface for Geographic 105

Fig. 8. Query based command

Fig. 9. Chrome automation query

The Fig. 8 demonstrates the takecommand() function, different kind of queries are
taken for the assistant and then respective responses for every query is given so that the
assistant respond when that particular query is called upon by the user at the time of
testing the model. Figure 10 gives a demonstration of all the chrome automation query
that has been coded in the model so that whenever the user is working with the chrome
engine then the user can use this command.
As a result, this is the entire scenario, starting with how the problem statement was
understood and how the dataset for the problem statement was selected to build the
functions that allow the assistant to speak and understand any command the user gives.
106 S. Pant and N. Panigrahi

The query here is the voice-based commands which get executed when the user says
them in the system.
Therefore, this paper gives an overview of the design and implementation that took
place in this voice-based command interface for a menu-driven system.

3 Results and Discussion

The term “voice recognition devices” is not new; in fact, these devices have been around
since the late 1980s (Terzopoulos et al.,2020) [7]. The training that has been done on
the model with increasingly modern datasets at each and every iteration has been what
has ultimately contributed to it achieving better results over time. While the model is
being trained on a substantial amount of data, it is able to comprehend the distinctions
that exist between each and every phrase, word, and phenome.

3.1 Testing Model by Creating a War Zone like Environment

In order to check the reliability, and responsiveness of the model and if it is feasible to
get deployed for the use in defense sector, we create a warzone like environment to test
out model.
Inside a room an environment has been created which is somewhat similar to a war
zone where there is a lot of disturbance.

• An audio27 is adapted from YouTube28 of a war sound which is played in full volume
on the mobile phone that is, iPhone 13, which generates 104.7dBA ~ 105dBA when
it is at it’s full volume.
• The audio was played for approximately five minutes.
• Meanwhile, two commands were given to the assistant first “Open Maps” and other
“Open new tab”
• The model was able to interpret both the commands in that scenario and was able to
execute the tasks.
• The snippet of the same is shown below in Fig. 8 where the model successfully opened
the base location in the map along with the other 2 new tabs.

Therefore, we have achieved the end result, by testing our model in this environment.
 Designing Smart Voice Command Interface for Geographic 107

Fig. 10. Result of warzone like experiment

3.2 Spectrogram and Waveform Samples for Spoken Voice

The spectrogram in Fig. 11, and Fig. 13 and waveform in Fig. 12, and Fig. 14 depicts
the waveform of the audio signal and audio quality of the voice signal being generated
from the assistant respectively. The color shade in Fig. 11 and Fig. 13 shows the heat
in the voice signal generated by the assistant. This gives an idea of how impactful the
generated voice signal is.

Fig. 11. Spectrogram of hello command

This measure has been evaluated and visualized to understand the audio quality
and the audibility of the assistant’s speech. The frequency on the y axis of the Fig. 14
108 S. Pant and N. Panigrahi

Fig. 12. Waveform of hello command

Fig. 13. Spectrogram of google website query

describes the frequency rate for the voice signal that is being generated when the assistant
responds to the user on a google website query. The Fig. 13 shows the voice signal when
the google website query response is being uttered by the assistant. The Fig. 14 simply
defines when and with how much pressure the voice is being generated from the assistant.
This will help to gather the information about the quality of audio that is coming out of
the assistant so that when it will work in a real-time scenario it works correctly.
 Designing Smart Voice Command Interface for Geographic 109

Fig. 14. Waveform of google website query

3.3 Comparative Analysis Based on Word Error Rate

The word error rate is a parameter which is used to check the efficiency of a speech
recognition model that is being used in several voice assistant such as Kaladi, Siri,
Alexa, Cortana, and Naksha. The Mathematical formula to calculate the basic word
error rate is as follows:
Selection + Inserton + Deletion
Word − Error − Rate = (1)
TotalNumberofWordsSpoken
Here, the word error rate of the model has been compared with an existing model in the
organisation to show how much the proposed model has progressed in contrast to the
earlier automatic speech recognition model.

Fig. 15. Graph containing word error rate %

Figure 15 shows that the word error rate of the kaladi based automatic speech recog-
nition model when trained and tested on the dataset is getting a word error rate of 25
percentage and the same when did with the Naksha- smart voice-based assistant for the
110 S. Pant and N. Panigrahi

menu driven system, the word error rate is being reduced by a significant value of 8.49
percentage, dropping down to 16.51 percentage. The smart voice-based assistant has
also been tested on difficult scenarios where the voice-based model has performed well
and provided effective results in the given dataset. Hence it can be seen that the current
model fulfils the respective use cases for which it was meant to be built.

Fig. 16. Line chart displaying the difference in word error rate

From these charts in Fig. 15 and Fig. 16 it can easily be inferred that the kaladi
based automatic speech recognition system which was used till now had a word error
rate of 25%, with which the assistant was facing a lot of issue in comprehending the
voice command in noisy environment.
On contrary, with the development of Naksha - voice assistant this parameter has
been reduced to 8.49% which is a significant change and upliftment in the features
of the already built automatic speech recognition model (ASR) earlier. The purpose
of building an advanced version of this kaladi based system was to achieve a system
which could understand the query-based command in the scenarios where there is a
lot of disturbance and ambient noise and can provide the better results to the user. The
spectrogram in Figs. 11 and 13 proved that the voice quality of the assistant is up to the
mark as the voice sample is under the frequency range of 20–20000 Hz of voice band
(Bentley et al.,2018) [8]. The contours formation in the voice sample of Fig. 11 and
Fig. 13 proves that as and when the sample is being executed every single detail of the
voice is clearly visible which tends towards the conclusion that user can comprehend
the voice of the assistant clearly.
 Designing Smart Voice Command Interface for Geographic 111

4 Conclusion
The main goal of this research was to build a model that can easily work on a machine
that is safe and secure and for working in geographic information system data. This
voice-based model has many pros if compared with the already existing model in the
organization. The speech model when tested in different scenarios could easily compre-
hend what the user said and was able to execute the respective command given by the
user. The word error rate of the proposed model was approximately 8% less than the pre-
vious model which made it easier to interpret the words quickly (Belanche et al.,2019)
[9].
The best part of this model is that many of the things that the user was not able to
change earlier now have full ability to change according to user’s will. The model can
be re-created according to the given use case.
The automatic speech recognition30 model is not a new concept rather it has been
in existence since the 1950s but as and when the time passed now every industry or
organization simply uses it according to their problem statement and needs. Similarly,
this research has been created to fill the void of having a smart voice-based assistant that
could fulfil the needs of defence organizations (Yan et al.,2022) [10].
Further on I’d like to start by quoting one of the famous quotes said by famous
scientist Albert Szent-Györgyi “Research is to see what everybody else has seen, and
to think what nobody else has thought”. Moving on the same line it can be figured that
some of the work that could be done to improve the voice assistant is by integrating
a layer of authenticity so that nobody instead the authoritative user can only have the
access to whole system just by the tone of voice which will be a far more advanced step
in this field where passcode, fingerprint or retina-based authentication are in use that can
easily be breached or manipulated by allies for an illegal purpose (Hansen et al.,2015)
[3]. With the help of a powerful voice-based authentication system soon one can not
only automate and multitask but can also safely work even in the tough environment
where there’s a risk of leaking of the data (Kepuska et al.,2018) [4].

Acknowledgement. Authors are thankful to Director, CAIR-DRDO for his support to carry out
this research. This research did not receive any specific grant from funding agencies in the public,
commercial, or not-for-profit sectors.

References
1. Panigrahi, N., Mohanty, S.P.: Brain Computer Interface. CRC Press (2022)
2. App.dimensions.ai (n.d.)
3. Hansen, J.H.L., Hasan, T.: IEEE Signal Process. Mag. 32, 74 (2015)
4. Kepuska, V., Bohouta, G.: 2018 IEEE 8th Annual Computing and Communication Workshop
and Conference (CCWC) (2018). https://doi.org/10.1109/ccwc.2018.8301638
5. Xie, B., Charness, N., Fingerman, K., Kaye, J., Kim, M.T., Khurshid, A.: J. Aging Soc. Policy
32, 460 (2020)
6. Amershi, S., et al., Proceedings of the 2019 CHI Conference on Human Factors in Computing
Systems - CHI ’19 (2019). https://doi.org/10.1145/3290605.3300233
112 S. Pant and N. Panigrahi

7. Terzopoulos, G., Satratzemi, M.: Informatics in Education 19, 473 (2020)

8. Bentley, F., Luvogt, C., Silverman, M., Wirasinghe, R., White, B., Lottrjdge, D.: Proceedings
of the ACM on Interactive. Mobile, Wearable and Ubiquitous Technologies 2, 1 (2018)
9. Belanche, D., Casaló, L.V., Flavián, C., Schepers, J.: Serv. Ind. J. 40, 203 (2019)
10. Yan, C., Ji, X., Wang, K., Jiang, Q., Jin, Z., Xu, W.: ACM Comput. Surv. (2022). https://doi.
org/10.1145/3527153
11. Panigrahi, N.: Geographical Information Science. Boca Raton Crc Press (2018)
12. Panigrahi, N., Computing in Geographic Information Systems. CRC Press (2014)
 Smart Garbage Classification

Aviral Jain , Vidipt Khetriwal , Hitesh Daga , and B. K. Tripathy(B)

School of Information Technology and Engineering, VIT, Vellore, TN, India

[email protected]

Abstract. Garbage recycling is a key aspect of maintaining our environment in

good condition. Poor waste management has the potential to have a significant
negative effect on the environment, on public health, and on the economy of the
nation. The garbage must be separated into groups with similar recycling processes
to make the recycling process much more effective and faster. Recycling, needless
to say, is a very important task for all the countries. Garbage categorization is the
most basic stage in enabling cost-effective recycling among the tasks required for
recycling. Some of the well-known deep learning models for trash categorization
include Densenet121, DenseNet169, InceptionResnetV2, MobileNet and Xcep-
tion architecture. In this article, we propose a procedure to recognize single trash
objects in images and categorize them into one of the recycling categories. A Con-
volutional Neural Network (CNN) with transfer learning is used. An analysis of
the results obtained from the study shows that the EfficientNet-b0 CNN performs
well under this scenario. The trash categorization problem for the target database
can be efficiently handled using deep learning approaches as they offer a reliable
foundation for image recognition with high consistency. Once a waste is put into
the bin, the top compartment scans and predicts which type of waste it is, and
then the respective lid of the bottom compartment opens, pushing the waste down.
This prototype model of the system is proposed, which can be used in real-time
implementation.

Keywords: Recycling · Convolutional neural network · Transfer learning ·

EfficientNet · Deep learning

1 Introduction

Recycling is quickly gaining traction as a necessary component of a sustainable society.

However, there is a significant hidden cost associated with the entire recycling process.
This is due to the selection, classification, and processing of recycled materials. Even
though many customers nowadays can sort their own garbage, they might not be sure
of which waste category to select when getting rid of a variety of items. In today’s
industrial and information-based society, finding an automated solution to recycling is
incredibly beneficial since it provides both environmental and economic advantages.
Using computer vision and artificial intelligence (AI), recognition, removal, and sorting
of things on a moving conveyor is possible. The vast majority of today’s best object
identification networks use CNN features ([20, 21]). The convolutional neural network

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 113–124, 2022.
https://doi.org/10.1007/978-3-031-23233-6_8
114 A. Jain et al.

(CNN), a widely used image classifier, is based on biological neural networks ([17, 18]),
which include many layers with neurons connected directly to neurons in the next layer
[26]. The use of CNN has the advantage of being independent of prior knowledge and
requiring less work in extracting features and design ([19, 22]). In image detection and
classification, the CNN has been a significant success ([23, 24]). A pre-trained model is
used as the basis for a new model in the machine learning technique known as transfer
learning [3]. A CNN named EfficientNet-B0 was trained by using more than a million
images from the ImageNet database. The network can categorize images into a thousand
different object categories. Hence, using this in our model as a base layer would lead
to faster optimization and hence better results. No human workforce is required in the
adopted and implemented methodology. Trash items can be directly thrown into the
main container, which acts as a common space for attached separate dustbins and where
segregation takes place ([4, 5]).

2 Literature Review
In the paper by Lam, K.N. et al. [9], by using SSD-MobileNetv2, the server can identify
and classify 3 types of garbage: bottles, nylon, and scrap paper. The SSD architecture is a
single convolution network that learns to anticipate and classify bounding box locations
in a single run. As a result, SSD can be trained from beginning to end. All garbage has
been labeled and the position is returned to the Four Degrees Of Freedom (4 DoF). 4 DoF
can pick up and return the garbage to the correct trash bin. With SSD-MobileNetv2, the
recognition and classification of garbage have achieved an elevated level of accuracy. A
system has been developed that effectively visually classifies and separates several types
of waste, a task that would typically require manual labor. It does this by utilizing the
most recent innovations in computer vision, robot control, and other sectors and taking
advantage of their maturity as proposed by Salmandor et al. [2] and Adedeji et al. [10]. A
system that can analyze pictures from a camera and command a robot arm and conveyer
belt to automatically classify several types of waste has been developed using current
technology according to A. P. Puspaningrum et al. [1] and White et al. [14]. To detect and
classify domestic garbage, a novel deep CNN based on the multimodal cascading method
was developed in the papers by Simonyan et al. [7], Meng et al. [12] and Thanawala
et al. [13]. They created a smart trash bin system as the platform’s front-end carrier of
household garbage disposal, communicating directly with residents and providing data
support. They also gathered 30 000 rubbish pictures from homeowners as a dataset for
model training and identified 52 distinct types of waste. A multi-target detection model
for garbage pictures was also established at the same time to improve detection precision.
The trials revealed that the suggested method can improve detection precision by more
than 10% on average, and it also performs well in terms of model size and detection
time. In the paper by Kang et al. [15], ResNet-34 was found to perform the best and was
tweaked to perfection. Resnet34 is a 34-layer convolutional neural network that can be
used to create a cutting-edge image classification model. This model has been pre-trained
on the ImageNet dataset, which contains 100,000+ images from 200 different classes.
However, it differs from traditional neural networks in that it uses residuals from each
layer in the subsequent connected layers. Three modifications were made to the ResNet-
34 model, including multi-feature fusion, residual unit feature reuse, and optimization
 Smart Garbage Classification 115

of the activation function, to address the problem of trash classification. The changes
are put to the test on a trash dataset containing 14 different sorts of garbage items. The
original model has a precision of 0.9859. ResNet34-A’s accuracy has been improved
to 0.9941 using multi-feature fusion. ResNet-34-B’s residual unit accuracy has been
enhanced to 0.9995. With the changed activation function, the accuracy of ResNet-34-
C has increased to 0.9928. The highest accuracy is 0.9996 for ResNet-34-ALL, which
integrates all three modifications. The automatic trash classification system is completed
with the suggested algorithm and associated hardware. This system’s average cycle time
for classifying is 0.95 s, and its accuracy for classifying is 0.9996.

2.1 Gaps in Literature

In [9], the performance of the model can be improved by the following methods: A. The
performance of classification can be improved by training more models. B. The labeling
of images should be done carefully. C. This model can recognize only 3 types of garbage,
which is not enough according to the real-world scenario. So, more types should be
added to compare and evaluate the training model. In the paper by Li et al. 2022 [6], the
accuracy of the model can be further improved. Waste detection and automatic sorting
need more attention. The method proposed by Kang et al. [15] has several limitations: The
classification in the case of small targets has a scope for improvement. The classification
criteria can further be extended, including more categories such as kitchen waste, etc.
Overall, a general gap that is observed is the lack of classification categories, lower
accuracy scores, and complex structures. These are some points researchers are aiming
to improve upon. We have tried to provide solutions to some of the limitations in our
work.

3 System Details
The overall architecture of the system is shown through a prototype and can be described
in the following steps. The name of the prototype which we have made is ‘smart dustbin’.

3.1 Waste Scanning Through Camera

The waste is scanned by using a web camera as in Fig. 1. When a person puts a waste in
the bin, the camera is ready to scan the waste so that it can be classified by the dustbin.
We have also provided a light source, so that there is no problem in scanning even when
it is dark.

3.2 Waste is Segregated and the Lid Opens

The waste is segregated from the image produced in Fig. 1. This is shown in (Fig. 2).
Here, after scanning, the waste is classified as one of the items, and then the respective
lid to that compartment opens.
116 A. Jain et al.

Fig. 1. Waste scanning through camera

Fig. 2. Waste is segregated and the lid opens

3.3 Moving of Hands and Trash Being Put into Respective Compartment

Once the lid is opened, two hands move in both the direction to push the waste down
the opened compartment. This is shown in Fig. 3.
The workflow of the system is depicted in Fig. 4.
As stated in the diagrams, and the workflow, when a garbage is thrown, there is a
camera present to capture an image of the garbage. The image is then processed to the
CNN model. After the image is identified, it is classified into one of the categories. After
 Smart Garbage Classification 117

Fig. 3. Successful classification

Fig. 4. Workflow diagram

this, the lid of the appropriate dustbin opens and the arms move to push the garbage into
the opened compartment.

4 Component Modules and Description

The overall system works on three modules. The first one is the hardware component, that
is, the main container which is being used. The trash is to be thrown in this custom-made
container. The top level has a camera attached, and the second level of the container is
internally sub-divided into various sections that will be used to segregate the waste into
distinct categories. Images of the trash are taken with the high-resolution camera and
transferred to the integrated Raspberry Pi controller for further analysis. The container
has some other components, like an LED light and robotic arms. The battery-powered
LED focuses on the trash to make sure that the trash is visible even in the dark. The
robotic arms can be said to be another sub-module which is responsible for pushing the
118 A. Jain et al.

waste into its respective container upon successful classification. Firstly, the arm moves
from left to right, sweeping the product to the extreme right, allowing it to fall into its
respective container, and then it moves from right to left for a similar sweep. Eventually,
the item will reach its respective container after both arm movements are completed.
The second main module is the software component, that is, the model that is being used
to classify the waste item. A deep neural network is used for this purpose, and the best
fitting model is saved as a “.h5” file. Now, this file is loaded into the Raspberry Pi module,
which is the third module connecting hardware and software components by providing
a control structure for the process of classification. It runs the machine learning models
for the classification process [11]. According to the classification results, the respective
lid opens, and hand movements begin.

5 Algorithmic Steps

Step 1: Data collection is done for waste classification.

Step 2: The practical imbalanced data is then processed and re-sampled.
Step 3: Augmentation, standardization, and preprocessing are done on the dataset.
Step 4: The ImageDataGenerator class, at every epoch, ensures the model gets fresh
iterations of the images.
Step 5: Post-completion of preprocessing, training of the model is done. Transfer learning
principles are implemented. Also, optimizers are used.
Step 6: Early stopping is used, and validation loss is monitored with a specific patience
level or value.

The dataset used was publicly available on Kaggle and is made from web scraping
the images. This dataset has 15,150 images from 12 different classes of household
garbage: paper, cardboard, biological, metal, plastic, green-glass, brown-glass, white-
glass, clothes, shoes, batteries, and trash. ([8, 26]) Some sample images in the dataset
are shown in Fig 5.

Fig. 5. Sample images in dataset

 Smart Garbage Classification 119

The algorithm used for image classification is straightforward. The process starts
with the creation of a dataset and preprocessing. The dataset was found to be highly
imbalanced and hence the distribution was improved by re-sampling the dataset [16].
The images are captured inside the main container in an isolated environment, so
there is no need for segmentation (to separate the object from the background). Image
augmentation is a method of altering original images by applying various transformations
to them, resulting in many altered copies of the same image. These picture augmenta-
tion approaches not only increase the amount of the dataset but also add variance to it,
allowing the model to generalize better on unknown data. The images are preprocessed
using ImageDataGenerator. It offers a variety of augmentation options, including stan-
dardization, rotation, shifts, flips, brightness changes, and more. It is designed to provide
real-time enhancement of the data. In other words, it generates augmented images while
the model is still learning. The ImageDataGenerator class ensures that the model gets
fresh iterations of the images at every epoch [25]. The input image must be resized in
accordance with the classifier input layer before being fed to the classifier.
After preprocessing is done, the image is passed to the network and hence the model
is trained. The concept of transfer learning is used here. Basically, as an optimization,
a model trained on one job is reconfigured on a second, similar job, allowing for rapid
progress while creating a model for the second job. The pre-trained model used here
is EfficientNet-B0, which is relevant to our system as it has learned rich feature rep-
resentations for a wide range of images. A flatten layer and a dense layer are further
added to the network and the model is compiled using the Adam optimizer. Optimizers
are modules that adjust model attributes like weights and learning rates to minimize
losses. Optimizers are generally preferred during training as they aid in obtaining faster
outcomes.
During the training of the model, the concept of early stopping is used, where vali-
dation loss is monitored with a patience of 10. Basically, it stops the model fitting before
the total number of epochs if the model isn’t improving. The best trained model is saved
to a “.h5” file and is used to perform classification in real time.

6 Result and Analysis

An accuracy of 99% was achieved on the training dataset, 92% on the validation dataset,
and 93% on the training dataset. To understand the exact distribution of the predictions
made by the model, we can generate a classification report using sklearn.metrics. This
provides different metrics like precision, recall, f1-score etc. to evaluate the performance
of the model.
Accuracy is a metric for classification models that measures the number of pre-
dictions that are correct as a percentage of the total number of predictions that are
made.
Precision counts the percentage that is correct among everything that has been pre-
dicted as a positive; Recall counts the number which the model succeeds to find from
among everything that actually is positive. Mathematically,

#of True Positives

Pr ecision = (1)
#of True Positives + #of False Positives
120 A. Jain et al.

#of True Positives

Recall = (2)
#of True Positives + #of False Negatives
Pr ecision ∗ Recall
F1 score = 2 ∗ (3)
Pr ecision + Recall
Precision is used to indicate the model’s performance. It determines the quality of
a ‘positive prediction’ that is obtained from the model. The recall is the measure of our
model’s correctly identifying true positives. The F1 Score gives the weighted average
of precision and recall. As a result, the F1 score accounts for both false positives and
false negatives. The classification report generated is shown in Fig. 6. It is observed that
items in some categories like clothes, shoes, trash, etc. are getting classified with high
accuracy. However, classification for items like metal and plastic is not performing very
well. This is due to the imbalanced nature of the dataset.

Fig. 6. Classification report

The table in Fig. 6 provides the classification of trash into different classes in the first
12 rows. While the columns under precision, recall and F1-measure are as explained in
the formulae provided, the last three are on the whole data set of 999 elements. The last
column provides the distribution of 999 elements into different classes (mentioned in
the first column) in the 12 rows.
The simulated working of the model is depicted in Fig. 7 and Fig. 8. We can see that
in this case, the model correctly classified the items as ‘plastic’ and ‘clothes’.
So, as we can see, the camera detection analysis helps us recognize the type of waste
successfully. This shows that the model was trained effectively and, hence, the object
identification and detection were made possible in a matter of seconds. The accuracy
 Smart Garbage Classification 121

Fig. 7. Plastic bottle being classified as ‘plastic’

Fig. 8. Handkerchief being classified as ‘clothes’

in some of the classes is better than in others. This smart, quick, and efficient method
of classification makes segregation a lot easier and faster. We are the future generation,
and it is our responsibility to look after the plant and dump waste in the allotted bins
122 A. Jain et al.

(biodegradable or non-biodegradable). Only if we take care of the planet will our future
generations be able to thrive in a stable way.
As the outcomes of this study show, the problem of trash picture categorization may
be solved with high accuracy using deep learning algorithms. The achieved accuracy was
93% on the testing dataset. The use of EfficientNet-B0 and the Adam optimizer helped
to attain this accuracy score. This accuracy can be improved by further preprocessing
the dataset.

7 Conclusions

This paper focuses highly on how we can segregate waste and differentiate it into partic-
ular kinds so that each kind can be identified and recycled. Every kind is re-used in some
way or other and we save our planet. The proposed model is much simpler than the other
proposals and the number of categories is increased to 12, which allows a more realistic
real-world implementation. The twelve categories include: battery, biological, brown-
glass, cardboard, clothes, green-glass, metal, paper, plastic, shoes, trash and white-glass.
Currently, the size of the dataset is relatively small and there are 12 categories, with each
category having only a few samples. Merging a few categories and scraping the internet
for more data can result in better accuracy. Even though the neural network is complex,
the process of real-time classification is amazingly fast, which increases the efficiency
of the proposed system.

References
1. Puspaningrum, A.P., et al.: Waste classification using support vector machine with SIFT-PCA
feature extraction. In: 2020 4th International Conference on Informatics and Computational
Sciences (ICICoS), pp. 1–6 (2020). https://doi.org/10.1109/ICICoS51170.2020.9298982
2. Salmador, A., Pérez Cid, J., Rodríguez Novelle, I.: Intelligent garbage classifier. Intelligent
Garbage Classifier. Int. J. Interact. Multimedia and Artificial Intelligence 1(1), 31–36 (2008)
3. Gao, M., Qi, D., Mu, H., Chen, J.A.: Transfer residual neural network based on resnet-34 for
detection of wood knot defects. Forests 12, 212 (2021)
4. Gondal, A.U., et al.: Real time multipurpose smartwaste classification model for efficient recy-
cling in smart cities using multilayer convolutional neural network and perceptron. Sensors
21, 4916 (2021). https://doi.org/10.3390/s21144916
5. Wang, H.: Garbage recognition and classification system based on convolutional neural net-
work vgg16garbage recognition and classification system based on convolutional neural net-
work VGG16. In: 2020 3rd International Conference on Advanced Electronic Materials,
Computers and Software Engineering (AEMCSE), pp. 252–255 (2020). https://doi.org/10.
1109/AEMCSE50948.2020.00061
6. Li, J., et al.: Automatic detection and classification system of domestic waste via multimodel
cascaded convolutional neural network. IEEE Trans. Industr. Inf. 18(1), 163–173 (2021)
7. Simonyan, K., Zisserman, A.: Very Deep Convolutional Networks for Large-Scale Image
Recognition, in arXiv:1409.1556 [cs], San Diego, CA, USA, pp. 1–14 (2015)
8. Mostafa, M.: Garbage Classification (12 classes). Dataset on Kaggle (2020). https://www.kag
gle.com/mostafaabla/garbage-classification
 Smart Garbage Classification 123

9. Lam, K.N., et al.: Using artificial intelligence and IoT for constructing a smart trash bin.
In: Dang, T.K., Küng, J., Chung, T.M., Takizawa, M. (eds.) FDSE 2021. CCIS, vol. 1500,
pp. 427–435. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-8062-5_29
10. Adedeji, O., Wang, Z.: Intelligent waste classification system using deep learning convo-
lutional neural network. Procedia Manuf. 35, 607–612 (2019). https://doi.org/10.1016/j.pro
mfg.2019.05.086
11. Khan, R., et al.: Machine learning and IoT-based waste management model. Comput. Intell.
Neurosci. 2021, 1–11 (2021)
12. Meng, S., Chu, W.T.: A study of garbage classification with convolutional neural networks.
In: 2020 Indo–Taiwan 2nd International Conference on Computing, Analytics and Networks
(Indo-Taiwan ICAN), pp. 152–157. IEEE (2020)
13. Thanawala, D., Sarin, A., Verma, P.: An approach to waste segregation and management
using convolutional neural networks. In: Singh, M., Gupta, P.K., Tyagi, V., Flusser, J., Ören,
T., Valentino, G. (eds.) ICACDS 2020. CCIS, vol. 1244, pp. 139–150. Springer, Singapore
(2020). https://doi.org/10.1007/978-981-15-6634-9_14
14. White, G., Cabrera, C., Palade, A., Li, F., Clarke, S.: WasteNet: Waste classification at the
edge for smart bins. arXiv preprint arXiv:2006.05873 (2020)
15. Kang, Z., Yang, J., Li, G., Zhang, Z.: An automatic garbage classification system based
on deep learning. IEEE Access 8, 140019–140029 (2020). https://doi.org/10.1109/ACCESS.
2020.3010496
16. Ziouzios, D., Tsiktsiris, D., Baras, N., Dasygenis, M.: A distributed architecture for smart
recycling using machine learning. Future Internet 12(9), 141 (2020)
17. Bhattacharyya, S., Snasel, V., Hassanian, A.E., Saha, S., Tripathy, B.K.: Deep Learning
Research with Engineering Applications. De Gruyter Publications (2020). https://doi.org/
10.1515/9783110670905
18. Maheshwari, K., Shaha, A., Arya, D., Rajasekaran, R., Tripathy, B.K.: 2 Convolutional neural
networks: a bottom-up approach. In: Bhattacharyya, S., Snasel, V., Hassanien, A.E., Saha,
S., Tripathy, B.K. (eds.) Deep Learning: Research and Applications, pp. 21–50. De Gruyter
(2020). https://doi.org/10.1515/9783110670905-002
19. Bose, A., Tripathy, B.K.: Deep learning for audio signal classification. In: Bhattacharyya, S.,
Hassanian, A.E., Saha, S., Tripathy, B.K. (eds.) Deep Learning Research and Applications,
pp. 105–136. De Gruyter Publications (2020). https://doi.org/10.1515/9783110670905-00660
20. Adate, A., Tripathy, B.K.: A survey on deep learning methodologies of recent applications.
In: Acharjya, D.P., Mitra, A., Zaman, N. (eds.) Deep Learning in Data Analytics. SBD, vol.
91, pp. 145–170. Springer, Cham (2022). https://doi.org/10.1007/978-3-030-75855-4_9
21. Kaul, D., Raju, H., Tripathy, B.K.: Deep learning in healthcare. In: Acharjya, D.P., Mitra,
A., Zaman, N. (eds.) Deep Learning in Data Analytics. SBD, vol. 91, pp. 97–115. Springer,
Cham (2022). https://doi.org/10.1007/978-3-030-75855-4_6
22. Tripathy, B.K., Parikh, S., Ajay, P., Magapu, C.: Brain MRI segmentation techniques based
on CNN and its variants. In: Brain Tumor MRI Image Segmentation Using Deep Learning
Techniques, pp. 161–183. Elsevier (2022). https://doi.org/10.1016/B978-0-323-91171-9.000
01-6
23. Prabhavathy, P., Tripathy, B.K., Venkatesan, M.: Analysis of diabetic retinopathy detection
techniques using CNN models. In: Mishra, S., Tripathy, H.K., Mallick, P., Shaalan, K. (eds.)
Augmented Intelligence in Healthcare: A Pragmatic and Integrated Analysis, pp. 87–102.
Springer Nature Singapore, Singapore (2022). https://doi.org/10.1007/978-981-19-1076-0_6
24. Adate, A., Tripathy, B.K.: S-LSTM-GAN: shared recurrent neural networks with adversarial
training. In: Kulkarni, A.J., Satapathy, S.C., Kang, T., Kashan, A.H. (eds.) Proceedings of the
2nd International Conference on Data Engineering and Communication Technology. AISC,
vol. 828, pp. 107–115. Springer, Singapore (2019). https://doi.org/10.1007/978-981-13-1610-
4_11
124 A. Jain et al.

25. Subramanian, M., Narasimha Prasad, L.V., Sathishkumar, V.E.: Hyperparameter optimiza-
tion for transfer learning of VGG16 for disease identification in corn leaves using bayesian
optimization. Big Data 10(3), 215–229 (2021)
26. Lu, W., Chen, J.: Computer vision for solid waste sorting: a critical review of academic
research. Waste Manage. 142, 29–43 (2022)
 Optical Sensor Based on MicroSphere
Coated with Agarose for Heavy Metal Ion
Detection

Asesh Kumar Tripathy1(B) and Sukanta Kumar Tripathy2

1
Department of Computer Science and Engineering (H), Koneru Lakshmaiah
Education Foundation, Vaddeswaram, Vijayawada 522502, AP, India
[email protected]
2
Department of Physics, Berhampur University, Bhanja Bihar,
Berhampur 760007, Odisha, India
[email protected]

Abstract. Micronutrients are essential for plants to flourish sustain-

ably. For optimum plant growth, the heavy meatal ions like Zn2+ ,
Cu2+ , N i2+ , etc. must be present in the right amounts. For example,
0.69mgKg −1 is the critical limit of Zn2+ ions in plants. It is the most
restricting micronutrient for wetland rice, and its absence poses a seri-
ous nutritional challenge. Zinc deficiency in edible plant parts causes
micronutrient deficiencies by impeding growth. In this paper, we show
the best way to build up a microsphere coated in agarose gel for the
detection of heavy metal (Zn2+ ) ions. Using the Finite-Difference Time-
Domain approach, the optimal configuration of the sensor is then created
in order to calculate the concentration of Zn2+ ion. The proposed sen-
sor is shown to have a sensitivity and LOD for Zn2+ ion detection of
3.1021au/ppm and 0.01924ppm, respectively.

Keyword: Heavy metal Zn2+ ion sensor, Agarose coated, Microsphere,

Resonance

1 Introduction
A decent crop in plants can be ensured by the soil’s optimal zinc concentration.
All living things require zinc, a micronutrient important for growth, develop-
ment, and defence. According to Indira Sarangthem et al. (2018), the critical
limits in soil nutrients, such as Zn2+ characteristics, to distinguish between
adequacy and deficiency are 0.69mgKg −1 . Rastegarzadeh and Rezaei (2008)
designed an optical sensor to detect Zn2+ ion by spectrophotometry. He used
Zincon as a sensing agent. The procedure metnoined requires laboratory set up
and the method has a limit of detection of 0.16M . Aksuner et al. (2011) devel-
oped a sensor membrane, in laboratory, that is capable of determining Zn2+ ions
with a LOD of 1.6μgL−1 . Hassana et al. (2021) used cystalline optical fibers as
a sensor for the Zn2+ ion concentration using the image processing method.
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022
M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 125–136, 2022.
https://doi.org/10.1007/978-3-031-23233-6_9
126 A. K. Tripathy and S. K. Tripathy

The sensor showed a sensitivity of about 73.47%. Gupta et al. (2016) synthe-
sized 2-((5-methylpyridin-2-ylimino)methyl)phenol (L1) chemosensor, in labora-
tory, for detection of Zn2+ ions with a LOD of 1.13 × 10−7 M . Abbasitabar et
al. (2011) showed how to determine Zn2+ using a very accurate and reversible
optical chemical sensor that uses dithizone as the chromoionophore immobilised
within a plasticized carboxylated PVC sheet.
The development of optical sensors has received significant analytical atten-
tion due to its advantages in terms of size, cost, ease of sample preparation, speed
of measurement, and lack of dependence on a reference solution as mentioned
by Fen et al. (2013). Ramdzan et al. (2020) mentions that future development
of innovative sensing composites with outstanding sensitivity and selectivity for
heavy metal ion detection and other practical sensing applications may greatly
benefit from the development of biopolymers and conducting polymers with SPR
sensors. Lim and Yoon (2015) emphasized on vigorous pursuance of research into
the creation of a water-quality measurement device that will allow us to deter-
mine the potability of water swiftly and precisely at a reasonable cost for the
general population. An experimentally proven fiber-optic interferometric sensor
for N i2+ detection was proposed by Raghunandhan et al. (2016). The lower
detection limit was found to be 0.1671M , while the detection sensitivity was
found to be 0.05537nm/M . In the concentration range up to 500M , the sug-
gested sensor displays a linear response. Heavy metal ion detection using an
optrode, an optic-chemical sensor, has been developed by Czolk et al. (1992). Ions
like Cd(II), Pb(II), or Hg(II) can be complexed to produce distinct absorbance
spectra, which alters the sensor’s reflection behaviour. In the instance of Cd,
the impact of immobilisation on the dye’s complexation properties has been
investigated (H). For Cd, the sensor’s detection limit is 3 × 10−6 mol/l, and
it is reversible (U). Zhang et al. (2020) summarized progress of optical fiber
heavy metal ion sensors through various measurement methods based on optical
absorbance, fiber grating, modal interference, plasmonic and fluorescence. The
sensing characteristics of fiber-optic heavy metal ion sensor were also analyzed.
With the sensor reported by Klimant and Otto (1992) , heavy metal ions in
the concentration range of 3 × 10−6 to 3 × 10−5 M can be determined quickly
and simply. To be fully reversible, the sensor must be regenerated using diluted
HN O3 . The sensor may be used for the online measurement of the total concen-
tration of heavy metal ions in industrial waste water. In their investigation of
heavy metal pollution in various Indian soils, Kumar et al. (2019) revised prior
knowledge on both the average heavy metal contamination in soil. The review
by Lu et al. (2018) provides an overview of current developments in the use of
electrodes modified with inorganic materials for the electrochemical detection
of heavy metal ions (HMIs) in contaminated soils and other mediums. In this
study, Jeong and Kim (2015) investigated whether aminosilane (APTES) might
be used as a colorimetric detecting agent for heavy metal ions in the aqueous
phase. The amine group in APTES has a strong affinity for bivalent metal ions
and is easily synthesised into the metal-amine complex, which might be trans-
formed into metal nanoparticles, through self-seed generation with silanization
of APTES.
 Optical Sensor Based on MicroSphere Coated with Agarose 127

The discussion shows that there is a need for highly accurate, reasonably
priced, and user-friendly HM ion concentration sensors. The organisation of this
document is as follows. The Sect. 2 describes the intended device’s sensing idea.
Section 3 discusses the device design process. The results are examined and our
proposed sensor is compared with other sensors that are already being used in
the literature in the Sect. 4. Section 5 concludes the discussion.

2 Sensing Principle
The FDTD approach, which is based on the Yee (1966) algorithm, is used to
investigate electromagnetic wave propagation since it reduces the amount of
compute and memory needed, according to Qiu and He (2000) and Qiu and He
(2001). An area of space is selected for field sampling in both space and time. At
t = 0, all fields in the sample region are zero. Examining mode radiation makes
use of the central difference approximation. The following update equations from
Sundaray et al. (2018) are used in our simulation.
 n+1  n−1   n 
 2  2 Δt n 

D̃ = D̃  + √ Hy   − Hy  
i,j i,j Δy. μ0 ω0 i+1 i−1
2 ,j 2 ,j
    (1)
Δt  n
 n
− √ Hx  + Hx 
Δx. μ0 ω0 i, j+1 i, j−1
2 2
(n+1) n−1
 
Hx  = Hx 
i, j+1
2 i, j+1
2
  n+1  n+1  (2)
Δt  2  2
− √ Ez  + Ez 
Δz. μ0 ω0 i,(j+1) i,j
(n+1) (n−1)
 
Hx  = Hx 
i+1 i+1
2 ,j 2 ,j
  n+1  n+1  (3)
Δt  2  2
− √ Ez  + Ez 
Δz. μ0 ω0 (i+1),j i,j

The x and y coordinate axes of the lattice space increments are referred to as Δx
and Δy. The coordinates of the sample sites are represented by the numbers i and
j, which stand for x and y, respectively. The time increment is represented by Δt
and coupled to the number n to localise a predetermined observation interval.
A perfectly matched layer (PML) was employed to ensure the uniqueness and
validity of the numerical solution of Maxwell’s equations inside the computation
domain. In this study, the spatial step (Δx, Δy) is 40 λ
and the temporal step
−12
(Δt) is 2 × 10 s. λ is the signal wavelength. Additionally, stability is attained
when the following conditions are met:
 − 12
1 1 1
Δt < + (4)
c Δx2 Δy 2
128 A. K. Tripathy and S. K. Tripathy

where c is the speed of light. Here, the loss in the proposed sensor caused
by transmission-related absorption or scattering is not taken into account. A
correction factor is included in the simulation to account for these losses.
According to Baaske and Vollmer (2012) optical microcavities have devel-
oped into one of the most sensitive micro/nanosystems biodetection technology.
Light enters our proposed structure through the glass rod and the microsphere.
According to Tripathy et al. (2022), light is trapped inside the micro-sphere and
circulates for many tens of thousands of times because of the material’s pre-
dominate absorption loss. If scattering losses at the microsphere’s boundary are
controlled and light absorption in the transparent material is kept to a mini-
mum, photons can go around their orbit thousands of times before exiting the
microsphere via the loss mechanism. According to Soria et al. (2011), the long
duration of imprisoned light is correlated with a long optical path length due
to the resonant character of the event. Additionally, a whispering gallery mode
(WGM) often has many more total internal reflections per orbit than is shown
in the Fig. 1 , making the polygonal optical route more similar to a circular opti-
cal path that circles the microsphere’s surface. This straightforward illustration
guides us to the general WGM resonance condition:
λr
N× = 2πR (5)
n
It asserts that a circular optical route with a length 2πR, a microsphere radius R,
and a microsphere refractive index n must fit an integer N number of wavelengths
λr. The WGM’s resonant frequency is then determined as follows:
c cN
ω = 2πf = 2π = (6)
λr nR

Even while all solutions have the same concentration, they all have differ-
ent refractive indices. As stated in Sun et al. (2019), carrier absorption is
the likely cause. The small optical resonator is utilised to detect changes in
the WGM resonance frequency that occur when Zn2+ ions attach to the
Dithizone(C13 H12 N4 S) coating. Figure 2 demonstrates how a Zn2+ ion’s bind-
ing slightly affects the resonance frequency. The shift takes place as a result

Fig. 1. Micro-Sphere with Total Internal Reflection

 Optical Sensor Based on MicroSphere Coated with Agarose 129

of the connected Zn2+ ion “pulling” a piece of the optical field outside of the
microsphere, lengthening the route by 2πΔl. According to:
Δω 2πΔl Δl
=− =− (7)
ω 2πR R
this increase in path length results in a shift in the resonance frequency Δω.
Amplitude

Resonance Ferquency

Fig. 2. The WGM path length is lengthened when a Zn2+ ion binds to the Dithizone
coating on the surface of the microsphere, and this is seen as a Δω resonance frequency
shift.

3 Sensor Design

Figure 3 depicts the proposed sensor’s configuration. By first coating the sphere
with agarose gel and subsequently with Zinc ion binding Dithizone, a novel opti-
cal microscopic sphere sensor that can detect the concentration of Zinc ions indi-
rectly was proposed. We investigated micro-sphere sensing with various sphere
and cylinder radii as well as its various lengths in order to determine the best
sensor parameters; the spectra are given in Figs. 4, 5, and 6.
According to Fig. 4, the sphere’s radius is 0.65mum and the wavelength with
the least reflected intensity is 640nm. Therefore, the speher’s diameter is kept
at 130μm. As stated in Han et al. (2019) , the microsphere is coupled with a
solid cylinder made of silicon dioxide (SiO2 ). The cylindrical glass rod’s radius
and length were adjusted, and it was found from Figs. 5 and 6 that a radius of
30μm and a length of 150μm, respectively, at 565nm and 640nm, offered the
least amount of reflection. The diameter and length of the cylindrical rod were
calculated to be 60μm and 150μm, respectively. It is first coated with precast
Agarose gel with RI 1.3661 and then with Zinc ion binding Dithizone with RI
1.684.
130 A. K. Tripathy and S. K. Tripathy

Dithizone
Agarose

10
μm
130 μm

150 μm

60 μm

Fig. 3. The proposed sensor’s design

Fig. 4. Interference spectra at various radii of Sphere

 Optical Sensor Based on MicroSphere Coated with Agarose 131

Fig. 5. Interference spectra at various radii of Cylinder

Fig. 6. Interference spectra at various lengths of Cylinder

The cylinder allows a plane wave with a wavelength between 400nm and
700nm to pass through. The microsphere absorbs some of the light at resonance,
reflecting back some of the remaining light. The modification of Zn2+ ions in
soil can alter the resonance environment, which moves the reflection spectrum’s
fringe. As a result, by keeping an eye on the reflection spectrum of the suggested
structure at the cylinder’s end, the Zn2+ ion concentration can be demodulated.
132 A. K. Tripathy and S. K. Tripathy

4 Results and Discussions

The concentration of Zn2+ affects the Refractive index (R.I.) of Zn2+ . There-
fore, in the device arrangement depicted in Fig. 3 of Sect. 3, we alter the refractive
index of a thin layer surrounding the Agarose coating in order to replicate the
various concentrations of Zn2+ ions. The refractive index changes as the con-
centration of Zn2+ ions rises, as illustrated in Table 1. The values less than or
equal to 70 ppm is taken from Fen et al. (2011). The values greater than 70
ppm is calculated as per the procedure mentioned in Tan and Huang (2015).
Due to the very small diemension of the suggested configuration, it is possible to
place the cylindrical structure into one end of a glass tube that is long enough.
The sensor’s observed reflection spectra for Zn2+ ions with different refractive
indices are shown in Figures 7 and 8. It is clear that the characteristic wavelength
changes whenever the refractive index of the Zn2+ ions changes, primarily as a
result of changes in the resonant state. Agarose is a type of hydrophilic poly-
mer, and as a result, it will take up water molecules from the soil, raising the
refractive index of the Zn2+ ions. As the Zn2+ ions’ refractive index rises and
the wavelength of the reflection spectra shifts, the resonant condition changes.
According to Fig. 8 the refractive index shifts from 1.3317 to 1.3318, causing a
0.00057 nm shift in the reflection spectrum.

Table 1. Variation of Zn2+ ions’ refractive indices

Concentration of Real Part of the Imaginary part

Zn2+ ion Refractive index, of the
(in ppm) n (± 0.0005) Refractive
index, k (± 0.0002)
0.5 1.3317 0.0002
1 1.3317 0.0002
10 1.3318 0.0005
30 1.3318 0.0009
50 1.3318 0.0009
70 1.3318 0.0020
90 1.3319 0.0028
110 1.3320 0.0046
130 1.3321 0.0056
150 1.3322 0.0061
170 1.3323 0.0065
 Optical Sensor Based on MicroSphere Coated with Agarose 133

Fig. 7. Interference spectra with different RI of Zn2+ ions

Table 2. Recent Literature On Sensor Performance For Zn2+ Ion Detection

Range LOD Sensitivity Ref

0.76 − 30.60µM 0.16µM Not Mentioned Rastegarzadeh
et al.
2.5 × 10−8 to 5.8 × 10−8 molL−1 8.0 × 10−9 molL−1 Not Mentioned Abbasitabar
et al.
0 − 10ppm 0.01ppm Not Mentioned Daniyal
et al.
0 − 170ppm 0.01924 ppm 3.1021au/ppm Our
proposed
Sensor

The Fig. 8 shows that the reflection spectra shift monotonically as the Zn2+
ion concentration increases. As stated in Han et al. (2019), the sensitivity for
the Zn2+ ion concentration measurement using spectral shift may be described
as the ratio of the variation in reflection wavelength to the variation in Zn2+ ion
concentration, i.e. S = dZn2+ Concentration
dλres
. According to calculations, the sensor’s
sensitivity is 3.1021au/ppm, and its Limit of Detection (LOD) is 0.01924ppm.
Table 2 demonstrates that our suggested sensor is the most sensitive to the con-
centration of the Zn2+ ion. As a result, it may be used to find the soil’s critical
limit of the Zn2+ ion. Because the Zn2+ ion will only stick to the Dithizone,
the coating ensures the sensor’s specificity. The Zn2+ ions are the only ones
that change the Refractive Index (RI), as other metal ions do not bind to the
Dithizone coating.
134 A. K. Tripathy and S. K. Tripathy

Fig. 8. When the plot in Fig. 7 is magnified for the tip at wavelength 565 nm, the shift
is clearly visible.

Fig. 9. Linear fitting of the Reflection and Zinc ions concentration of the sensor

5 Conclusion

Finally, a novel design of a Zn2+ ion detection sensor based on an Agarose

Coated Micro-Sphere Resonator has been put forth. It was demonstrated that
the sensor could detect Zn2+ ion concentrations between 0ppm and 170ppm.The
sensor’s sensitivity to ion concentration is 3.1021au/ppm. The predicted lowest
detectable Zn2+ ion concentration is 0.01924ppm.
 Optical Sensor Based on MicroSphere Coated with Agarose 135

References
Abbasitabar, F., Zare-Shahabadi, V., Shamsipur, M., Akhond, M.: Development of an
optical sensor for determination of zinc by application of PC-ANN. Sens. Actuators,
B Chem. 156(1), 181–186 (2011). https://doi.org/10.1016/j.snb.2011.04.011
Aksuner, N., Henden, E., Yenigul, B., Yilmaz, I., Cukurovali, A.: Highly sensitive
sensing of zinc (II) by development and characterization of a PVC-based fluorescent
chemical sensor. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 78(3), 1133–1138
(2011). https://doi.org/10.1016/j.saa.2010.12.065
Baaske, M., Vollmer, F.: Optical resonator biosensors: molecular diagnostic and
nanoparticle detection on an integrated platform. ChemPhysChem 13(2), 427–436
(2012). https://doi.org/10.1002/cphc.201100757
Czolk, R., Reichert, J., Ache, H.: An optical sensor for the detection of heavy metal
ions. Sens. Actuators B Chem. 7(1–3), 540–543 (1992). https://doi.org/10.1016/
0925-4005(92)80360-A
Fen, Y.W., Yunus, W.M.M., et al.: Characterization of the optical properties of heavy
metal ions using surface plasmon resonance technique. Opt. Photonics J. 1(03), 116–
123 (2011). https://doi.org/10.4236/opj.2011.13020
Fen, Y.W., Yunus, W.M.M., Talib, Z.A., Yusof, N.A.: Fabrication and evaluation of
surface plasmon resonance optical sensor for heavy metal ions detection. In: 2013
IEEE 4th International Conference on Photonics (ICP), pp. 114–116 IEEE (2013).
https://doi.org/10.1109/ICP.2013.6687085
Gupta, V.K., Singh, A.K., Kumawat, L.K., Mergu, N.: An easily accessible switch-on
optical chemosensor for the detection of noxious metal ions Ni (II), Zn (II), Fe (III)
and UO2 (II). Sens. Actuators, B Chem. 222, 468–482 (2016). https://doi.org/10.
1016/j.snb.2015.08.063
Han, C., Ding, H., Zhao, C., Chen, C.: Ultrafast miniature humidity sensor based on
single-sided microsphere resonator. J. Lightwave Technol. 37(21), 5493–5499 (2019).
https://doi.org/10.1364/JLT.37.005493
Hassana, O.S., Al-azawi, R.J., Mahdi, B.R.: Image processing technique for zinc ion
sensing using a crystalline fiber sensor. Eng. Technol. J. 39(10), 1539–1543 (2021).
https://doi.org/10.30684/etj.v39i10.2136
Indira Sarangthem, L., Sharma, D., Oinam, N., Punilkumar, L.: Evaluation of critical
limit of zinc in soil and plant. Int. J. Curr. Res. Life Sci. 7(08), 2584–2586 (2018)
Jeong, U., Kim, Y.: Colorimetric detection of heavy metal ions using aminosilane. J.
Ind. Eng. Chem. 31, 393–396 (2015). https://doi.org/10.1016/j.jiec.2015.07.014
Klimant, I., Otto, M.: A fiber optical sensor for heavy metal ions based on immo-
bilized xylenol orange. Microchim. Acta 108(1), 11–17 (1992). https://doi.org/10.
1007/BF01240367
Kumar, V., et al.: Pollution assessment of heavy metals in soils of india and ecological
risk assessment: a state-of-the-art. Chemosphere 216, 449–462 (2019). https://doi.
org/10.1016/j.chemosphere.2018.10.066
Lim, S.-H., Yoon, S.: Sensors and devices for heavy metal ion detection. In: Kyung, C.-
M. (ed.) Smart Sensors for Health and Environment Monitoring. KRS, pp. 213–232.
Springer, Dordrecht (2015). https://doi.org/10.1007/978-94-017-9981-2 9
Lu, Y., Liang, X., Niyungeko, C., Zhou, J., Xu, J., Tian, G.: A review of the identifica-
tion and detection of heavy metal ions in the environment by voltammetry. Talanta
178, 324–338 (2018). https://doi.org/10.1016/j.talanta.2017.08.033
Qiu, M., He, S.: Numerical method for computing defect modes in two-dimensional
photonic crystals with dielectric or metallic inclusions. Phys. Rev. B 61(19), 12871
(2000). https://doi.org/10.1103/PhysRevB.61.12871
136 A. K. Tripathy and S. K. Tripathy

Qiu, M., He, S.: FDTD algorithm for computing the off-plane band structure in a
two-dimensional photonic crystal with dielectric or metallic inclusions. Phys. Lett.
A 278(6), 348–354 (2001). https://doi.org/10.1016/S0375-9601(00)00795-7
Raghunandhan, R., et al.: Chitosan/PAA based fiber-optic interferometric sensor for
heavy metal ions detection. Sens. Actuators, B Chem. 233, 31–38 (2016). https://
doi.org/10.1016/j.snb.2016.04.020
Ramdzan, N.S.M., Fen, Y.W., Anas, N.A.A., Omar, N.A.S., Saleviter, S.: Devel-
opment of biopolymer and conducting polymer-based optical sensors for heavy
metal ion detection. Molecules 25(11), 2548 (2020). https://doi.org/10.3390/
molecules25112548
Rastegarzadeh, S., Rezaei, V.: An optical sensor for zinc determination based on zincon
as sensing reagent. Sens. Actuators, B Chem. 129(1), 327–331 (2008). https://doi.
org/10.1016/j.snb.2007.08.016
Soria, S., et al.: Optical microspherical resonators for biomedical sensing. Sensors 11(1),
785–805 (2011). https://doi.org/10.3390/s110100785
Sun, P., Chen, Y., Gao, C., Liu, X., Yang, X., Xu, M.: Heavy metal ion detection
on a surface plasmatic resonance based on the change of refractive index. In: 9th
International Symposium on Advanced Optical Manufacturing and Testing Tech-
nologies: Optoelectronic Materials and Devices for Sensing and Imaging, vol. 10843,
pp. 365–373 SPIE (2019). https://doi.org/10.1117/12.2506976
Sundaray, M., Tripathy, S., Das, C.: FDTD analysis of diffraction efficiency in a holo-
gram for application in optical fiber communication. Optik 154, 325–330 (2018).
https://doi.org/10.1016/j.ijleo.2017.10.003
Tan, C.-Y., Huang, Y.-X.: Dependence of refractive index on concentration and tem-
perature in electrolyte solution, polar solution, nonpolar solution, and protein solu-
tion. J. Chem. Eng. Data 60(10), 2827–2833 (2015). https://doi.org/10.1021/acs.
jced.5b00018
Tripathy, A.K., Tripathy, S.K., Sundaray, M.: An ultra-sensitive optical sensor based
on agarose coated microscopic sphere to detect cu2+ ion in soil. Comput. Electron.
Agric. 202, 107424 (2022). https://doi.org/10.1016/j.compag.2022.107424
Yee, K.: Numerical solution of initial boundary value problems involving Maxwell’s
equations in isotropic media. IEEE Trans. Antennas Propag. 14(3), 302–307 (1966).
https://doi.org/10.1109/TAP.1966.1138693
Zhang, Y.-N., Sun, Y., Cai, L., Gao, Y., Cai, Y.: Optical fiber sensors for measure-
ment of heavy metal ion concentration: a review. Measurement 158, 107742 (2020).
https://doi.org/10.1016/j.measurement.2020.107742
 Influential Factor Finding for Engineering
Student Motivation

Sourajit Ghosh1 , Aritra Sinha2(B) , Arpita Roy2 , and Biswarup Neogi2

1 Department of Computer Science and Engineering, JIS University, Kolkata, WB, India
2 Department of Electronics and Communication Engineering, JIS College of Engineering,
Kalyani, WB, India
[email protected], [email protected]

Abstract. The case study encompasses the evaluation of the motivational factors
of a student to pursue and complete his/her engineering degree program. In order
to analyze the factors, a large amount of data have been collected from the students
of various engineering institutes. Two categories of factors have been found and
the most influential attribute has also been realized by a statistical analysis of the
data set. The two categorized factors are: Intrinsic and Extrinsic. And, either of
them has been seen to influence the motivation of the students. The study shows
that students are mostly bothered by the anxiety of being outperformed by their
fellow classmates.

Keywords: Motivation · Data · Factor analysis · Attributes · Linear regression ·

Student motivation · Intrinsic · Extrinsic

1 Introduction
Motivation is a word, that is derived from the word ‘motive’, which means the need of a
person. It also means the desire, wants, or drive of a human being [1]. It is the process of
motivating people to achieve their personal or collective goals [2]. This study has tried
to find the most influential factor in the motivation of engineering students [3].
The engineering degree program is a robust, rigorous process that runs for four years
in India. The motivation level of a student can’t be the same for the whole four years.
Sometimes students are highly motivated, sometimes they are not. Different levels of
motivation have been observed during various phases of the course. This motivation
level [4] does fluctuate because of some external reasons or attributes. In this study, we
have analyzed those attributes and categorized them into factors and in the end, will
find the most influential factor in engineering students’ motivation. In order to do this
analysis,

i. a questionary was prepared in accordance with educators and psychologists. It was

distributed amongst different students of various engineering institutes excluding
1st-year students. There was no bar on gender or stream of engineering. After getting
enough responses from the students, accepting more entries was stopped.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 137–147, 2022.
https://doi.org/10.1007/978-3-031-23233-6_10
138 S. Ghosh et al.

ii. After that, the data were analyzed with a statistical tool [5]. From the analysis, the
data were categorized into two major factors.
iii. Then, the study tried to find out the most influential attribute that impacts engineering
student motivation the most.

While browsing through hitherto relevant studies on the means and the strategies
[6] to motivate students, it has found a lack of analysis regarding the assessment of the
influential factors in student motivation and the assessing the dominating attribute.
Here, the related studies are reported in section II, followed by the experiment in
section III. Sections IV and V are the discussion and conclusion respectively.

2 Related Studies

Adam N. Kirn a Ph.D. scholar at Clemson University has disserted a study on the moti-
vation of students in his work [7]. His work tells, about a sequential explanatory study
of various methods that can motivate engineering students. How the current behavior
of engineering students is influenced by long-term motivation. The most common indi-
cator of students’ performance is academic performance and it does not consider the
needed fundamental motivation. Students can use their logical resources efficiently as
it is somewhat triggered by academic performance. The relevant features (e.g. Expec-
tations, values, future scopes) of student motivation in relation to students’ long-term
goals and short-term work assignments were examined in the first stage of the study. In
the second stage, three more factors of student motivation, are expectations of success
in engineering courses, current recognition as a student of an engineering program, and
ideas about their future as an engineer come under the study. Along with these their
problem-solving ability also comes under the scrutiny of this study. The third stage of
the study inspects the motivational profiles of senior engineering students in important
and key subjects.
In order to do the study, a specific group was formed. The first phase result showed
that the expectation of students and their insights about the future differentiate students
with different long-term goals. In the second stage, it was seen that students’ technical
problem-solving ability correlates with the recognition of students’ future. The senior
engineering students were divided into groups according to their expectations, problem-
solving ability, and their future perceptions. Long-term goals and measures of students
were asked in the fourth stage of the study. This phase is a continuation study on previous
works by inspiring the engineering student experience with their Future Temporal Out-
look (FTP). The result of the fourth stage showed that some of the students had a clear
perception on the FTP and some others didn’t have a specific goal beyond completion
of the course program.
Development of plans beyond just completing the course with a large variety is also
seen in the study. Also, a well-defined future for students creates greater aspects of their
current works and tasks. It helps them to improve their own performance.
Understanding the relationship between student motivation and current behavior
helps engineering educators raise their interest in engineering and prepare students to
become effective engineers. In the final phase of this work, they explore how students
 Influential Factor Finding for Engineering Student Motivation 139

looked upon a technical problem to solve the problem efficiently. The result shows that
motivation across the different time scales actually determines the students’ perception
of problem-solving. In addition, assessment by the students of engineering issues may
be based on the student’s integrated cultural observation of engineering issues.
In her research on student motivation [8], Linda S. Lumsden of ERIC Clearinghouse
on Educational Management, Eugene, and Oreg have summarized regaining student
motivation. She has some offer-specific strategies that can be used in the classroom, etc.
Address issues outside the classroom and admit this throughout the school Policies and
practices can also stimulate or satisfy a student’s hunger for learning. James P. Rafini
[9] encourages educators to investigate the idea of “win or lose” in many schools. He
makes a suggestion for structural changes and classroom strategies aimed at empowering
students’ motivation. Recalling that “classrooms are not islands,” Martin L.[10] Mar and
Carol Migily have school-wide policies, practices, and procedures that affect student
motivation. They suggest the process by which the principal can start moving the school
Apart from emphasizing relative abilities “Learning, Achievement, Effort.“ How does
Carol A. Ames focus? Motivational concepts and processes are suitable for everyday
use Teacher issues and decisions. Jere Brophy shows this example Four categories of
motivational strategies available to teachers stimulate your interest in learning. (1) Taking
care of students’ Expectations for success; (2) Provide external motivation. (3) Take
advantage of existing essential motivations. (4) Stimulate Student motivation to learn.
Hermione H. Marshall [11] clearly different motivational directions for the three-fifth
graders’ teachers.
After going through several previous pieces of research on this very ground, it was
found that all of them comprise the ways to motivate students and the relevant strategies as
well [12]. No hitherto study has been found that analyzes what are the influential factors
that impact the students’ motivation and which attribute is predominant. Therefore, we
made our research in this very arena to assess the aforementioned.

3 Experiment
In order to do this analysis, some questions were set that can extract the students’ per-
spectives [13] on their program. The questions were set in such a way, that students will
answer the questions without knowing the purpose. Keeping this in mind, with the help
of consultation of psychiatrists and educators, a questionary was set, in which there were
some demographical questions and some of them were motivation-oriented. Amongst the
twenty-two questions, eighteen questions were taken for the analysis (Table 1).

Table 1. Questions and corresponding variables.

Sl. no. Question Variable Name

1 I am enjoying learning engineering V1
2 My personal goals and objectives are linked to my learning V2
(continued)
140 S. Ghosh et al.

Table 1. (continued)

Sl. no. Question Variable Name

3 It has always bothered me that, other students will outperform me in V3
the evaluation process
4 It makes me concerned about how I will fare on the engineering V4
exam
5 I try to figure out why I’m having trouble learning engineering V5
subjects
6 I’ll be nervous when it’s time to take the test(s) V6
7 It is critical and valuable for me to achieve high grades in the V7
exam(s)
8 I am very interested in my study and put in a lot of effort to learn it V8
9 I use a variety of approaches to ensure that I fully understand the V9
course
10 The subject I’m learning can help me find a great job V10
11 I expect to outperform other students in a technical subject(s) V11
12 It worries me to think about poor performance in the exam(s) V12
13 I am concerned about how my Engineering performance will affect V13
my overall grade
14 I despise even thinking about the evaluation V14
15 It is important to me how I will apply the engineering that I study in V15
my daily life and in the future
16 All of my technical knowledge is related to or relevant to my V16
existence
17 I am confident in my course abilities and competencies V17
18 I am satisfied with my progress in understanding the subject(s) V18

The other four questions were kind of demographical, from where students’
demographical presence can be looked upon. Those questions were (Table 2):

Table 2. Demographical questions

1 Year First/second/third/fourth
2 Gender Male/Female
3 Last semester marks
4 I study to Learn/Stay Away from social humiliation/Peer or family
pressure/Earn or rewarded
 Influential Factor Finding for Engineering Student Motivation 141

From the above-mentioned questions, a google form was created and distributed
amongst the students of different engineering institutes to collect data. The window was
kept open for around 15 days so that the responses from the students could be received
and registered. Students’ response was collected on a Likert scale of 5 scales, where 1
denotes very low, 2 stands for low, 3 is moderate, 4 implies high and 5 means very high.
In this period, 638 student responses were collected and put in Raosoft to check
the data sufficiency. The recommended sample size is directly proportionate with the
sample size, remarks Raosoft. It suggests that at least a sample size of 377 is necessary
to carry out a study on a large amount of population. A larger sample size ensures a
lower margin regarding the error and a higher confidence level as well. From now on,
the questions will be called attributes, that influence a student’s motivation. In the first
phase of analysis, a reliability test was run on the data set to check the reliability. During
the reliability tests, Chronbach’s Alpha value of 0.9 was maintained which indicates the
data’s reliability (Table 3).

Table 3. Chronbach’s alpha values

Rotation Chronbach’s alpha

1st 0.913
2nd 0.905
3rd 0.900

Cronbach’s alpha is a measure of internal consistency, or how closely related a group

of items is. It is regarded as a scale reliability metric. Then, the attributes were analyzed
to be categorized by factor analysis. For reliability testing and factor analysis, a statistical
tool named Statistical Package for Social Science [SPSS] was used. SPSS is a tool [14],
that provides descriptive and inferential statistics on the data researchers have collected
from surveys or observations. It can analyze a huge pool of data with ease and this is
very much user-friendly also [15].
The received data set was opened in SPSS and run the factor analysis method on
the said dataset so that relevant attributes can be extracted and the irrelevant attributes
can be eliminated by finding out the attributes with low communalities, cross-loading,
or low item-total correlation.
In the first phase of the factor analysis, the dataset was to be checked if it was
significant to run the factor analysis. For this, a KMO & Bartlett’s test was done. The
critical measurement of KMO sampling adequacy is 0.6, i.e., the value needs to be above
0.6 to be significant. From the given data set it was seen the KMO sampling adequacy
is well above 0.9, this tells the dataset is substantial (Table 4).
From the first rotation of the factor analysis method, it can be stated that there are
two dominating factors in which all our attributes can be categorized. And from the
pattern matrix of the first rotation, it was seen that only one attribute (V13) has cross-
loading. Therefore, the attribute was removed from the list and the second rotation of
factor analysis was run. From the second rotation, no low communality attribute or low
142 S. Ghosh et al.

Table 4. KMO sampling adequacy

Rotation KMO sampling adequacy

1st 0.931
2nd 0.927
3rd 0.922

item-total correlation was found amongst the attributes. V7 attribute once again showed
cross-loading. Therefore, it had to be removed as well. After removing the V7 and V13,
the factor analysis was run for the third time, and no attributes with low commonality,
cross-loading, and low item-total correlation were found. Thus, after the third rotation,
the attributes can be clearly categorized into two broad categories (Table 5).

Table 5. Two factors

Factor 1 Factor 2
I am enjoying learning engineering It has always bothered me that, other students
outperform me in the evolution process
My personal goals and objectives are linked to It makes me concerned about how I will fare
my learning on the engineering exam
I am very interested in my study and put a lot I try to figure out why I’m having trouble
of effort to learn it learning engineering subjects
I use a variety of approaches to ensure that I I’ll be nervous when it’s time to take the
fully understand the course test(s)
The subject I am learning can help me find a It worries me to think about poor performance
great job in the exam(s)
I expect to outperform other students in a I despise even thinking about the evolution
technical subject(s)
It is important to me how I will apply the
engineering that I study in my daily life and in
the future
All my technical knowledge is related to or
relevant to my existence
I am confident in my course abilities and
competencies
I am satisfied with my progress in
understanding the subjects

After analyzing the two factors, it can be seen that the attributes under Factor 1 are
occurring from a student’s own self, which we named Intrinsic and the attributes under
Factor 2 are caused due to external effects like family or peer pressure. Those are named
 Influential Factor Finding for Engineering Student Motivation 143

extrinsic. The scree plot also graphically acknowledges the two-factor theory. The scree
plot graphs the eigenvalue against the factor number. From the third factor on, it can be
seen that the line is almost flat, meaning each successive factor is accounting for smaller
and smaller amounts of the total variance (Fig. 1):

Fig. 1. Scree plot

At this stage, the method proposed two clear factors that are influencing engineering
student motivation: 1. Intrinsic Factors. 2. Extrinsic factors.
Now in the next phase, the most influential attribute of all is to be found. To find
this the linear regression [16] method was applied in SPSS. From the output of linear
regression, the standard error values of the attributes were compared.

3.1 Logistic Regression

In statistics, the method used for modeling a relationship between a scalar response and
one or more dependent or independent variables (also known as explanatory variables),
is known as linear regression. In linear regression, when there is one explanatory variable
present, the model is known as simple linear regression; on the other hand, if there are
more than one explanatory variable present, it is called multiple linear regression [17].
The first regression analysis method that undergoes in-depth research and sees a lot
of use in actual application is linear regression [18]. This is because models with linear
dependency on their unknown parameters are simpler to fit than models with non-linear
dependency on their parameters and because it is simpler to determine the statistical
characteristics of the resulting estimators.
144 S. Ghosh et al.

Table 6. Standard error

Variable Std. Error

V13 0.36
V17 0.44
V4 0.29
V9 0.38
V6 0.25

V3 0.24

V10 0.33
V16 0.38
V2 0.33
V14 0.27
V7 0.26
V12 0.26
V5 0.26
V8 0.36
V18 0.37
V15 0.36
V11 0.33

Linear regression is one of the most vital algorithms in supervised machine learning
[19]. It is applied in model predicting, forecasting, etc [20].
By applying the linear regression algorithm in the dataset, a list of standard errors of
each of the attributes was made. Table 6 depicts that the standard error value of the V3
attribute is the least. This indicates that V3 is the nearest attribute to the predicted model
or regression line. That means this is the very attribute that influenced engineering student
motivation the most. This is empirical by the study as it can be noted that the present-day
rat race, the cut-throat competition in every sector, the fear of being undervalued or being
left out during the process of achieving career goals in the future, or even securing a
job as per his capability and crave-all of these and more serve to be the reason behind
the fluctuating and sometimes, depleting motivation levels during different phases of
engineering students’ career.
The V3 attribute falls under the Extrinsic Factor category.

4 Discussion
The focus of our study was to find the most influential factor in engineering students’
motivation, i.e., the attribute which drives the engineering students’ motivation [19–
24] the most in their four-year marathon course. From the analysis, we have found the
 Influential Factor Finding for Engineering Student Motivation 145

attributes fall under two categories of Factors: 1. Intrinsic and 2. Extrinsic and the most
influential attribute (V3) that motivates the students, the most is also determined.
From the received answers the attributes were analyzed for factor analysis and factor
finding. Since the questionary was distributed amongst very few localized institutes, the
results received may contain location bias. To get even more precise results the study
may be carried out on a broader scale all over the state or the country.
The collected data used in the study comprises every stream of engineering course,
therefore, there was no scope for specification in the analysis. But there is an aspect to
doing this analysis on every particular stream of engineering or any general course. From
the received dataset we can categorize the data into two genders: Male and Female. A
huge difference was found between male and female students’ input. This implies that
engineering is still a male-dominated sector of education at least in our locality. If this
study can be done over a larger number of students, over a broader region like a state or
a country, the numbers can differ and upon that analysis can be done [4, 22, 25–29].
Therefore, there is still a lot of scope for research in the future on this topic. Our data
can be very much gender-biased or region biased. We can further do trade-based analysis,
gender-based analysis, or region-based analysis on the same topic. In that case, the dataset
can be used to assess the reason for the engineering trade being male-dominated. That
can give far more precise results regarding our research topic.

5 Conclusion
While analyzing the database, only two attributes were eliminated through factor anal-
ysis. After factor analysis was done, the attributes were classified into two following
categories, the first one being the intrinsic factor and the other one being extrinsic factor.
After analyzing the standard error of each attribute, to find out which one influenced a
student to keep motivated toward the course program, the attribute with the least standard
error should have the highest impact on the students. The study found that the V3 attribute
“It has always bothered me that, other students outperform me in the evaluation process”
is the most impactful on students’ motivation. Now, from the previous factor analysis,
the V3 attribute can be classified under the extrinsic factor.

References
1. Atkinson, J.W.: An introduction to motivation. Van Nostrand, Princeton, N.J. (1964)
2. Geen, R.G.: Social motivation. In: Colman, A.M. (ed.) Companion Encyclopedia of
Psychology, pp. 522–541. Routledge (2019). https://doi.org/10.4324/9781315002897-30
3. Savage, N., Birch, R., Noussi, E.: Motivation of engineering students in higher education.
Eng. Educ. 6(2), 39–46 (2011)
4. Mubeen, S., Norman, R.E.İD.: The measurement of motivation with science student. Eur. J.
Educ. Res. 3(3), 129–144 (2014)
5. Pivk, M., Le Diberder, F.R.: Plots: a statistical tool to unfold data distributions. Nucl. Instrum.
Methods Phys. Res., Sect. A 555(1–2), 356–369 (2005)
6. Yusuf, M.: The impact of self-efficacy, achievement motivation, and self-regulated learn-
ing strategies on students’ academic achievement. Procedia Soc. Behav. Sci. 15, 2623–2626
(2011)
146 S. Ghosh et al.

7. Kirn, A.N.: The influences of engineering student motivation on short-term tasks and long-
term goals. Doctoral dissertation, Clemson University (2014)
8. Lumsden, L.S.: Student Motivation. Research Roundup, vol. 10, issue (3), n3 (1994)
9. Raffini, J.P.: Winners Without Losers: Structures and Strategies for Increasing Student
Motivation to Learn. Allyn & Bacon, 160 Gould Street, Needham Heights, MA 02194 (1993)
10. Maehr, M.L., Midgley, C.: Enhancing student motivation: a schoolwide approach. Educ.
Psychol. 26(3–4), 399–427 (1991)
11. Marshall, H.H.: Motivational strategies of three fifth-grade teachers. Elem. Sch. J. 88(2),
135–150 (1987)
12. Cho, M.H., Heron, M.L.: Self-regulated learning: the role of motivation, emotion, and use
of learning strategies in students’ learning experiences in a self-paced online mathematics
course. Distance Educ. 36(1), 80–99 (2015)
13. Babo, R., Dey, N., Ashour, A.S. (eds.): Workgroups eAssessment: Planning, Implementing
and Analysing Frameworks. Springer (2021)
14. Bala, J.: Contribution of SPSS in social sciences research. Int. J. Adv. Res. Comput. Sci. 7(6),
250–254 (2016)
15. Acharjya, D., Anitha, A.: A comparative study of statistical and rough computing models in
predictive data analysis. Int. J. Ambient Comput. Intel. 8(2), 32–51 (2017)
16. Dey, N., Wagh, S., Mahalle, P.N., Pathan, M.S. (eds.): Applied Machine Learning for Smart
Data Analysis. CRC Press (2019)
17. Freedman, D.A.: Statistical Models: Theory and Practice. Cambridge University Press,
Cambridge (2009). https://doi.org/10.1017/CBO9780511815867
18. Krupinski, E.A.: Medical imaging. In: Chen, J., Cranton, W., Fihn, M. (eds.) Handbook of
Visual Display Technology, pp. 545–558. Springer, Cham (2016). https://doi.org/10.1007/
978-3-319-14346-0_186
19. Bindu, K.H., Raghava, M., Dey, N., Rao, C.R.: Coefficient of Variation and Machine Learning
Applications. CRC Press (2019)
20. Das, S.K., Das, S.P., Dey, N., Hassanien, A.E. (eds.): Machine Learning Algorithms for
Industrial Applications. Springer, Cham (2021)
21. Labib, W., Abdelsattar, A., Ibrahim, Y., Abdelhadi, A.: What motivates students to study
engineering? a comparative study between males and females in Saudi Arabia. Educ. Sci.
11(4), 147 (2021)
22. Saadon, N.F.S.M., Ahmad, I., Pee, A.N.C., Hanapi, C.: The implementation of augmented
reality in increasing student motivation: a systematic literature review. In: IOP Conference
Series: Materials Science and Engineering, vol. 854, no. 1, p. 012043. IOP Publishing (2020)
23. Benson, L., Morkos, B.: Student motivation and learning in engineering. In ASEE 120th
Annual Conference, p. 13 (2011)
24. Brown, P.R., Matusovich, H.M.: Unlocking student motivation: development of an engi-
neering motivation survey. In: 2013 ASEE Annual Conference & Exposition, pp. 23–1284
(2013)
25. Ng, B.L., Liu, W.C., Wang, J.C.: Student motivation and learning in mathematics and science:
a cluster analysis. Int. J. Sci. Math. Educ. 14(7), 1359–1376 (2016)
26. Rahman, A., Muktadir, M.G.: SPSS: an imperative quantitative data analysis tool for social
science research. Int. J. Res. Innov. Soc. Sci. V, 300–302 (2021)
27. Sweet, S.A., Grace-Martin, K.: Data Analysis with SPSS, vol. 1. Allyn & Bacon, Boston,
MA, USA (1999)
28. Basto, M., Pereira, J.M.: An SPSS R-menu for ordinal factor analysis. J. Stat. Softw. 46, 1–29
(2012)
29. Liu, R.X., Kuang, J., Gong, Q., Hou, X.L.: Principal component regression analysis with
SPSS. Comput. Methods Programs Biomed. 71(2), 141–147 (2003)
 Influential Factor Finding for Engineering Student Motivation 147

30. Bennett, C., Ha, M.R., Bennett, J., Czekanski, A.: Awareness of self and the engineering
field: student motivation, assessment of ‘fit’and preparedness for engineering education. In:
Proceedings of the Canadian Engineering Education Association (CEEA) (2016)
31. Panisoara, G., Duta, N., Panisoara, I.O.: The influence of reasons approving on student
motivation for learning. Procedia Soc. Behav. Sci. 197, 1215–1222 (2015)
32. Yacob, A., Saman, M.Y.M.: Assessing the level of motivation in learning programming among
engineering students. In: The International Conference on Informatics and Applications
(ICIA2012), pp. 425–432. Malaysia:[sn] (2012)
 Prediction of Software Reliability Using Particle
Swarm Optimization

Getachew Mekuria Habtemariam1 , Sudhir Kumar Mohapatra2(B) ,

and Hussien Worku Seid1
1 Addis Ababa Science and Technology University Addis Ababa, Addis Ababa, Ethiopia
[email protected], [email protected]
2 Sri Sri University, Cuttack, Odisha, India

[email protected]

Abstract. The quality of Software comprises many features constituting of soft-

ware reliability. Estimating of software reliability in the initial stage of platform
establishment will allow a software professional in originating tables well as defect
long-suffering software. Testing and maintaining Software is terribly exorbitant
and strenuous, and it has been predicted that about half of software establishment
expenses are designated to validating of the software. In view of this we propose
nature inspires methods of Particle Swarm Optimization (PSO) based model to pre-
dict software failure. The proposed model is compared with some existing bench-
mark techniques like Neural Networks (NN), Support Vector Machine (SVM),
Logistic Regression, K-Nearest Neighbour (KNN), Random Forest, and genetic
algorithm (GA). The dataset considered for experiments are taken from NASA
Promise Software Engineering Repository projects. The prediction generated by
PSO is more accurate as compared with other benchmark techniques.

Keywords: Software reliability · Particle swarm optimization · Parameter

estimation

1 Introduction

A reliability of software is one of the indispensable attribute of system perfection. In con-

sideration of that, more and more scholars are giving attention to it. Various researchers
are investigating the reliability prediction of software utilizing various search algorithms,
as a consequence of our dependability on software systems is rising, we are coming to
have more exposure to the impairment caused because of software collapses. Software
dependability is a quality element and establishing a sound software creation is a chal-
lenging duty. Consequently, reliability is a quality cornerstone and is recognized as
the capability of a software to achieve its expected functions adequality in a period of
given time and a given circumstance. Performance of software declines may happen in
accordance with environmental factors, but gracefully [1].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 148–156, 2022.
https://doi.org/10.1007/978-3-031-23233-6_11
 Prediction of Software Reliability 149

The measurement of software reliability is intended to help in the subsequent ways:

i. Determining if a proper amount of evaluation is done or not; if done then we should

move to next stage, if not continue evaluating. For instance, we can determine to
stop the checking when the reliability reaches its margins.
ii. Organizing our maintenance task according to the acquired data or in some cases
we can predict the possible scenarios for the subsequent type of the product; i.e.,
should the subsequent type be functioned as expected or not.
iii. As a complication of optimization character of the problems meticulously investi-
gate and evaluate all possible fact and figure in order to reach accurate and sound
prediction.

Increasing gains or decreasing disasters has always been a prime importance in

software reliability issues. For various area of discipline, acomplication of enhancement
issues which maximizes science and technology advancement. Frequently, instances of
computing issues which might be need an optimization method in software engineering
particularly in software reliability and other like in power energy renovation and delivery,
in electronic design, in interconnected structure deign, and in reload of atomic activator.
To increase or decrease a performance in order to identify the best, solution for a given
problem, there are various methods that could be accomplish optimization methods.
Even though a broad scope of optimization procedures that could be applied, which
is not the vital one that is recognized to a promising of the best for any situation. In
an optimization approach which is appropriate for one problem might not be so fit for
another problem; it relies on various elements, for instance, in case the operation is
dissimilar and its concavity which is convex or concave. So as to settle an issue, we
have to recognize variety optimization approaches therefore, the system developer good
enough for choosing the procedure which is the most tailored through the characteristic
of the optimized problem [2, 3].
Concentration in using developmental data processing to settle software reliability
prediction as well as software reliability problems prolonged in the current years. The
evaluation of the model parameters employing Genetic Algorithms and Particle Swarm
Optimization for Software developments is straightforward as well as simple to learn
[4].
Particle swarm optimization encompasses a very simple idea, and models which are
executed with a small number of lines of computer code. It requires simply elementary
computational operators and is analytically economical in connection with both storage
and speed requirements. The initial testing has found the application to be efficient with
various types of problems [5, 6].
In the next section, we discuss some of the related work for our research. In Sect. 3
we address the proposed PSO model followed by the experiment and result discussion
in Sect. 4. Section 5 also discuss concludes and future insight of the research.

2 Related Work
Reliability of software has regularly been among the major issues to the scholars for elon-
gated a period of time. Fenton discusses about the software measurement and metrics
150 G. M. Habtemariam et al.

[7]. Reliability has been eminently determining humankind ever meanwhile we adapted
to form groups or communities among ourselves. Communitie sreveal the interrelation-
ship; as well reliability is at the centre of interrelationship. Nowadays, software has tend
to the indispensable parts of our lives. Evaluating or estimating reliability has ever been
an intuitive operation. In the event of software, investigates have been perform to create
this operation more scientific instead of intuitive. The current paper enhances a bit in
this series of investigates [10].
Software reliability is an optimization problem which is no single model can answer
the problem of software reliability which mean one model fit for one problem it may
not be fit for another problem. Therefore, following optimization techniques for solving
such type problem is crucial. Optimization approaches motivated by Swarm intelligence
have tend to more common during the previous decade. Swarm intelligence perhaps
used to various aspects of software engineering [10].
Developmental intelligence-based approaches, for instance Genetic Algorithm and
Particle Swarm Optimization perhaps an answer to such type issues. PSO is a search
space method utilized in computing discipline and in engineering field to obtain the
appropriate answers to optimization the problems and that was innovative through the
community characteristics of bird swarming and fish schooling. PSO has its origins
in imitation of life and community way of thinking, as well in engineering field and
computing technology.It exploits a community of individual that glide over the problem
hyperspace within given paces [8].
In current assessments, Particle Swarm Optimization (PSO) is another search
method, and regularly outshined from Genetic Algorithm as when used to several prob-
lems. This takes up the request of how PSO challenges with Genetic Algorithms in the
context of developmental structural testing [9].
PSO plan of action is exploited to pay attention of software establishment undeviating
quality development exhibiting issues to a great degree applied as a part of the docu-
menting in the Logarithmic, Exponential, Power, S-Formed and Converse polynomial
model [2].
Particle swarm enhancement, as a novel global optimization technique, has been
used in parameters estimation and quadratic programming. So, it is available for Particle
swarm enhancement to optimize different type software reliability models parameters
[11].

3 Reliability Prediction Algorithm Using PSO

Software reliability prediction proposed algorithm.

Algorithm 1. Initializing the particle

2. Repeat
2. Calculate the fitness value (ft)of each particle
3. if ft > pBest

Update pBest with ft

4. Assigned best pBest value of the particle to gBest

 Prediction of Software Reliability 151

5. Calculate velocity and update the position

6. Until no updation is possible in gBest

The above PSO algorithmstops when there is no update in the gBest value. The initial
value are the software error of individual projects.The fitness function is the normalised
mean squre error(NRMSE).
 
  
 n −y 2 y
NRMSE =  i=1 iy i 2 (1)
i=1 yi

The software error data is a pair having time and cumulative software failure infor-
mation is present. In Eq. 1, n is represent the number of failure data of a software, and
yi represent is the actual error data of particle I where as yi is the predicted error.

4 Experimental Result and Comparison

The algorithm is implemented on python language. The experiment is carried out in a 11th
Gen Intel(R) Core(TM) i7-1165G7 @ 2.80 GHz 2.80 GHz machine. For our research
purpose, we use the following Metric Data Program (MDP) [12] dataset from NASA
Promise Software Engineering Repository projects which support software developers in
inspection test status and estimating schedules and used to verifying software reliability
models (Table 1).

Table 1. Metric data program (MDP) dataset

Project Language KLoc No. module Defect module

0 CM1 C 20 505 10.0
1 KC3 Java 18 458 9.0
2 KC4 Perl 25 125 49.0
3 MC1 C&C++ 63 9466 0.7
4 MC2 C 6 161 32.0
5 MW1 C 8 403 8.0
6 PC1 NaN 40 1107 7.0
7 PC2 NaN 26 5589 0.4
8 PC3 c 40 1563 10.0
9 PC4 NaN 36 1458 12.0
10 PC5 c 164 17168 3.0

Figure 1 shows the lines of code, number of module size and defect module which is
important for our research in order to exhibit the efficiency of detecting of error capacity
by different software reliability methods.
152 G. M. Habtemariam et al.

Fig. 1. Metric data program (MDP) dataset

Based on the above dataset we were chosen seven classic Software Reliability pre-
diction technique to calculate detection of error accuracy. The testing results we obtained
on each data extracted from its formulas and algorithms which is the maximum capacity
of detecting software error.

Table 2. Detecting error capacity of the techniques

Detecting error capacity

Deft. module NN SVM Logitic.reg KNN Ran.forst GA PSO
CM1 (10) 78.89 92.01 85.23 87.5 91.7 92.67 93.89
KC3 (9) 88.01 77.56 78.98 90.23 77.89 90.67 91.78
KC4 (49) 83.12 91.03 81.45 83.56 88.61 73.35 89.56
MC1 (.7) 91.03 88.23 90.32 78.76 90.12 77.36 93.02
MC2 (32) 87.89 90.68 90.56 77.89 86.45 91.61 92.01
MW1 (8) 83.45 79.67 88.36 82.29 73.36 91.45 93.65
PC1 (7) 86.28 86.37 91.07 82.57 85.21 89.51 90.57
PC2 (.4) 74.84 78.31 81.27 82.24 87.31 91.25 94.01
PC3 (10) 82.23 85.23 87.23 90.01 77.25 88.25 93.12
PC4 (12) 78.23 76.12 83.35 81.59 84.21 90.78 92.35
PC5 (3) 90.12 89.45 90.37 78.29 89.27 91.38 93.45

From Table 2, we use ten-fold cross-validation technique and average the results
over the folds. We made a comparison to our algorithm with seven different machine
learning techniques such as Neural Networks (NN), Support Vector Machine (SVM),
Logistic Regression, K-Nearest Neighbour (KNN), Random Forest, Genetic Algorithm
(GA) and Particles Swarm Optimization (PSO). The outcome indicates that Particles
Swarm Optimization achieve the maximum result. As we know software reliability
is an optimization problem and results shows that PSO outshines in each project we
examined Thus, PSO have shown their ability to provide an adequate and best error
detecting capacity for predicting reliability of a software (Fig. 2).
Table 3 displays the Statistical Summary of mean, standard deviation, minimum and
maximum result of the experiment which shows Particles Swarm Optimization achieve
the maximum result (Fig. 3).
 Prediction of Software Reliability 153

Fig. 2. Detecting error capacity of the ML techniques

Table 3. Statistical summary of the experiment

NN SVS log. r KNN rad. f GA PSO

Count 11 11 11 11 11 11 11
Mean 84.01 84.2 86.2 83.18 84.67 88.03 92.49
Std 5.2 5.98 4.37 4.37 5.96 6.43 1.42
Min 74.84 76.12 78.98 77.89 73.36 73.35 89.56
25% 80.56 78.99 82.4 80.18 81.05 88.88 91.9
50% 83.45 86.37 87.23 82.29 86.45 90.78 93.02
75% 87.95 90.06 90.34 85.53 88.94 91.41 93.55
Max 91.63 92.01 91.07 90.23 91.7 92.67 94.01

Finally, we conclude from the above experiment Particle Swarm Optimization is

broadlyemployed in multiple researchdomains and real-world utilizationas a tremendous
optimization practicewhich is imitating the natural behaviour, of swarm as well Particle
Swarm Optimization technique has a collection of swarm particles that fly around n
dimensional problem space in search of an optimal solution.
154 G. M. Habtemariam et al.

Statistical Summary of the experiment

100
90
80
qu 70
an 60
50
- 40
30
20
10
0
NN SVS log.r KNN rad.f GA PSO
Machine Learning Techniques

count Mean Std Min 25% 50% 75% Max

Fig. 3. Statistical summary of the experiment

It is regularly noble for us to assess the relationships of the attributes in our dataset
using Particle Swarm Optimization into machine learning project as some machine learn-
ing techniques like linear regression and logistic regression will perform inadequately
if we have hugely associated with attributes. Figure 4 shows the relationships between
each machine learning techniques (Table 4).

Table 4. Relationships between each machine learning techniques

NN SVS log. r KNN rad. F GA PSO

NN 1 0.38 0.53 −0.31 0.02 −0.25 −0.26
SVS 0.38 1 0.49 −0.23 0.61 −0.36 −0.2
log. r 0.53 0.49 1 −0.56 0.12 0.08 0.14
KNN −0.31 −0.23 −0.56 1 −0.44 0.12 −0.02
rad. f 0.02 0.61 0.12 −0.44 1 −0.29 −0.06
GA −0.25 −0.36 0.08 0.12 −0.44 1 0.54
PSO −0.26 −0.36 0.14 −0.02 −0.06 0.54 1
 Prediction of Software Reliability 155

0
NN svs log.r KNN rad.f GA PSO
-1

-2

NN svs log.r KNN rad.f GA PSO

Fig. 4. Correlations between each machine learning techniques

5 Conclusions
Because of the increasing desire of software with highly reliable and safety software,
therefore, software reliability prediction emerges as more and more important. Software
reliability is a vital component of software quality. Here we design a novel concept
of particle swarm optimization (PSO) algorithm which is used for software reliability
prediction. The proposed model is tested with 11 number of dataset taken from NASA
Promise Software Engineering Repository. The prediction given by PSO is compara-
tively more accurate as compared to other techniques. The error-detecting capability is
varying from 89.56% to 94.01%. Whereas in other models it is much comparatively less.

References
1. Malhotra, R., Negi, A.: Reliability modeling using particle swarm optimization. Int. J. Syst.
Assur. Eng. Manage. 4(3), 275–283 (2013)
2. Shin, S.M., Uroosa, S.: Predicting software reliability using particle SWARM optimization
technique. Asia-Pac. J. Convergent Res. Interchange 1(3), 17–30 (2015)
3. de Almeida, B.S.G., Leite, V.C.: Particle swarm optimization: a powerful technique for solving
engineering problems. In: Ser, J.D., Villar, E., Osaba, E. (eds.) Swarm Intelligence – Recent
Advances, New Perspectives and Applications. IntechOpen (2019)
4. Sheta, A.: Reliability growth modeling for software fault detection using particle swarm opti-
mization. In: 2006 IEEE International Conference on Evolutionary Computation, pp. 3071–
3078. IEEE (2006)
5. Eberhart, R., Kennedy, J.: A new optimizer using particle swarm theory. In: MHS’95. Proceed-
ings of the Sixth International Symposium on Micro Machine and Human Science, pp. 39–43.
IEEE (1995)
6. Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of ICNN’95-
international conference on neural networks, vol. 4, pp. 1942–1948. IEEE (1995)
7. Fenton, N.: Software measurement: a necessary scientific basis. IEEE Trans. Software Eng.
20(3), 199–206 (1994)
156 G. M. Habtemariam et al.

8. Del Valle, Y., Venayagamoorthy, G.K., Mohagheghi, S., Hernandez, J.C., Harley, R.G.: Par-
ticle swarm optimization: basic concepts, variants and applications in power systems. IEEE
Trans. Evol. Comput. 12(2), 171–195 (2008)
9. Windisch, A., Wappler, S., Wegener, J.: Applying particle swarm optimization to software test-
ing. In: Proceedings of the 9th Annual Conference on Genetic and Evolutionary Computation,
pp. 1121–1128 (2007)
10. Ahuja, N.G.T.: A review on particle swarm optimization for software reliability. Environment
3(3), 213–214 (2014)
11. Can, H., Jianchun, X., Ruide, Z., Juelong, L., Qiliang, Y., Liqiang, X.. A new model for
software defect prediction using particle swarm optimization and support vector machine. In:
2013 25th Chinese Control and Decision Conference (CCDC), pp. 4106–4110. IEEE (2013)
12. Banga, M., Bansal, A., Singh, A.: Proposed hybrid approach to predict software fault detection.
Int. J. Performability Eng. 15(8), 2049 (2019)
 An Effective Optimization of EMG Based
Artificial Prosthetic Limbs

Dwarika Nath Choudhury(B) and Narayan Nayak

Silicon Institute of Technology, Bhubaneswar, India

[email protected]

Abstract. Humans value with their physical parts are the greatest of all the pos-
sessions. The human hand is capable of a broad range of dexterous maneuvers that
allow us to interact with our surroundings and communicate with one another. In
this paper our main aim is to attempt for restoration of amputated limbs with artifi-
cial limbs for millennia. The difficulty of replacing a lost human limb, particularly
a hand, lets one properly understand the complexities in a human. In this paper,
early designs are drawn in order to finalize the model for 3d modelling. 3d mod-
els are then designed using CATIA v5 modelling software. The size of modelled
limbs are same as that of a human being. Using a 3d printer, each component of
the limb is printed. 3d printed components are cured with light in order to remove
moisture from it. Actuators are then set up with the limbs. The controller is then
programmed with the EMG sensors. The analog signals from the EMG sensors
are amplified resulting in motion of the limbs. The angular rotation angles are then
recorded by keeping different subjects on it. Grasping capacity is calculated and
compared with that of human limbs. The optimized result is recorded.

1 Introduction

A number of ancient prosthetic devices from many cultures throughout the world have
been found, illustrating the development of prosthetic technology. The development of
prosthetic limb design has been somewhat gradual up until recently. Simple prosthetic
devices can be looked of as early breakthroughs like the wooden leg. History demon-
strates that prostheses have traditionally been passive tools that provide nothing in the
way of control or movement. Modern prosthetic hands have been created to closely
resemble natural limbs in terms of both shape and function. Although the bionic hand
has lately been lauded as a victory of engineering prowess, it still falls short of the gen-
uine thing and as a result, there are a number of obstacles preventing the upper limb
amputee community from adopting it. The prosthetic hand is unable to achieve the com-
plete acceptance of its users, which is the ultimate objective of any prosthesis. The topic
of myoelectric prosthetic arms will be covered in this thesis. The goal is to create a
machine that performs human arm functions.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 157–171, 2022.
https://doi.org/10.1007/978-3-031-23233-6_12
158 D. N. Choudhury and N. Nayak

A large amount of prosthesis are present in current world. Some of them are as
follows:

1. Passive Prostheses
Simple, immobile limbs known as passive prostheses are designed to help
amputees regain their fundamental functioning and aesthetic appeal. A straight-
forward passive prosthesis is an item like a wooden “pirate” peg leg.
2. Mechanical Control Prostheses
Control of mechanical powered prosthesis is accomplished through a harness
fastened to the user. They typically consist of a straight forward tool like a hook
that is connected to shoulder and elbow. Although these gadgets are very straight
forward, they continue to be the most common kind of prosthesis in use today.
3. Myo-signal Controlled Prostheses
Myo-signal controlled prostheses track the electromyography signals that are
produced when muscles contract. Through electrodes attached to the muscle, these
impulses are monitored.
These signals are then modulated and sent to the controller. These signals are
then amplified and processed by micro-controllers for proper control of the limbs.
4. Brain Interface
The best type of control is through brain signals. Generally, EEG sensors are
used to fetch the neural signals. These are signals are then amplified and passed
through certain instructions so as to work ideally.

2 Literature Review
Kato et al. [1] proposed a model of communication between brain and prosthetic limbs
using myo-sensors. The signals were then modulated and amplified using a controller
board. The relation between muscle readings and movement of the limbs were then
established. Hussein et al. [2] introduced manufacturing of tiny parts related to prosthesis
using 3d printer. In North America small manufacturers constructed mechanical limbs
using 3d printed parts. Moreno et al. [3] implemented cheap and affordable mechanism
of prosthesis. He also introduced leg prosthesis. They proposed a IOT embedded real
time operating system for the prosthesis. The paper throws more light on circuits to
process and achieve ideal prosthesis. Melchiorri et al. [4] focuses on development of
prosthetic hands. The paper is well suited to a single domain. The use of these type
of hand can also be implemented in humanoid robots. Full replication of human hand
is targeted to achieve through this publication.Clement et al. [5] emphasizes on design
of light weight prosthetic hands and weight distribution of the prosthetic limbs. The
control of lateral balancing while movement of hand is Kaplanoglu et al. [6] The shape
memory alloy (SMA) wires that make up the finger tendon act as muscle pairs to flex and
extend the finger joints as needed. Three finger’s four degrees of freedom are actively
used. Abhishek et al. [7] introduced introduced the use of mathematical-algorithms to
recognize human gestures. They how gestures of hand can turn pages, scroll up and
down, etc. Atique et al. [8] introduces a cost effective prosthetic hand using myoelectric
signals. Analog signals are simulated and result is improvised using x and y coordinates.
The motions of limbs are processed accordingly.
 An Effective Optimization of EMG 159

3 Design and Manufacturing

A prosthetic tendon’s design was finalized after examining several actuation strategies.
The wires (green lines) attach to the fingers and are tightened by the motors. The fingers
open and close by a result of pulling on the tendons. To make it more portable and
attachable, the electric motors are entirely enclosed within. The ideal location for these
motors is as close to the fingers (Figs. 1 and 2).

Fig. 1. Early design mechanism (1)

Fig. 2. Early design mechanism (2)

160 D. N. Choudhury and N. Nayak

3.1 Cad Model

Catia is the application software used in modelling of prosthetic limbs (Figs. 3, 4, and
5).

Fig. 3. Catia model offinger prototype

Fig. 4. Catia model of thumb prototype

 An Effective Optimization of EMG 161

Fig. 5. Catia model of palm prototype

The below-displayed closed loop is made by the tendons wrapping around specially
made, 3D-printed servo horn. The tendon is pulled while the servo motor turns in one
direction, closing the finger (Table 1). The motor is turned counterclockwise to release
the finger (Figs. 6 and 7).

Fig. 6. Catia model of forearm prototype (inner)

162 D. N. Choudhury and N. Nayak

Fig. 7. Catia model of forearm prototype (outer)

Table 1. Parameters of CAD modelling

Parameter Dimensions
Length of profile 370 mm
Thumb (height) 40 mm
Thumb (width) 18.5 mm
Index-finger (length) 75 mm
Index-finger (width) 24 mm
Middle-finger (length) 78 mm
Middle-finger (width) 27 mm
Ring-finger (length) 73 mm
Ring-finger (width) 24 mm
Pinkie-finger (length) 64 mm
Pinkie-finger (width) 20 mm
Forearm (height) 163 mm
Forearm (width) 60 mm
Wrist size 42 mm
Palm (height) 58 mm
Palm (width) 53 mm

3.2 Manufacturing and Assembly

All components have been printed using Flashforge Creator pro 2. This type of 3-D
printer produces small parts with high precisions. The printer automatically prints an
outer cover for each part as a layer of protection. Each of the screws used within the
model like finger joints, are 3mm in diameter. Polypropylene is used for 3-D printing.
 An Effective Optimization of EMG 163

Pin-holes were done using a drill after printing. The printer is perfect for printing of tiny
subjects. It gives very high precision while printing (Figs. 8, 9 and 10).

Fig. 8. 3D printed part (i)

Fig. 9. 3D printed parts (ii)

164 D. N. Choudhury and N. Nayak

Fig. 10. 3D printed part (iii)

4 Electrical Components and Design

These motors can movie around 360° in clockwise and anticlockwise direction. The
angular accuracy of each servo motor influences how each finger moves when the limbs
move to open and close finger. Relatively cheap servos are used to decrease production
cost. Using higher quality servos would increase strength of limbs and accuracy.
The Arduino Uno is used initially for as micro-controller. The EMG sensors are
connected with the controller. The controller receives analog signals in real time which
after processing results in rotation of servos, ultimately resulting in movement of limbs.
Initially this micro-controller is used in order to make the mib cost-effective.

4.1 Electromyography Sensing

Electromyogram device board is used for detecting and monitoring live activity of the
muscle. Three electrodes and a tiny PCB are included. Three electrodes are used as: two
electrodes help in monitoring voltage potential, while the third one is used as ground
point. As a user flexes, the interior systems converts electrical signals into a correspond-
ing ironed signal, then used as input to a microcontroller’s analogue to digital convertor
(Fig. 11).
 An Effective Optimization of EMG 165

Fig. 11. Connection flow chart.

When a user flexes, it creates an analog signal that is amplified by the EMG sensor.
The Arduino board uses this analog signal to create a movement. This moves servos by
whose tension, the limbs cause the limbs to move (Fig. 12).

Fig. 12. Servo processor interfacing

The following code says that the signal pin is connected to pin 10 (pwm) in aurduino
will turn a servo motor from 0° to 360, wait for 15ms delay, then turn it back from 360
to 0° back.
166 D. N. Choudhury and N. Nayak

5 Artificial Intelligence
5.1 Gesture Recognition
Gesture focus is the method used to apprehend and analyze human physique behavior.
The in-flip helps in making a channel between the computing device and the consumer.
Gesture cognizance is helpful in processing the records that can’t be conveyed through
verbal or written content.

• Training library contains different hand gestures used in training.

• Library includes determination of the centroid.
• Hand Detection: first the photo is taken from web camera. The image is taken at
30 frames per second. The distance between the camera and the hand should range
between 30 cm to 120 cm.

TEST AND RESULTS

5.2 Grasping Capacity (According to Size)

Testing with small essential objects and comparing it with grasping capacity of human
hand and plotting a graph.
Aim of the experiment:- To test the maximum grasping capacity of our designed
prosthetic limb.

Fig. 13. Flow chart of grasping capacity

Procedure:- small objects which are really essential for an amputee or prosthetic
limb user were taken into account such as glass of water, pen, pendrive, wallet, etc.
 An Effective Optimization of EMG 167

These objects were taken for the experiment and then their circumference were noted
down by the traditional thread and ruler method and noted down. Then for each object
it was marked upto what circumference our hand can grasp objects. After all readings
were taken a graph was plotted to compare it with the functionality of the human hand
(Fig. 13) (Table 2).

Table 2. Weight of different parts of prosthetic hand

Parts Weight (in g)

3D printed parts 156
Servomotors (5) 5 * 13 = 65
Lithium Polymer battery 180
Rest parts: Wires, Tendons, Arduino Nano 50
Total 451

5.3 Analysis of EMG Signals

Analysis of EMG signals that are detected from the user muscles have been analyzed
by taking the EMG signals of 10 different persons (Subject) to check the rotation of
servomotor, as the EMG signals vary from person to person.
As physiological, anatomical and biochemical characteristics of persons are different
from each other, so that the range of EMG signals changes from person to person. Some
factors like height, weight of a person, strength of muscles, placement of electrodes also
influence the values of EMG signals. So, to analyses the changing value the readings of
EMG signals for 15 s of different 10 persons have been taken and presented in graph
(Tables 3, 4 and 5).

Table 3. Basic data list taken from different persons

Subject Height Weight (kg) Arm length (cm)

01(M) 5 3 92 25
02 5 5 88 24
03 5 8 61 27
04 6 66 27.3
05 6 110 27
06 5 10 100 26.4
(continued)
168 D. N. Choudhury and N. Nayak

Table 3. (continued)

Subject Height Weight (kg) Arm length (cm)

07 5 11 73 24.5
08 4 9 39 23.6
09 4 8 52 23
10 5 3 61 23.6

Table 4. EMG Signals of different subjects in each second and rotation angle (a)

Time (s) Subject-1 Subject-2 Subject-3 Subject-4 Subject-5 Servo rotation angle (°)
1 408 405 276 234 208 170
2 377 370 279 210 201 170
3 380 382 284 199 202 170
4 395 392 290 208 195 170
5 380 325 286 244 199 170
6 343 340 290 239 207 170
7 331 328 297 236 212 170
8 342 340 283 222 224 170
9 311 250 292 238 242 170
10 309 108 279 241 225 170
11 309 208 273 232 229 170
12 322 120 285 241 235 170
13 320 150 298 238 237 170
14 323 310 283 245 228 170
15 319 122 274 241 231 170

From the above characteristics, it is observed that the servo mechanism performs
almost the same for all subjects (i.e., subject-1, subject-2, subject-3, subject-4, subject-
5, subject-6, subject-7, subject-8, subject-9, sub-10) irrespective of their body weight,
height and arm length. Though the servo mechanism holds good for all different subjects,
it gives a good response for the operation. So, it can be concluded that this setup can
easily be used for any amputees (Fig. 14).
 An Effective Optimization of EMG 169

Table 5. Signals of different subjects in each second and rotation angle of servo motor (b)

Time (s) Subject-6 Subject-7 Subject-8 Subject-9 Subject-10 Servo rotation angle (°)
1 158 250 208 246 209 170
2 178 256 216 236 211 170
3 159 236 213 235 225 170
4 15 260 212 224 229 170
5 154 265 200 258 227 170
6 152 261 194 256 210 170
7 181 287 184 254 237 170
8 161 290 188 243 239 170
9 160 268 189 256 222 170
10 131 261 193 258 235 170
11 153 252 195 256 248 170
12 135 250 200 242 244 170
13 131 244 252 242 234 170
14 144 246 255 242 235 170
15 146 246 253 230 236 170

6 Conclusion

This project aims in design of the ideal prosthetic limbs controlled by EMG signal. Each
of the finger has three axis of rotation. Each finger is powered by using servo motors
controlled by micro-controller. The prosthetic limb is processed by the EMG signals.
The described EMG setup can operate only one finger so if we want to operate all five
fingers of the designed prototype then we need 5 EMG sensors as well as 15 electrodes.
This implies that a total of 11 electrodes are required while complete functioning of the
limbs.
When EMG signals of 10 different persons were taken it is observed that the servo
mechanism performs almost the same for all persons irrespective of their body weight,
height and arm length. The servo motor is rotating about 170° which helps to move the
finger of the prosthetic hand.
Though the servo mechanism holds good for all different subjects, it gives a good
response for the operation. So, it can be concluded that this setup can easily be used for
any amputees without any maloperation.
170 D. N. Choudhury and N. Nayak

Fig. 14. Characteristics cure of EMG signals

References
1. Kato, R., Yokoi, H., Arai, T.: Real-time learning method for adaptable motion-discrimination
using surface EMG signal. In: 2006 IEEE International Workshop on Intelligent Robots and
Systems (IROS), pp. 2127–2132 (2006)
2. Ahmed, S.F., Kamran Joyo, M., Mahdi, H.F., Kiwarkis, I.J.: Muscle fatigue detection and
analysis using EMG sensor. In: 2020 IEEE 7th International Conference on Engineering
Technologies and Applied Sciences (ICETAS), pp.1–4 (2020)
3. Cifuentes, C.A., Braidot, A., Frisoli, M., Santiago, A., Frizera, A., Moreno, J.: Evaluation of
IMU ZigBee sensors for upper limb rehabilitation. In: Converging Clinical and Engineering
Research on Neurorehabilitation, pp. 461–465 (2013)
4. Palli, G., Scarcia, U., Melchiorri, C., Vassura, G.: Development of robotic hands: the UB
hand evolution. In: 2012 IEEE/RSJ international conference on intelligent robots and systems,
pp. 5456–5457
5. Clement, R.G.E., Bugler, K.E., Oliver, C.W.: Bionic prosthetic hands: a review of present
technology and future aspirations. Surgeon 336–340 (2011)
 An Effective Optimization of EMG 171

6. Akgun, G., Cetin, A.E., Kaplanoglu, E.: Exoskeleton design and adaptive compliance control
for hand rehabilitation. In: Transactions of the Institute of Measurement and Control, pp. 493–
502 (2020)
7. Rai, V., Sharma, A., Rombokas, E.: Mode-free control of prosthetic lower limbs. In:2019
International Symposium on Medical Robotics (ISMR), pp. 1–7 (2019)
8. Atique, M.D., Moin, M.S., Siddique, R.: A cost-effective myoelectric prosthetic hand, journal
of prosthetics and orthotics, pp. 231–235 (2018)
Communications
 Performance Analysis of Fading Channels
in a Wireless Communication

Pradyumna Kumar Mohapatra1(B) , Saroja Kumar Rout2 , Ravi Narayan Panda3 ,

Anudeep Meda4 , and Banoj Kumar Panda5
1 Department of ECE, Vedang Institute of Technology, Bhubaneswar, Odisha, India
[email protected]
2 Department of IT, Vardhaman College of Engineering (Autonomous), Hyderabad, Telangana,
India
3 Department of ECE, GIFT, Bhubaneswar, Odisha, India
4 Department of IT, Vardhaman College of Engineering (Autonomous), Hyderabad, India
5 Department of ECE, GIET, Gunupur, Odisha, India

Abstract. Wireless communication relies heavily on system performance. It is

very important to consider channel behavior when designing wireless communica-
tion systems. Fading, scattering, interference, and other channel aspects affect the
received signal quality. In MATLAB 2015 simulations, Rayleigh, Rician, and Nak-
agami fading channel models were compared for fading envelope, signal power,
and channel capacity. Multipath fading environments require the use of parameters
such as source velocity and pdf to analyze and design digital communication sys-
tems. The present study analyzes and simulates wireless channel behavior under
different distributions of fading.

Keywords: Fading channel · Channel capacity · Nakagami channel

1 Introduction

Although wireless technology has exploded in popularity, some unavoidable conditions,

signal attenuation, and barriers have made it difficult for the system to achieve its best
results. An antenna that transmits and receives wireless signals from a simple line of
sight to complex barriers like buildings, mountains, etc. In contrast to fixed channels,
mobile channels are unpredictable and very different, due to their randomness. Faded
communication is characterized by several multipath mechanisms’ different times of
arrival at the receiver, which is particularly bothersome in wireless communications [1–
3]. The mobility of the transmitter and receiver as well as the signal bandwidth are a
few of the parameters that can have an impact on a channel’s fading and multipath delay
nature [4]. In wireless communication, some of the undesired signals may interfere with
the useful signal which may affect the performance of the channel. An effective channel
equalizer is needed which is positioned at the receiver side to eliminate these multipath
effects. Thus the role of the equalizer is one of the vital parts of wireless communication.
Channel equalization itself is an optimization problem [5]. In training neural networks

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 175–183, 2022.
https://doi.org/10.1007/978-3-031-23233-6_13
176 P. K. Mohapatra et al.

for channel equalization, evolutionary algorithms were used extensively [6–11]. The
node location algorithm adopts distance measurement as a TDOA technique and the
static sensor node can localize when it comes into the reception range of the channel
[12].
Using a vehicular environment, this paper develops Rayleigh, Rician, and Nakagami-
m fading algorithms. Frequency selective fading is categorized under multipath time
delay spread in addition to flat fading. Fast fading is observed on the channel when
its coherence time is small compared to its delay constraint. In the Nakagami fad-
ing model, multipath scattering with large delays is considered, with multiple reflected
waves clustered together. Each reflected wave within a cluster has a random phase, but
all waves have approximately equal delay times. Hence, each cumulated cluster signal
has a rayleigh envelope. Multipath reception is considered the cause of Rayleigh fading.
Using a Rayleigh distribution, the Rayleigh fading model predicts that a signal’s magni-
tude will vary randomly through the transmission medium. A Rayleigh fading effect is
most commonly observed when line-of-sight propagation is not dominant between trans-
mitters and receivers. As a result of the Rician model, the dominant wave is composed
of several dominant signals, for instance, the line of sight and the ground reflection. A
deterministic process is then used to treat the combined signal, with shadow attenuation
also applying to the dominant wave.
The rest of this work is structured as follows:
The technique for Signal propagation is affected by fading channels described in
Sect. 2. In Sect. 3, the study’s findings are given and thoroughly discussed. Section 4
marks the conclusion of the research work.

2 Fading Channels
Signal propagation is affected by fading channels due to elements including shadowing,
multipath propagation, and geographic conditions. Rayleigh, Rician, and Nakagami,
channels [13] are some of the fading channels that are commonly used.
The following expressions can be used to determine the capacity of the Rayleigh,
Rician, and Nakagami fading channels:
  
1 bits
C = log2 det In + 2 HRx H (1)
σ Hz

Here, C refers capacity of the channel, mutual information denoted as In whereas the
channel matrix identified as H and the signal envelope mentioned as Rx .

2.1 Performance Analysis of Rayleigh Fading

The transmission and reception of wireless signals are subject to LOS conditions. How-
ever, in specific towns LOS conditions are difficult to achieve; consequently, the receiver
detects multipath signals. The amplitudes and phases of these multipath signals are dis-
tributed. In the absence of LOS, these signals are subject to Rayleigh fading at the
 Performance Analysis of Fading Channels in a Wireless Communication 177

receiver side [14] Summation of two Gaussian noise signals in quadrature is governed
by Rayleigh distributions [15–17]:
 
h − h2
fh (h) = p(h) = e 2σ 2 h≥0 (2)
σ2
Ithis case, σ is the root mean square of the voltage signal received before envelope
detection. In-phase and quadrature components of the received signal are demodulated.
Using the following formula, we can calculate the envelope of a received signal:

x(t) = P 2 (t) + Q2 (t) (3)

Here, P(t) and Q(t). Ardefined as the in-phase and quadrature Signal. Based on the
central limit theorem, signals can be entirely described by their means and autocorrelation
functions,

2.2 Description of the Performance of Rician Fading Channel

An envelope distribution that exhibits fading at small scales is rician when the majority
of the signal component is stationary (non-fading), such as a line-of-sight transmission
channel. In this case, a dominating signal that is steady has random multipath components
that are coming at various angles superimposed on it. This has the result of increasing the
random multipath’s dc component at the output of an envelope detector. Several weaker
multipath signals interact to produce the Rician distribution. An enveloped Rayleigh sig-
nal is formed when the dominant signal becomes weaker. Rician distributions degenerate
into Rayleigh distributions when the dominant component disappears. The transmitted
signal can be written as the following equation when such a path exists.
N −1
c(t) = lj cos(wc t + wsj t + ϕj ) + ks cos(wc t + ws t) (4)
j=1

Here, ks is denoted as the strength of the direct component,ws represents the Doppler
shift along the LOS path [18, 19].
The pdf’s derivation in this situation is comparable to that in the Rayleigh scenario.
It is possible to identify the mean and autocorrelation function of Q(t) and P(t) if N
is large enough. They are independent Gaussian processes. The presence of the direct
component in the Rician situation prevents the mean values of P(t) and Q(t) from being
zero. By demodulating the signal s(t), the envelope r(t) of P(t) and Q(t) is derived [2].
In this instance, the Rician density function of the envelope is given by:
 2   
h h + ks2 hks
fh (h) = p(h) = 2 exp − I 0 h≥0 (5)
σ 2σ 2 σ2
178 P. K. Mohapatra et al.

2.3 Performance Analysis of NAKAGAMI-M Fading Channel.

Channel estimation is lacking in the Rayleigh and Rician distributions over long distances
and at higher frequencies. Nakagami distributions, on the other hand, include a density
function coupled to a parametric gamma distribution. As a condition for evaluating
channel behaviour, A LOS component must exist between the transmitter and receiver.
[16, 20].
Here, m is the scaling parameter that describes the amount of fading. As soon as m
goes to ∞, the Nakagami fading channel is no longer available. Equation (6) showed
the probability density function for the said channel:
 
2  m  2m−1 − mh2
fh (h) = p(h) = mh e 
h≥0 (6)
(m) 
Here,
2
m= 2
(7)
E h2 − 

In Eq. (7), the spread parameter is represented as  where  = E h2 (Fig. 1).

3 Experimentation and Result Analysis

3.1 Simulation and Discussion of Rayleigh Fading Channel
In Fig. 2, we can see the envelope of fading and power of the Rayleigh channel. The
signal envelope shows a gradual increase in fading when the user’s speed increases.
Thus, fading is a key problem in wireless communication. In the 2nd part of this figure,
signal power is maximum then substantially decreases. Figure 3 also shows the channel
capacity of the Rayleigh channel with or without channel state information of both
transmitter and receiver.

Fig. 1. Fading envelope and power for a Rayleigh channel

 Performance Analysis of Fading Channels in a Wireless Communication 179

Fig. 2. Channel capacity

3.2 Simulation Results of Rician Fading Channel

Since the same signal is created for both this fading and the preceding Rayleigh example,
the components, envelope, and RF signal values will all be the same. For the Rician
distribution, the probability density function plot will resemble that in Fig. 3.

3.3 Simulation Results of Nakagami-M Fading Channel

Fading parameter for Nakagami channel denoted as m (m ≥ 0.5). It is also called a

shape parameter and the standard Gamma function is identified as (.). A variety of
fading circumstances are covered by the Nakagami-m distribution, it is a one-sided
Gaussian distribution when m = 0.5. The condition for Rayleigh distribution when
m = 1, when m < 1, a more severe fading scenario than Rayleigh fading is used by
the Nakagami model (Fig. 4 and Fig. 5). The capacity variation to SNR is depicted in
Fig. 6. In general, channel state information describes how power decays with distance,
180 P. K. Mohapatra et al.

Fig. 3. Comparison of fading envelope of Rician and Rayleigh distribution for different value K

interference, fading, scattering, etc. Knowing CSI improves the reliability of the signal
transmitted compared to not knowing CSI. As a result, it plays a significant role in
defining the communication relationship. From transmitter to receiver and from receiver
to transmitter, the CSI varies. A comparison graph between the Nakagami and AWGN
channel with CSI of both transmitter and receiver is also shown in Fig. 6.
Performance Analysis of Fading Channels in a Wireless Communication 181

Fig. 4. Fading envelope of Nakagami-m Distribution

Fig. 5. Fading Power of Nakagami-m Distribution

182 P. K. Mohapatra et al.

Fig. 6. Channel capacity with or without CSI for Nakagami- channel

4 Conclusion
Using Rayleigh, Rician, and Nakagami-m Distributions of probability density functions
for respective fading channel models, this paper compares the respective probability den-
sity functions against channel capacity. According to the simulation results, increasing
vehicle speed leads to increased fading in the signal envelope. Consequently, the amount
of fading increases as the signal sinks below the threshold at higher speeds. However,
some fading models perform better than others even after Doppler Effect degrades the
channel capacity. Using better distribution in these models, the channel’s capacity can
be further enhanced. Compared to Rayleigh and Ricchian fading channels, the probabil-
ity density function of the Nakagami-m fading channel grows. In Nakagami-m fading
channel, signal amplitudes of multiple independently dispersed Rayleigh-fading signals
with identical distributions are added together.

References
1. Kumar, S., Gupta, P.K., Singh, G., Chauhan, D.S.: Performance analysis of rayleigh and rician
fading channel models using matlab simulation. Int. J. Intelli. Sys. Appl. 5(9), 94 (2013)
2. Prabhu, G.S., Shankar, P.M.: Simulation of flat fading using MATLAB for classroom
instruction. IEEE Trans. Educ. 45(1), 19–25 (2002)
3. Khan, M. J., Singh, I., Tayal, S.: BER Performance using BPSK modulation over rayleigh
and rician fading channel. In: 2022 IEEE 11th International Conference on Communication
Systems and Network Technologies (CSNT), pp. 434–437. IEEE (April 2022)
4. Bellorado, J., Ghassemzadeh, S., Kavcic, A.: Approaching the capacity of the MIMO Rayleigh
flat-fading channel with QAM constellations, independent across antennas and dimensions.
IEEE Trans. Wireless Commun. 5(6), 1322–1332 (2006)
 Performance Analysis of Fading Channels in a Wireless Communication 183

5. Mohapatra, P.K., Jena, P.K., Bisoi, S.K., Rout, S.K., Panigrahi, S.P.: Channel equaliza-
tion as an optimization problem. In: 2016 International Conference on Signal Processing,
Communication, Power and Embedded System (SCOPES), pp. 1158–1163. IEEE (October
2016)
6. Mohapatra, P.K., Rout, S.K., Bisoy, S.K., Sain, M.: Training Strategy of Fuzzy-Firefly based
ANN in Non-linear Channel Equalization. IEEE Access (2022)
7. Panda, S., Mohapatra, P.K., Panigrahi, S.P.: A new training scheme for neural networks and
application in non-linear channel equalization. Appl. Soft Comput. 27, 47–52 (2015). https://
doi.org/10.1016/j.asoc.2014.10.040
8. Mohapatra, P., Samantara, T., Panigrahi, S.P., Nayak, S.K.: Equalization of Communication
Channels Using GA-Trained RBF Networks. In: Saeed, K., Chaki, N., Pati, B., Bakshi, S.,
Mohapatra, D.P. (eds.) Progress in Advanced Computing and Intelligent Engineering. AISC,
vol. 564, pp. 491–499. Springer, Singapore (2018). https://doi.org/10.1007/978-981-10-6875-
1_48
9. Mohapatra, P., Sunita, P., Panigrahi, S.P.: Equalizer Modeling Using FFA Trained Neural
Networks. In: Soft Computing: Theories and Applications, pp. 569-577. Springer, Singapore
(2018)
10. Pradyumna, M., et al.: Shuffled Frog-Leaping Algorithm trained RBFNN Equalizer. Int. J.
Comp. Info. Sys. Indu. Manage. Appl. 9, pp. 249–256 (2017)
11. Kumar Mohapatra, P., et al.: Application of Bat Algorithm and Its Modified Form Trained
with ANN in Channel Equalization. Symmetry 14(10), 2078 (2022)
12. Rout, S.K., Rath, A.K., Bhagabati, C.: Energy efficient dynamic node localization technique
in wireless sensor networks. Indian J. Sci. Technol. 10(15), 1–8 (2017)
13. Panchal, A., Dutta, A.K.: Performance analysis and design of MIMO power NOMA with
estimated parameters error statistics along with SIC and hardware imperfections. IEEE Trans.
Veh. Technol. 70(2), 1488–1500 (2021)
14. Narayana, M., Bhavana, G.: Performance analysis of MIMO system under Fading Channels
(Rayleigh and Rician) Using SVD PCA and FSVD. journal of engineering technology 5(2),
116–126 (2016)
15. Zhang, D., Zhou, P., Jiang, C., Yang, M., Han, X., Li, Q.: A stochastic process discretization
method combing active learning Kriging model for efficient time-variant reliability analysis.
Comput. Methods Appl. Mech. Eng. 384, 113990 (2021)
16. Tang, L., Hongbo, Z.: Analysis and simulation of Nakagami fading channel with MAT-
LAB. In: Asia-Pacific Conference on Environmental Electromagnetics, 2003. CEEM 2003.
Proceedings, pp. 490–494. IEEE (November 2003)
17. Gvozdarev, A.S.: The novel approach to the closed-form average bit error rate calculation for
the Nakagami-m fading channel. Digital Signal Processing 127, 103563 (2022)
18. Sijbers, J., den Dekker, A. J., Scheunders, P., Van Dyck, D.: Maximum-likelihood estimation
of Rician distribution parameters (2004)
19. Mounika, I.S.D., Sharma, D., Sharma, P.K.: Analysis of different detection techniques of
MIMO in future generation of wireless communication. In: international journal of pure and
applied mathematics 114(12), 419–426 (2017)
 Power Conscious Clustering Algorithm Using
Fuzzy Logic in Wireless Sensor Networks

Sanjaya Kumar Sarangi1(B) , Arabinda Nanda2 , Manas Ranjan Chowdhury3 ,

and Subhadra Mishra4
1 Department of Computer Science and Applications, Utkal University,
Bhubaneswar, Odisha, India
[email protected]
2 Department of Computer Science and Engineering, Krupajal Engineering College,
Bhubaneswar, Odisha, India
3 Department of Computer Science and Engineering, Trident Academy of Technology,
Bhubaneswar, Odisha, India
4 Department of Computer Science and Applications, Odisha University of Agriculture and

Technology, Bhubaneswar, Odisha, India

Abstract. The most challenging difficulties in Wireless Sensor Networks (WSNs)

is energy conservation. Data fusion is the process of combining data from several
sources to produce a single scenario. It has the ability to save a large amount of
sensor energy while simultaneously enhancing sensing data accuracy. Reduced
energy usage is mainly essential and difficult examination issue in WSNs. Com-
putational intelligence (CI) principles like localization, grouping, power-sensitive
routing, job arrangement, and protection are now widely employed WSNs & other
applications. WSNs are organized into clusters to optimize data-collecting effi-
ciency while minimizing energy dissipation. Clustering allows you to organize
a deployed network into a connected hierarchy while simultaneously balancing
network load and increasing the system’s lifespan. Each sensor node in a cluster-
based WSN sends the data it has acquired to the cluster coordinator for the cluster
in which it is located. The expanse among cluster heads & events, with the energy
of clusters, is fuzzified in order to choose clusters for data uploading and fusing
using fuzzy logic. The cluster heads apply the fugitive logic technique in confined
resolution building as well, and the outcomes of the local decision-making are then
relayed to the base station. As a result of this research, a fuzzy logic-based power-
conscious active clustering method is developed, that amplifies system duration
assessed by means of Last Node Dies (LND). Results of this simulations carried
out illustrate the usefulness of this approach.

Keywords: Wireless sensor networks · Fuzzy logic · Cluster head

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 184–193, 2022.
https://doi.org/10.1007/978-3-031-23233-6_14
 Power Conscious Clustering Algorithm Using Fuzzy Logic 185

1 Introduction
Many new applications, such as armed surveillance, environmental supervising, & intel-
lectual shipping system, have been made possible as a result of recent advancements in
low-WSN, which are becoming increasingly common. As a typical WSN application,
event detection has received a great deal of attention in the past few years. Diverse wire-
less sensor networks (DWSN) are systems that are made up of different types of sensors
that differ in terms of energy consumption, computing power, and storage space, among
other characteristics. Cluster leader is dominant than cluster associate in terms of all
assets, including control, storage, communiqué, & data dispensation; this reduces the
transparency of cluster associate by allowing cluster heads to perform all of the expensive
computations, thereby reducing the overhead of cluster members. In particular, when
using a Hierarchical Wireless Sensor Networks(HWSN), for example a cluster depen-
dent system, cluster leader [1] is more influential than cluster component in terms of
every source, including control, storage, communiqué, & information processing. Thus,
load-balancing capability and network lifetime can both be significantly enhanced as
a result. It is widely used in a variety of fields, including sensor network, robotics,
video & image processing [2], and it is particularly effective in combining information
from multiple sources into a single unified picture. Data fusion has several advantages at
WSN because it is an efficient way to collaborate between several sensors. As a physical
data fusion design, clustering has become increasingly popular, as it set sensors into
numerous groups so as to accomplish the system scalability goal [3]. Clustering is a fre-
quently used physical architecture of data fusion. Each group has a cluster head (CH),
which performs and acts as a relay for data fusion. The cluster head therefore use more
power than the usual sensors; a stronger sensor is therefore more possible to chosen as
the cluster head. With the introduction of the membership concept [4], it is possible to
handle faulty data in a proper manner. Data fusion theoretical thinking framework that
adds a new notion of membership and allows for the right handling of flawed data is
known as fuzzy reasoning. Because the sensing information from a solitary sensor is
frequently indistinct & incomplete, it is complicated to derive the last fusion result from
these imperfect data by performing an exact quantitative calculation on the imperfect
data, as is the case with many other types of sensing data in general. To produce fuzzy
output, fuzzy logic makes use of the membership degree to fuzzily transform partial
data. This data is then combined with fuzzy system to create even more fuzzy produc-
tion. Using it is a simple and effective method of dealing with data that contains some
degree of uncertainty. It establish the novel concept of association degree, that allows
for the appropriate handling of imperfect data.[5] fuzzy reasoning is a supposed analysis
system for information integration that set up the narrative concept of association quan-
tity, which allows for the appropriate handling of imperfect data. Because the sensing
information from a particular antenna is frequently indistinct & incomplete, it is hard to
derive the last fusion result from these imperfect data by performing an exact quantitative
calculation on the imperfect data. To produce fuzzy output, fuzzy logic makes use of the
membership degree to fuzzily transform partial data. This data is then combined with
fluffy system to create even more fuzzy production. Using it is a simple and effective
method of dealing with data that contains some degree of uncertainty.
186 S. K. Sarangi et al.

The organized papers are as follows: Sect. 2. Related Works, Sect. 3. Proposed
Model, Sect. 4. Simulation Setup and evaluation, Sect. 5. Conclusions.

2 Related Works

Low-energy adaptive clustering hierarchy (LEACH) [6, 7] is a steering practice that uses
a grouping system wherein cluster arrangement is random, adaptive, & self configuring
in the field of routing protocols. Application-specific data transmission is controlled at
the local level. All of the sensor nodes in LEACH are grouped into clusters, with a cluster
head at the centre of each cluster (CH). The CH transfers collected information to the base
station (BS) using code division multiple access (CDMA), wherever the data is desirable,
using TDMA (time division multiple access) scheduling. When the access point is by far
the clusters and the amount of data would be sent is big, it suffers from a disadvantage. in
this case, the data transmission process consumes a lot of energy. Another disadvantage
is that, due to the adaptive nature of the clusters, the initialization stage for group a
specified round will have little impact on LEACH’s overall performance.
Centralized LEACH (LEACH-C) [8] is a centralized approach that provides an
improved cluster by using a consistent grouping stage of the CH collection. It have
stable stage, just like LEACH. For larger networks, however, they lack scalability and
robustness. Using Multi Criteria Decision-production, a strategy called Reliability based
Enhance Techniques for Ordering Preference by Ideal Similarity Solutions (RE-TOPSIS)
[9] in blend with Fuzzy supports viable and dependable CH choice (MCDM). It like-
wise utilizes the notable LEACH convention to take into consideration one time CH
choice or planning dependent on RE-TOPSIS rank file esteem in each group. During
each round of the LEACH arrangement state period, this methodology dispenses with
the requirement for CH choice. The Gupta convention [10], which depends on 3 param-
eters: centrality, focus, and energy for concentrated bunch head political race, has been
proposed to conquer LEACH’s limits. As indicated by reenactment results, the organi-
zation lifetime was altogether expanded, beating LEACH. It proposes a fluffy rationale
based bunch head political decision system (CHEF). Neighborhood distance and energy
levels were utilized by CHEF. As indicated by reenactment results, bunch heads are
appropriated more equally than LEACH. Subsequently, the organization’s future has
been broadened. The fundamental disadvantage of CHEF is that it can’t be utilized to
assemble multi-jump courses in CHs.[11] suggested a power, regression routing algo-
rithm to enhance lifetime of the network, assuming the cluster members has no resource
constraints and more power than some other sensors. It proposes a technique for model-
ing the multiple step information broadcast problem in WSNs, which is caused by more
data failure & low power competence, with the goal of providing reliable end-to-end
information broadcast at a lower price. The authors proposed a Distributed Learning
Automaton (DLA) based algorithm to conserve the difficulty as best route trouble with
numerous limitations. DLA’s ability to find the smallest number of nodes while main-
taining the required QoS specifications is exploited in the proposed solution [12] to find
the smallest number of nodes while still maintaining the required QoS specifications.
Develop and demonstrate a data gathering scheme for wireless sensor networks (WSNs)
that ensures service quality while also optimizing system concert metrics such as power
 Power Conscious Clustering Algorithm Using Fuzzy Logic 187

expenditure, operation competence, and overall reliability. EPDC adjusts the prospect
of group head appointment & evenly allocate the cluster leader diagonally the system
in order to provide connectivity between cluster members, but it is not practical for
practical WSNs due to its computational complexity. When using the Energy competent
group arrangement procedure (ECGA), cluster leader was chosen based on the amount
of data transmitted among every node and neighboring elements. [13] Has proposed an
algorithm for forming energy-efficient clusters of atoms and molecules. The projected
technique necessitates compound matrix procedures, which was never capable of being
performed by sensor nodes. When using the FLGAP procedure, the stable condition
stage is the similar as when using the LEACH. The lone time there is a conflict is during
the early stages of cluster organization. With the FLGAP (Fuzzy Logic group arrange-
ment procedure), a cluster head (CH) was selected for non-CH members using Fuzzy
Logic. The CH chance value was calculated using three parameters in order to select a
cluster head (CH). A solution to the problem of sensors with uniform distinctiveness,
which is inconvenient for a wide range of applications, has been proposed in [14]. It
is necessary for both regions to gather responsive &non-responsive data in a manner
that is separate from one another. Suitable for sensible purposes such as sensors with
mixed characteristics (i.e., where each sensor node has a different characteristic from
the others). The authors [15] have anticipated two methods, one is based on the central
(Fuzzy–C) approach which was central grouping algorithm dividing the entire network
into a predetermined quantity of spectral partitioning strategy clusters. The sink node
is supposed to have total information regarding system arrangement. The descend node
is divided the sensors into k-numbers groups and connects to all the CHs. The others
use a disseminated grouping method by using the neighboring information of a node.
A central algorithm is used as benchmarks for the evaluation of the performance of
the distributed algorithm. The centralized algorithm can improve performance with full
knowledge of network topology. A Fuzzy Logic (MCFL) [16] Multi-clustering Strategy
was developed to reduce energy dissipation and boost network life. The MCFL approach
combines sensor nodes with multiple clustering algorithms at different times, reducing
the number of messages that are sent from each node and base station to other nodes
and maintaining network energy. The authors [17] used fuzzy approach with the input
to the fuzzy system are angle, distance and energy and output is the chance of Cluster
head selection. After the cluster head selection, IDA* search algorithm is used to find
the shortest path to send the data packets from source node to sink node via cluster head
node. In [18] the authors used mamdani fuzzy inference method to select the cluster
head (CH) node. Once after the cluster head is selected, the Best first search algorithm
is used to find the shortest route path to send data to reach at the sink node.

3 Proposed Model
A. Preliminaries
The derived system model is implicated for the proposed sensor network.

1. The sensors don’t have GPS projections; therefore they were not aware of their
location (node localization algorithm used to find location of node).
188 S. K. Sarangi et al.

3 Proposed Model

Wireless Sensor Nodes

Residual Energy

(Mamdani)
Fuzzy Rules CH Selecon
FIS

Node Centrality Knowledge bases

Data transmission
Sink Node

Fig. 1. Proposed model

2. The nodes have the ability to adjust the broadcast control based on the expanse.
3. Radio relations are symmetric, which means that in any situation, two sensors X &
Y will converse by means of the equivalent broadcast authority.
4. The strength of the wireless radio signal can be used to calculate distance.
5. The most common cause of node failure is power exhaustion (Fig. 1).

B. Fuzzy Inference System

This method uses widely accepted and simplest Mamdani FIS. The inputs to the
system are residual energy and node centrality. The output of the system is fuzzy cost.
Residual Energy
The residual energy of node calculated using Eq. 1.

Ere = Ein − Eco (1)

where,
E re : Residual Energy
E in : Initial Energy of a node
E co : Consumed Energy of a node
Node Centrality
Node centrality is calculated as the sum of distance of the shortest paths between the
node and all other neighboring nodes in the networks shown in Eq. 2.
1
Nc (x) =  (2)
z d(z, x)

where, d(z, x) is the distance between the node and neighboring node.
The Mamdani FIS has the following steps:
 Power Conscious Clustering Algorithm Using Fuzzy Logic 189

1. Fuzzification:
This is the first step of Mamdani FIS. In this step convert the crisp inputs residual
energy, node centrality and output fuzzy cost into fuzzy membership grade using
triangular and trapezoidal fuzification method (Table 1) and (Fig. 2).

Table 1. Parameters and membership grade

Parameters Membership Grade

Inputs Residual Energy poor, good, excellent
Node Centrality poor, good, excellent
Output Fuzzy Cost VeryPoor(VP), MediumPoor (MP), poor, NoMedium (NM),
Medium(M), Mediumgood (MG), Good, Very Good(VG), excellent

Fig. 2. Fuzzy Membership: (a) Node centrality (b) Residual energy (c) Fuzzy cost

2. Rule generation and Rule Evaluation

Fuzzy if…then rules are generates are shown in Table 2. For an example, Rule1
can be read as: If (residual energy) is poor and (node centrality) is poor then fuzzy
cost is very poor.
190 S. K. Sarangi et al.

Table 2. Fuzzy If…Then Rules

The fuzzy if…then rules are evaluated using the Eqs. (3) and (4).
 
µXUY (z) = Max µX (z) , µY (z) (3)

 
µX ∩Y (z) = Min µX (z) , µY (z) (4)

3. Defuzzification
This is the final stage of FIS. The evaluated fuzzy rules are changes to crisps value.
The output of the defuzzification is the node cost. The centroid method is used for
defuzzification.The defuzzified value Z* using COA is shown in Eq. 5.

∗ µA (x).x.dx
Z = (5)
µA (x).dx
The output of defuzzification value is less than or equal to the pre-assigned threshold,
then the node is selected as cluster head.

Proposed fuzzy based clustering approach:

Step-1: [Initial Round]

1. Start
2. BS choose CHs arbitrarily & transmit the CH_Msg
3. Cluster arrangement & information transmission will be done in this step
4. Every sensor calculate the remaining power & node centrality and propel this
information to BS from CH
5. End

Step-2: [General Rounds]

1. Start
2. fuzzy cost ← measured by BS by means of sensor centrality & residual power
 Power Conscious Clustering Algorithm Using Fuzzy Logic 191

3. BS select CHs depending on the cost of fuzzy & transmit the message
4. Cluster arrangement & information transmission will be done in this step
5. Every sensor calculate the remaining power & node centrality and propel this
information to BS from CH
6. End

4 Simulation Setup and Evaluation

The proposed model is evaluatexd using Network Simulator NS-2. An area of (1000 *
1000) square meter has been set aside for the installation of network test beds for 400
nodes. The proposed system is assessed with respect to energy efficiency and total data
received. The simulation parameters are shown in Table 3 below:

Table 3. Simulation parameters

Sl. No Parameters Description

1 Network Area (1000x 1000) sq.m
2 Number of Nodes 400
3 Data Packet size 2500 bytes
4 Channel bandwidth 2 Mbps
5 Energy 100 J / node
6 No of Cycle Simulations 10

Figure 3 shows that, when compared to LEACH, this Fuzzy Power Conscious Clus-
tering Algorithm (FPCCA) can process about 26.8% more data if FND (First Node
Dies) is taken into account, although the BS receives roughly 15.3% and 17.9% more
data through Half of Nodes Alive (HNA) and Last Node Dies (LND).

8
Data Received

6
4 LEACH
2 FPCCA
0
FND HNA LND

Fig. 3. Data received through Base Station (BS) during FND, HNA and LND

Figure 4 shows energy consumption of nodes. It is clear from the below figure, the
energy consumption of FPCCA is less as compared to LEACH as the node increases.
192 S. K. Sarangi et al.

Consumpon(J)
Energy
1
LEACH
0 FPCCA
0 100 200 300 400
Number of Nodes

Fig. 4. Energy consumption

5 Conclusions

WSNs routing has gotten a lot of attention recently, and it presents a distinct prob-
lem when evaluated to typical information routing in wired networks. We significantly
analyzed research findings on grouping in WSN using fuzzy logic. We discussed an
energy-efficient dynamic clustering methodology in this research. The softness of the
fuzzy approach allows it to be easily modified for various network and node conditions
by simply shaping the fuzzy sets. Sensor nodes are grouped together in a cluster-based
routing system to ensure that sensed data is efficiently sent to the sink. The selection of
cluster heads is centralized here, although data collecting is spread. When compared to
LEACH, our method will extend the sensor system’s lifetime while also ensuring the
optimal quantity of groups in each round. This algorithm is straightforward and requires
little computing power. As a result, this technique can be employed effectively in larger
WSNs. More research will be done to extend this method to suit the QoS (Quality of
Service) requirements for WSNs.

References
1. Abbasi, Younis, M.: A survey on clustering algorithms for wireless sensor networks. Comput.
Commun. 30(14–15), 2826–2841 (2017)
2. Zadeh, A.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965)
3. Manjunatha, P, Verma, A.K., Srividya, A.: Multi-sensor data fusion in cluster based wireless
sensor networks using fuzzy logic method. In: Proceedings of the IEEE Region 10 and the
3rd International Conference on Industrial and Information Systems (ICIIS 2008), pp. 1–6.
IEEE, Kharagpur, December (2018)
4. Zhang, Y., Wang, J., Han, D.: Wu Hand Zhou R: Fuzzy-logic based distributed energy-efficient
clustering algorithm for wireless sensor networks. Sensors 17(7), 1554 (2017)
5. Heinzelman, W.B., Chandrakasan, A.P., Balakrishnan, H : An application-specific protocol
architecture for wireless microsensor networks. IEEE Trans. Wireless Commun. 1(4), 660–
670 (2012)
6. Wang, Q., Guo, S., Hu, J., Yang, Y.: Spectral partitioning and fuzzy C- means based clustering
algorithm for big data wireless sensor networks, EURASIP J. Wireless Commun. Netw. 1,
54 (2018)
 Power Conscious Clustering Algorithm Using Fuzzy Logic 193

7. Murugaanandam, S., Ganapathy, V.: Reliability-based cluster head selection methodology

using fuzzy logic for performance improvement in WSNs. IEEE Access, 7, 87357–87368
(2019)
8. Gupta, I., Riordan, D., Sampalli, S.: Cluster-head election using fuzzy lsogic for wireless sen-
sor networks. In: Proceedings of 3rd Annual Communication Networks Services Research
Conference, vol. 255, no. 260, 255–260, May 2005
9. Mostafaei, H.: Energy-efficient algorithm for reliable routing of wireless sensor networks.
IEEE Trans. Ind. Electron. 66(7), 5567–5575 (2019)
10. Younis, O., Fahmy, S.: HEED: a hybrid, energy-efficient, distributed clustering approach for
ad hoc sensor networks. IEEE Trans. Mobile Comput. 3(4), 366–379 (2014)
11. Aslam, N., Phillips, W., Robertson, W., Sivakumar, S.: A multi-criterion optimization tech-
nique for energy efficient cluster formation in wireless sensor networks. Inf. Fusion 12(3),
202–212 (2018)
12. El Alami, H., Najid, A: Fuzzy logic based clustering algorithm for wireless sensor networks.
Int. J. Fuzzy Syst. Appl. 6(4), 63–82 (2017)
13. Wang, Q., Guo, S., Hu, J., Yang, Y.: Spectral partitioning and fuzzy C-means based clustering
algorithm for big data wireless sensor networks. EURASIP J. Wireless Commun. Netw. 54–64
(2018)
14. Mirzaie, M., Mazinani, S.: M: MCFL: an energy efficient multi- clustering algorithm using
fuzzy logic in wireless sensor network. Wireless Netw. 24(6), 2251–2266 (2018)
15. Lenka, S., Nanda, A., Pradhan, S.K.: QoS provisioning based cost-effective routing in WSN
based IoT Network using fuzzy IDA* technique. In: IEEE 2021 8th International Conference
on Computing for Sustainable Global Development (INDIACom), pp. 591–595 (2021)
16. Lenka, S., Pradhan, S.K., Nanda, A.: Quality of service (QoS) enhancement of IoT based
wireless sensor network using fuzzy best first search approach. In: IEEE 2022 International
Conference on Intelligent Controller and Computing for Smart Power (ICICCSP), pp. 1–6
(2022)
 Cryptanalysis on “An Improved
RFID-based Authentication Protocol
for Rail Transit”

Suresh Devanapalli1(B) and Kolloju Phaneendra2

1
Department of Mathematics, Rajiv Gandhi University of Knowledge Technologies,
Basar 504107, Telangana, India
[email protected]
2
Department of Mathematics, University College of Science, Osmania University,
Hyderabad 500004, India

Abstract. In recent years, Radio-Frequency Identification (RFID) sys-

tems are widely used in many applications. RFID-based authentication
protocol plays a major role in data protection network communication.
So, authentication protocol should ensure security against all well-known
attacks such as password guessing attacks, insider attacks, and imperson-
ation attacks. In the literature, several protocols are proposed to address
the problem of authentication using RFID-tags. In this paper, we analyze
the recently proposed Zhu et al.’s protocol and show that it is vulnerable
to a known session-specific temporary information attack and imperson-
ation attack. In addition, it shows that their scheme is lack scalability.

Keywords: Security · RFID · Authentication · Key agreement

1 Introduction

In the recent days, Radio Frequency Identification (RFID) technology becomes

more popular due to a contactless automatic identification of object and also
its low cost. The advantage of RFID system is that it will simultaneously rec-
ognize massive amounts of information. Therefore, RFID-based solutions are
widely applying to many applications, includes healthcare to monitor patient’s
health. However, the most important problem with the RFID system is that
an adversary can access the tag information, which gives rise to privacy and
forgery problems. After decades of development, RFID technology has gradu-
ally become the mainstream technology for various identification applications.
At present, researchers have put forward a variety of solutions for the potential
security problems in RFID systems.
In 2005, Rhee et al. [1] proposed challenge-response based authentication
protocol based-on static identity that changes a tag response using a hash
function and the random number. In the same year, Lee et al. [2] proposed
low-cost authentication protocol that improves database operation quantity for
c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022
M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 194–203, 2022.
https://doi.org/10.1007/978-3-031-23233-6_15
 An Improved RFID-based Authentication Protocol for Rail Transit 195

resynchronization while solving an asynchronous problem between a tag and a

database. However, later Lim et al. [3] pointed out the forward security issue, in
Lee et al. [2] proposed protocol, which rise because of usage of XOR operation
to update the identity. In addition, if an attacker intentionally blocks the pre-
vious communicated message, the same value is given as a response to a query
of a reader so as not to satisfy indistinguishability and allow location tracing.
Dimitriou et al. [4] proposed a Lightweight RFID protocol to protect against
traceability and cloning attacks that can ensure forward security by updating a
tag identity using a one-way hash function. Later, Lim et al. [3] pointed out that
if we keep a current tag ID (identity) only, a database needs additional hash
operations as many as the number of saved tags every time to identify a tag.
Moreover, despite its design based on a dynamic ID, a database cannot distin-
guish a tag when the last message is blocked by an attacker, because it does not
consider control of asynchronous status and resynchronization. Further, Lim et
al. [3] proposed a dynamic ID-based mutual authentication protocol designed to
meet requirements of both indistinguishability and forward security by ensuring
the unlinkability of tag responses among sessions. Henrici et al. [5] proposed
a simple and efficient authentication protocol for low-cost RFID system. Their
protocol is based on a hash function embedding in a tag and a random num-
ber generator on a back-end server to protect the user information privacy, the
user location privacy, and the replay attack. Their scheme also provides a simple
method for the data loss. However, later Yang et al. [6] pointed out that Henrici
et al. [5] proposed protocol can not resist against the man-in-the-middle attack.
The attacker can be located between a legitimate tag and a legitimate reader and
obtain the information from the tag. Thus, the attacker easily can be authenti-
cated by the legitimate reader before the next session. Weis et al. [7] proposed
two hash-based authentication schemes: hash-lock scheme and extended hash-
lock scheme. However, Yang et al. [6] pointed out the major security drawbacks
in Weis et al. [7] proposed scheme that they are insecure against eavesdropping
attack since the attacker can track ID and impersonate the tag to a legitimate
reader. Further, Yang et al. [6] proposed an enhanced authentication protocol
for low-cost RFID, which describe privacy and security risks and how they apply
to the unique setting of low-cost RFID devices. Cho et al. [8] proposed protocol
to secure against brute-force attack: a hash-based RFID mutual authentication
protocol using a secret value which emerged with the de-synchronization prob-
lem. Tsudik [9] proposed a YA-TRAP: Yet Another Trivial RFID Authentication
Protocol, with objective to provide tracing resistance tag authentication through
monotonically increasing timestamps on the tag.
Recently, the medical field is becoming an emerging area of research, partic-
ularly RFID technology in healthcare applications. In the recent years, several
authentication protocols are presented in the literature for medical applications.
In 2007, Chien et al. [10] proposed mutual authentication protocol for RFID
conforming to Electronic Product Code(EPC) Class 1 Generation 2 standards.
But, in 2010, Yeh et al. [11] showed that Chien et al. [10] proposed protocol
is vulnerable to DoS attacks. Due to the bad properties, the claimed security
196 S. Devanapalli and K. Phaneendra

objectives are also not met. Moreover, in each time of tag access, all records
kept in the database needs to be computed and verified one by one to pinpoint
the matching record which overloads the database and pulls down the overall
performance. Then, Yeh et al. [11] proposed a securing RFID systems conform-
ing to EPC Class 1 Generation 2 standard. Unfortunately, In 2011, Habbi et
al. [12] proved that Yeh et al. [11] proposed scheme is vulnerable to tracing
attacks, obtains the most important secret value, does not provide backward
untraceability and untraceability of a tag. Further, Habbi et al. [12] proposed an
improvement on Yeh et al. [11] proposed scheme. Later, Alavi et al. [13] found
that some vulnerabilities still there in Yeh et al. [11] proposed scheme, such
as traceability and forward traceability attacks. In 2013, Khedr’s [14] proposed
SRFID: A hash-based security scheme for low cost RFID systems. Dehkordi et
al. [15] proposed an improved version of Cho et al.’s protocol [16] that eliminates
weaknesses of Cho et al.’s [16]. Later, Alavi et al. [13] showed that the protocols
proposed in [14–16] are still has some privacy concerns and are not resistance
against backward traceability and forward traceability attacks. Hoque et al. [17]
proposed enhancing privacy and security of RFID system with serverless authen-
tication and search protocols in pervasive environments. However, Deng et al.
[18] pointed out that protocol cannot offer any protection against data desyn-
chronization attack. Chen et al.’s [19] proposed a novel mutual authentication
scheme based on quadratic residues for RFID systems. However, Doss et al.
[20] showed that Chen et al.’s [19] proposed protocol is insecure against the tag
impersonation attacks, replay attacks and location privacy compromise.
In 2015, Srivastava et al. [21] proposed a hash-based RFID mutual authen-
tication protocol for TMIS, and claimed that the protocol is effective against
a various active and passive attacks such as forged attacks, replay attacks, and
so on. However, Li et al. [22] pointed out the weaknesses of Srivastava et a.l
[21] RFID tag authentication protocol such as reader stolen/lost attacks, lack
of mutual authentication between reader and server, low efficiency. Later, Zhou
et al. [23] and Benssalah et al. [24] analyzed the security flaws of Li et al. [22]
that it fails to protect against tag and reader anonymity, strong forward trace-
ability attack, replay attack, de-synchronization attack, data integrity vulner-
ability and impersonation attack. Next, Zhou et al. [23] proposed a quadratic
residue-based RFID authentication protocol with enhanced security for TMIS.
Benssalah et al. [24] proposed an enhanced authentication protocol. In 2018,
Zheng et al. [25] proposed a new mutual authentication protocol in mobile RFID
for smart campus and claimed that their protocol will provide forward security,
anti-counterfeit, anti-replay, anti-tracking, anti-eavesdropping, anti-man-in-the-
middle attack, de-synchronize, anti-DoS. But, in 2019, Safkhani et al. [26] proved
that Zheng et al. [25] protocol is vulnerable to impersonation attack, replay
attack, traceability attack, anonymity and the attacks that the adversary can
control the time.
In 2019, Safkhani et al. [26] proposed a new secure authentication protocol
for TIMS and smart campus. They claimed that their protocol is secured against
various known attacks and provides distinguished properties. But, in 2020, Feng
 An Improved RFID-based Authentication Protocol for Rail Transit 197

[27] and Zhu et al. [28] analyzed the security flaws of Safkhani et al.’s protocol
[26] and showed that it fails to provide forward secrecy. Next, Feng [27] proposed
a new authentication protocol based on quadratic residues for RFID systems.
Recently, Zhu et al. [29] proposed an improved RFID-based authentication pro-
tocol in order to withstand these security issues, and claimed that their scheme
is secure against all possible known attacks.

1.1 Motivation and Contribution

The contribution of this paper is manyfold:
– We analyze the security limitations of the recently proposed Zhu et al.’s
RFID-based authentication and key agreement scheme, and this scheme is,
unfortunately, fails to prevent known session temporary information attack.
– In addition, we have demonstrated that their protocol lack of scalability.

1.2 Organization of the Paper

The rest of the paper is organized as follows: Sect. 2 presents the preliminaries,
including the Secure Requirements and the threat model. We then review the
recently proposed Zhu et al.’s scheme [29] in Sect. 3. In Sect. 4, we show that
Zhu et al.’s scheme [29] is vulnerable to various attacks. Finally, we wind up the
paper in Sect. 5.

2 Preliminary
2.1 Secure Requirements
To secure a RFID system, the applied authentication protocol should meet the
following security demands.
1. Untraceability: A tag should not be traced, e.g., by correlating the tag’s
messages in two different sessions.
2. Resistance to replay attacks:An adversary cannot get any benefits by
replaying old messages.
3. Forward secrecy: Even if the secrets of a tag is exposed to an adversary,
the adversary can hardly identify the previous messages of the tag.
4. Mutual authentication:The protocol parties should authenticate with each
other so as to prevent any impersonation attack.
5. Synchronization:If a protocol relies on shared values for authentication, an
adversary may cause desynchronization problems. For example, if the server
updates the shared values but the tag does not, the server may not be able
to authenticate the tag in future. Such desynchronization attacks should be
resisted.
6. Scalability: If a protocol requires the server to use exhaustive search for
authentication, the protocol is of low efficiency. Worse than that, an adversary
can launch a time measurement attack [30], e.g., tracing a tag based on its
authentication time.
198 S. Devanapalli and K. Phaneendra

2.2 Threat Model

This article considers the widely accepted Dolev-Yao threat model [31] in which
an adversary has control over the communication channels between the protocol
parties, i.e., the tag-to-reader channel and the reader-to-server channel, and can
eavesdrop, modify, delete or add messages in between the communication.

Table 1. The symbols in Zhu et al.’s scheme [29]

Symbols Description
IDk , : the identity of kth tag
RIDk , : the identity of kth Reader
T : a time stamp
ΔT : the time delay
EskS (.) : encryption function with private key skS
DpkS (.) : decryption function with public key pkS
Rx : the random number generated by x
G : a cyclic group
g : a generator of G
H(.) : a secure one-way collision avoiding hash function
⊕ : the bitwise XOR operation
 : a concatenation

3 Review of Zhu et al.’s Protocol

In this section, we briefly review Zhu et al.’s scheme [29]. Their scheme has
the following phases:set up phase, authentication phase. The symbols used in
Shuming et al’s. scheme are listed in Table 1.

3.1 Set up Phase

The server chooses the cyclic group G with a generator g, and publishes them.

3.2 Authentication Phase

In this phase, we briefly review the authentication and key agreement of Zhu et
al. protocol. The summary of the authentication phase is shown in Fig. 1.

Step1: As the start, the tag generates a fresh random number Rt and computes
g Rt , then the tag sends the query and g Rt to the reader.
 An Improved RFID-based Authentication Protocol for Rail Transit 199

Fig. 1. The summary of Zhu et al.’s Protocol

Step2: When the reader receives the message from the tag, the reader stores
g Rt for later use. Then the reader generates a random number Rr , computes
g Rr and sends it with g Rt to the server.
Step3: Upon receiving g Rr and g Rt , the server generates a random number Rs
and computes g Rs . The server signs {g Rt , g Rs , g Rr } with its private key skS
and send the signed message M S to the reader.
Step4: With public key pkS, the reader can decrypt the message M S and check
whether the message is legal. If so, the reader transfers the message M S to
the tag.
Step5: When receiving the message M S from the reader, the tag check whether
the message is legal.If so, the tag computes N1 = H(IDk g Rt Rs g Rt ) and
sends it to the reader.
200 S. Devanapalli and K. Phaneendra

Step6: The reader computes N2 = H(RIDk g Rr Rs N1 g Rr ) and transfers

(N1 , N2 ) to the server.
Step7: With (N1 , N2 , g Rt , g Rs , g Rr ), the server searches its database to see
whether there are RIDk and IDk that match N1 = H(IDk g Rt Rs g Rt ) ,
N2 = H(RIDk g Rr Rs N1 g Rr ). If not, the reader or the tag will be recog-
nized as legitimate and the protocol terminates. Otherwise the server com-
putes N3 = H(RIDk g Rs Rr g Rr ), N4 = H(IDk g Rs Rt N1 g Rt ) and sends
them to the reader.
Step8: The reader computes N3 = H(RIDk g Rs Rr g Rr ) and checks whether
N3 = N3 , if so, the tag and the server achieves mutual authentication. Then
the reader transfers N4 to the tag.
Step9: The tag computes N4 = H(IDk g Rs Rt N1 g Rt ) and checks whether
N4 = N4 , if so, the reader and the server achieve mutual authentication.
Finally session key SK = g Rt Rr is defined.

4 Weakness of Zhu et al.’s Protocol

4.1 Known Session-Specific Temporary Information Attack
According to [32–36] all the session keys must be secured even if the session
random numbers of the user are compromised to an adversary A. Assume that
the session random number Rt chosen by tag is unexpectedly revealed to an
attacker A. Then, Zhu et al.’s [29] Protocol has the following drawback:
– Since tag and reader computes a session key SK as SK = g Rt Rr , an attacker
A can compute the session key SK using known session random number Rt .
– Adversary A intercepts the message {query, g Rt } sent to the reader (In step1),
and checks whether g Rt matches with Rt . If it matches, A confirms that Rt
corresponds to {query, g Rt }. The adversary A sends reply message {query,
g Rt } to reader without any modifications. In this case, neither server nor
reader can identify the message {query, g Rt } as a replied one. From the
message M S = {g Rt , g Rs , g Rr }skS , the adversary A can decrypt the message
M S with public key pkS. And also, the adversary A knows g Rr , and he/she
can compute SK as SK = g Rt Rr using g Rt . As a result, A can successfully
impersonate the legal tag.

4.2 Lack of Scalability

Zhu et al.’s protocol uses exhaustive search for authentication. In the step 7,
the server searches for a match IDk that satisfies N1 = H(IDk g Rt Rs g Rt ).
To do so, the server has to retrieve each IDk from its database, computes
H(IDk g Rt Rs g Rt ) and compares it with N1 . Similarly, the server has to retrieve
each RIDk , computes N2 = H(RIDk g Rr Rs N1 g Rr ) and compares it with N2 .
Assume that there are m readers and n tags in the system. The server has to
perform m+n+22 hash operations on average to find the match IDk and RIDk .
It is obvious that the computation cost is linear with the number of readers
 An Improved RFID-based Authentication Protocol for Rail Transit 201

and tags in the system so this protocol is lack of scalability. In addition, as

mentioned in Sect. 2.1, because the server needs to perform exhaustive search
operation to authenticate the reader and the tag, this protocol also suffers from
time measurement attacks.

5 Conclusion
In this paper, we have first reviewed the recently proposed Zhu et al. protocol
and then shown that their protocol fails to prevent known session temporary
information attack. Also, we have demonstrated that their protocol lack of scal-
ability. In the future, we aim to design a novel and more secure RFID-based
authentication protocol using Elliptic curve cryptography for Telecare Medicine
Information System to withstand the security flaws found in Zhu et al. Scheme.

References
1. Rhee, K., Kwak, J., Kim, S., Won, D.: Challenge-response based RFID authenti-
cation protocol for distributed database environment. In: Hutter, D., Ullmann, M.
(eds.) SPC 2005. LNCS, vol. 3450, pp. 70–84. Springer, Heidelberg (2005). https://
doi.org/10.1007/978-3-540-32004-3 9
2. Lee, S.M., Hwang, Y.J., Lee, D.H., Lim, J.I.: Efficient authentication for low-cost
RFID systems. In: Gervasi, O., et al. (eds.) ICCSA 2005. LNCS, vol. 3480, pp.
619–627. Springer, Heidelberg (2005). https://doi.org/10.1007/11424758 65
3. Lim, J., Oh, H., Kim, S.: A new hash-based RFID mutual authentication protocol
providing enhanced user privacy protection. In: Chen, L., Mu, Y., Susilo, W. (eds.)
ISPEC 2008. LNCS, vol. 4991, pp. 278–289. Springer, Heidelberg (2008). https://
doi.org/10.1007/978-3-540-79104-1 20
4. Dimitriou, T.: A lightweight RFID protocol to protect against traceability and
cloning attacks. In: First International Conference on Security and Privacy for
Emerging Areas in Communications Networks (SECURECOMM2005), pp. 59–66.
IEEE (2005)
5. Henrici, D., Muller, P.: Hash-based enhancement of location privacy for radio-
frequency identification devices using varying identifiers. In: IEEE Annual Confer-
ence on Pervasive Computing and Communications Workshops, 2004. Proceedings
of the Second, pp. 149–153. IEEE (2004)
6. Yang, J., et al.: Mutual authentication protocol for low-cost RFID. In: Workshop
on RFID and Lightweight Crypto, pp. 17–24. WRLC (2005)
7. Weis, S.A., Sarma, S.E., Rivest, R.L., Engels, D.W.: Security and privacy aspects of
low-cost radio frequency identification systems. In: Hutter, D., Müller, G., Stephan,
W., Ullmann, M. (eds.) Security in Pervasive Computing. LNCS, vol. 2802, pp.
201–212. Springer, Heidelberg (2004). https://doi.org/10.1007/978-3-540-39881-
3 18
8. Cho, J.S., Yeo, S.S., Kim, S.K.: Securing against brute-force attack: a hash-based
RFID mutual authentication protocol using a secret value. Comput. Commun.
34(3), 391–397 (2011)
9. Tsudik, G.: Ya-trap: yet another trivial RFID authentication protocol. In: Fourth
Annual IEEE International Conference on Pervasive Computing and Communica-
tions Workshops (PERCOMW2006), p. 4. IEEE (2006)
202 S. Devanapalli and K. Phaneendra

10. Chien, H.-Y., Chen, C.-H.: Mutual authentication protocol for RFID conforming
to EPC class 1 generation 2 standards. Comput. Stand. Interfaces 29(2), 254–259
(2007)
11. Yeh, T.-C., Wang, Y.-J., Kuo, T.-C., Wang, S.-S.: Securing RFID systems conform-
ing to EPC class 1 generation 2 standard. Expert Syst. Appl. 37(12), 7678–7683
(2010)
12. Habibi M.H., Gardeshi, M.: Cryptanalysis and improvement on a new RFID mutual
authentication protocol compatible with EPC standard. In: 2011 8th International
ISC Conference on Information Security and Cryptology, pp. 49–54. IEEE (2011)
13. Alavi, S.M., Baghery, K., Abdolmaleki, B., Aref, M.R.: Traceability analysis of
recent RFID authentication protocols. Wireless Pers. Commun. 83(3), 1663–1682
(2015)
14. Khedr, W.I.: SRFID: a hash-based security scheme for low cost RFID systems.
Egyptian Inf. J. 14(1), 89–98 (2013)
15. Masoud Hadian Dehkordi and Yousof Farzaneh: Improvement of the hash-based
RFID mutual authentication protocol. Wireless Pers. Commun. 75(1), 219–232
(2014)
16. Cho, J.-S., Jeong, Y.-S., Park, S.O.: Consideration on the brute-force attack cost
and retrieval cost: a hash-based radio-frequency identification (RFID) tag mutual
authentication protocol. Comput. Math. Appl. 69(1), 58–65 (2015)
17. Hoque, M.E., Rahman, F., Ahamed, S.I., Park, J.H.: Enhancing privacy and secu-
rity of RFID system with serverless authentication and search protocols in perva-
sive environments. Wireless Pers. Commun. 55(1), 65–79 (2010)
18. Deng, M., Yang, W., Zhu, W.: Weakness in a serverless authentication protocol
for radio frequency identification. In: Wang, W. (ed.) Mechatronics and Automatic
Control Systems. LNEE, vol. 237, pp. 1055–1061. Springer, Cham (2014). https://
doi.org/10.1007/978-3-319-01273-5 119
19. Chen, Y., Chou, J.-S., Sun, H.-M.: A novel mutual authentication scheme based
on quadratic residues for RFID systems. Comput. Netw. 52(12), 2373–2380 (2008)
20. Doss, R., Sundaresan, S., Zhou, W.: A practical quadratic residues based scheme
for authentication and privacy in mobile RFID systems. Ad Hoc Netw. 11(1),
383–396 (2013)
21. Srivastava, K., Awasthi, A.K., Kaul, S.D., Mittal, R.C.: A hash based mutual RFID
tag authentication protocol in telecare medicine information system. J. Med. Syst.
39(1), 153 (2015)
22. Li, C.-T., Weng, C.-Y., Lee, C.-C.: A secure RFID tag authentication protocol with
privacy preserving in telecare medicine information system. J. Med. Syst. 39(8),
77 (2015)
23. Zhou, Z., Wang, P., Li, Z.: A quadratic residue-based RFID authentication protocol
with enhanced security for TMIS. J. Ambient. Intell. Humaniz. Comput. 10(9),
3603–3615 (2019)
24. Benssalah, M., Djeddou, M., Drouiche, K.: Security analysis and enhancement of
the most recent RFID authentication protocol for telecare medicine information
system. Wireless Pers. Commun. 96(4), 6221–6238 (2017)
25. Zheng, L., et al.: A new mutual authentication protocol in mobile RFID for smart
campus. IEEE Access 6, 60996–61005 (2018)
26. Safkhani, M., Vasilakos, A.: A new secure authentication protocol for telecare
medicine information system and smart campus. IEEE Access 7, 23514–23526
(2019)
27. Zhu, F.: SecMAP: a secure RFID mutual authentication protocol for healthcare
systems. IEEE Access 8, 192192–192205 (2020)
 An Improved RFID-based Authentication Protocol for Rail Transit 203

28. Zhu, F., Li, P., He, X., Wang, R.: A novel lightweight authentication scheme for
RFID-based healthcare systems. Sensors 20(17), 4846 (2020)
29. Zhu, R., He, X., Xie, J., Zhang, Z., Wang, P.: An improved RFID-based authen-
tication protocol for rail transit. In: 2020 IEEE 14th International Conference on
Big Data Science and Engineering (BigDataSE), pp. 65–72. IEEE (2020)
30. Avoine, G., Coisel, I., Martin, T.: Time measurement threatens privacy-friendly
RFID authentication protocols. In: Ors Yalcin, S.B. (ed.) RFIDSec 2010. LNCS,
vol. 6370, pp. 138–157. Springer, Heidelberg (2010). https://doi.org/10.1007/978-
3-642-16822-2 13
31. Dolev, D., Yao, A.: On the security of public key protocols. IEEE Trans. Inf. Theory
29(2), 198–208 (1983)
32. SK Hafizul Islam: Design and analysis of an improved smartcard-based remote user
password authentication scheme. Int. J. Commun Syst 29(11), 1708–1719 (2016)
33. He, D., Kumar, N., Khan, M.K., Lee, J.-H.: Anonymous two-factor authentication
for consumer roaming service in global mobility networks. IEEE Trans. Consum.
Electron. 59(4), 811–817 (2013)
34. Canetti, R., Krawczyk, H.: Analysis of key-exchange protocols and their use for
building secure channels. In: Pfitzmann, B. (ed.) EUROCRYPT 2001. LNCS, vol.
2045, pp. 453–474. Springer, Heidelberg (2001). https://doi.org/10.1007/3-540-
44987-6 28
35. Cheng, Z., Nistazakis, M., Comley, R., Vasiu, L.: On the indistinguishability-based
security model of key agreement protocols-simple cases. Cryptology ePrint Arch.
2005, 129 (2005)
36. Mishra, D., Das, A.K., Mukhopadhyay, S.: A secure user anonymity-preserving
biometric-based multi-server authenticated key agreement scheme using smart
cards. Expert Syst. Appl. 41(18), 8129–8143 (2014)
 A Novel Approach to Detect Rank Attack in IoT
Ecosystem

Amardeep Das1,2(B) , Nibedita Adhikari1 , Pradeep Kumar Bhale3 , and S. K. Rath4

1 Department of Computer Science and Applications, Utkal University, Bhubaneswar 751004,
India
[email protected], [email protected]
2 Department of Computer Science and Engineering, C.V. Raman Global University,
Bhubaneswar 752054, India
3 Department of Computer Science and Engineering, Indian Institute of Technology,
Guwahati 781039, India
[email protected]
4 Department Admin/Computer Science, BPUT, Rourkela 769015, India

Abstract. Internet of Things (IoT) is a popular technology in which everyday

objects are converted into intelligent gadgets. These gadgets have sensing, com-
putation, and networking capabilities. These gadgets, equipped with a variety of
IoT applications, are primarily focused on automating various functions and are
attempting to give inanimate physical things the ability to behave independently of
any human interaction. Existing and future IoT applications are extremely promis-
ing for enhancing consumer convenience, productivity, and automation. To build
such an IoT ecosystem in an ever-expanding manner, we require high levels of
security, privacy, and authentication. The IoT ecosystem and associated routing
protocols are composed of several threads. In an IoT network, a lightweight Rank
attack detection system is suggested by this research. With the aid of the Contiki
Cooja simulator, we implement our proposed lightweight security solution. Exper-
iments using simulation demonstrate that the proposed method is lightweight and
can detect the Rank attack with comparable performance.

Keywords: Internet of Things · Rank attack · Intrusion Detection System

(IDS) · Contiki Cooja simulator

1 Introduction
Kevin Ashton, a British technological innovator, is credited with inventing the Internet
of Things (IoT). The IoT is a huge network of machines/objects that are all connected
to each other. The connected machine/objects may be low-resource devices that share
and communicate to one another via wired or wireless connections [1]. A different study
says that IoT ecosystem evolution market share and prospects have grown by 31% since
2016 and that 8.4 billion IoT-enabled objects will be in use in 2017 [2, 3]. According
to the information gathered from the resources, the ecosystem of the Internet of Things
will expand by the year 2020, when there will be 30 billion gadgets. This demonstrates

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 204–221, 2022.
https://doi.org/10.1007/978-3-031-23233-6_16
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 205

both how quickly we are integrating them into our daily lives and why it is so important
to treat their security seriously.
The vast majority of Internet of Things ecosystem components and applications are
dependent on Internet Protocol version 6, which was released to support these technolo-
gies (IPv6). Because the sensor nodes have limited resources, the routing protocol that
is employed must be as efficient as possible in terms of both the amount of energy it
consumes and the amount of computational power it can muster. The routing protocol for
the IoT ecosystem is also known as Low-power and Lossy Networks (RPL) [4]. It was
developed to address routing issues in the IoT ecosystem. RPL has been designated by
the Internet Engineering Task Force (IETF) as the preferred routing protocol for the IoT
ecosystem [5]. Thus, RPL is widely employed in several IoT applications. This protocol
was designed from the ground up to work well over slow and error-prone connections,
and it succeeds admirably. Among RPL’s multiple impressive characteristics are its
adaptability to different routing metrics and goal functions; its ability to handle compli-
cated interconnections and multi-topology routing operations; and the adaptability of its
trickling method and control message style and frequency.
RPL contains a number of configurable security measures to safeguard its control
messages. These approaches ensure authenticity, integrity, and confidentiality. Assailants
are nevertheless able to get control of the genuine nodes despite the fact that they are not
immune to being tampered with and are not physically protected [6]. Threats that lower
the quality of IoT applications service can be launched using these compromised nodes.
RPL is utilized in numerous IoT applications and can be attacked in different ways
[7, 8]. Cyber-attacks like Rank, black-hole, Sybil, and Sinkhole attacks are able to target
the RPL mostly because of design defects in the topology generation of the RPL [9–12]
All of the above assaults are the most significant obstacles to the actual use of RPL in
the IoT ecosystem. Despite the protection provided by the MAC layer, RPL continues to
struggle significantly with the issue of internal assaults. Destructive attacks like the rank
attack occur when an attacker node successfully creates a fake topology and coerces its
neighbors into rerouting traffic to itself. Despite the fact that the rank attack has been the
subject of a noteworthy amount of analysis in the routing path, no study has investigated
the consequences of assaults of this kind on the different topologies of the IoT network.
Additionally, no study covers how this assault on the RPL networks is made viable. This
gave us the motivation to carry out this study and close the noted research gap. The
highlights and new contributions of our study are outlined here.
We examine both the non-attack and the attack scenario in relation to rank estima-
tion using the objective functions (OFs) and the DODAG Information Object (DIO)
communication methods.
We analyze rank attack change with threshold modifications introduced by the Rank
change scenario with minimum computational and communication overhead.
The suggested method offers novel lightweight security solutions to improve Rank
attack detection in IoT ecosystems. It gives High True Negative Rate (TNR) and True
Positive Rate (TPR) with minimum energy consumption.
The suggested solution is implemented on both small- and large-scale IoT networks
to examine the effectiveness and efficiency of the desired solution.
206 A. Das et al.

The comprehensive analysis of the suggested security solution that was carried out
using the Contiki cooja simulation.
The remainder of the paper is structured as follows. Section 2 introduces the generic
IoT network architecture, RPL as IoT routing protocol, Rank attack in IoT ecosystem, and
Intrusion Detection System (IDS) for IoT ecosystem. The relevant research is presented
and discussed in Sect. 3. In Sect. 4, an explicit overview of the proposed security model
is presented. The experiments and results are described in Sect. 5. The research paper is
finally ended in Sect. 6, which also discusses conceivable future directions.

2 Background

In this section, we offer a concise introduction to the architecture of the IoT net- work,
RPL as routing protocol, rank attack, and IDS for the IoT ecosystem.

2.1 Generic IoT Network Architecture

There are five parts to the IoT network architecture: Business, Application, Process-
ing, IoT-Internet Connection, IoT Access Network, and Perception. The network stack
remains the same, as illustrated in Fig. 1, with the exception of the upsurge of the adap-
tion layer, modification of the routing strategy at the IoT access network layer, use of
IEEE 802.15.4 as the MAC layer, and other similar changes.
The layers and functions associated with each layer are depicted in Fig. 1. With
the exception of the extra adaption layer, the IoT access network layer is functionally
equivalent to the network layer in conventional networks. The Application layer of
the Internet of Things commonly implements a protocol known as CoAP (Confined
Application Protocol). This protocol was designed particularly for limited devices such
as LLN devices and is used by the majority of IoT applications. In the part on the RPL
algorithm that came before this one, [13, 14] you may find a description of the routing
protocol that is used.

2.2 RPL Protocol

IoT ecosystems that are built on Low Power and Lossy Networks (LLN) will use this
routing protocol. As a practical protocol, RPL supports a variety of communication
modes, including [15]. Reactive and proactive RPL are the two classes of RPL that
could be found in the IoT ecosystem. RPL provides path as needed. The disadvantage is
that the time needed to find a path to an IoT node increases as more and more connections
are made. The routes are provided by proactive RPL before another IoT node needs them.
In addition to this, it engages in the trading of various control messages in order to get
access to local knowledge about the area and find new routes [16]. RPL procedure as
displayed in Fig. 2.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 207

Fig. 1. Generic IoT network architecture

2.3 Rank Attack in IoT

This section’s main goal is to talk about the two most important routing-based attacks.
As shown in Fig. 3, attacks can be put into three groups based on the IoT ecosystem
resources, IoT ecosystem topology, and IoT ecosystem traffic. This research paper used
an IoT ecosystem scenario to test and look at the rank attack.

2.4 IDS for IoT Ecosystem

It consists of cooperating elements that detect attacks and harmful behavior in the IoT
ecosystem. IDSs are as follows:

– Payload-based IDS (PIDS): It analyses inbound/outbound payload information.

Packets are statistically tested.
– Header-based IDS (HEIDS): It analyses only the packet header portion of network
packets. Comparing the other IDS, its tests and uses the perceived network packet
payload.
– Statistical anomaly-based IDS (SAIDS): It collects the IoT ecosystem information
and learns what normal user behavior is. Statistical tests are used to figure out if a
behaviour is right or wrong based on how it is seen. SAIDS can be categorized into
two main groups like profile-based anomaly detection and threshold-based anomaly
detection.
208 A. Das et al.

Internet
Downward Direction Rank Direction

Rank Decreases Upward Direction

1 1
RPL Messages: DAO, DIS, CC

RPL Messages: DAO, DIS, CC

2
2 2 2 2

3 3 3
3 3
3 3 3

4
4 4

Fig. 2. Execution of RPL routing protocol in IoT ecosystem.

RPL Attack

Traffic Related Attack Topology Related Attack Resources Related Attack

Identity Attack Black-hole Attack Version Number Attack

Traffic Analisys Sink-hole Attack Floading Attack
Sniffing Wormhole Attack DAG Inconsistent
Decreased Rank Attack DAO Inconsistency
DIS Attack

Fig. 3. Various types of attacks in RPL routing protocol in the IoT ecosystem

– Signature-based IDS (SIDS): This type of IDS uses a database of intrusion detection
signatures that have already been seen. Every intrusion leaves a trail, such as the files
and folders that were accessed, the type of data packets that were sent, and failed
logins, etc. This IDS compares the behavior of known attack imprints in the IoT
ecosystem. If attack imprints or signature match is found, a notification is given so
that the right steps can be taken.
– Network-based IDS (NIDS): Analyses network traffic and monitors many hosts to
identify malicious activities.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 209

– Host-based IDS (HIDS): In order to identify potentially malicious behaviour, a HIDS

will monitor file-system changes, system calls, and application logs (such as binaries,
password files, and so on), as well as host-based actions and states.

3 Related Work

Kamble et al. [17] enclose recently provided a survey on various protection methods
against vulnerabilities and safe routing algorithms in RPL-based Internet of Things
technology. They divided the attacks into two basic categories (i.e., direct and indirect
assault). In the case of an explicit invasion, the assailants are actively involved in the
attack. Examples of direct assaults in the IoT Ecosystem include routing table overflow
assaults and flooding attacks. When an invasion occurs, it is not the intruder who is
attacking directly. Therefore, the malicious node relies on other nodes to overwhelm
IoT networks with assaults such as increasing rank attacks, version number attacks, and
inconsistency in DAG formation attacks.
Raoof et al. [16] include different forms of routing-related attacks in the IoT ecosys-
tem: Sybil attacks, Selective forward attacks, Black-hole attacks, etc. It is important
to note that several metrics are used for evaluating various assaults (e.g., data packet
delivery rate, data packet delay, power consumption, etc.). This research study lessens
the severity of the effects of a routing assault without adding any additional defence
measures to the routing protocol. But the low TPR and TNR are a weakness of this
paper.
Le et al. [18] recognized attacks including rank, sinkhole, DIS and local repair using
IDS. The Extended Finite State Machine (EFSM) in this IDS was produced using a
semi-automatic characterization approach. The DIO suppression attack, a brand-new
method of attacking the RPL protocol, was first described by Perazzo et al. [19]. This
attack targets the ability of IoT nodes to send and receive updated DIO control messages,
which are necessary for the discovery of optimal routing pathways and the elimination
of inefficient or erroneous paths.
Wallgren et al. [20] utilized the Cooja simulation in order to carry out a variety
of attacks on RPL. These attacks included a selective forwarding attack as well as a
sinkhole attack. They have also exhibited many capabilities of the IPv6 protocol and its
application in the IoT enable IDS. This method’s primary drawback is that it consumes
more power than local processing.
Glissa et al. [21] presented SRPL, a safe routing mechanism utilized by RPL rout-
ing protocol. This approach prevents deviant nodes from generating a false topology by
altering node rankings. They have implemented a rank threshold and hash chain authen-
tication approach to thwart internal attacks likes selective forwarding attacks, blackhole
attacks, and sinkhole attacks. This research paper concludes that internal attacks depend-
ing on malicious RPL metrics must be avoided. The SRPL strategy, meanwhile, has
several shortcomings. First and foremost, regulated messages produced by SRPL incur
more overhead in contrast to RPL. Additionally, it requires greater processing power
and storage space.
Osman et el. [22], an ML-based model was introduced that uses the Light Gradient
Boosting Machine (LGBM) to find VN attacks. Methodology components focused on
210 A. Das et al.

creating a dataset of VN attacks, extracting features from that dataset, using a logistic
regression-based classification technique, and optimizing the algorithm’s specifications
were presented. The suggested model performed well but required at least 347,530
bytes of memory, which is beyond the capability of most IoT devices. This suggests that
in order to execute the model in a centralized fashion, this system requires additional
devices with significant memory resource capacities.
In order to protect the RPL routing mechanism against rank and VN at- tacks, the
authors of [23] developed a blockchain-based structure. ML A-based detection systems
modules in RPL network are connected securely through the blockchain network. To
identify potential rank and VN attacks, it deploys an XGBoost classifier on a private
blockchain network. After that, smart contracts are utilized to evaluate these efforts and
provide real-time warnings to harmful nodes in the IoT networks. The outcomes of
the performance tests demonstrated that the blockchain-based solution that was offered
boosted the accuracy reached by the ML algorithms when it came to forecasting the
assault.
Loo et al. [24] demonstrated a new type of routing attack known as a Topology
Attack, which alters node operation by separating the optimum topology of the nodes.
Local Repair Attack and Rank Attack are two subtypes of Rank at- tack. They have
furthermore implemented an IDS architecture that makes use of RPL finite state machines
in order to keep an eye on the nodes. Nevertheless, this strategy has the drawback of
using more memory and consuming more power than necessary. IoT network and RPL
attacks have been studied by et al. [25]. They have demonstrated that several assaults are
practicable on low-resource devices due to the simplicity and small size of IoT protocols.
In [26], the authors provide a lightweight solution which they call INTI (Intrusion
detection of SiNkhole attacks on 6LoWPAN for Internet of Things). This strategy, which
combines trust, reputation, and sinkhole attack detection using Rooming IDS in the
IoT ecosystem with a watchdog, is a sandwich of many approaches. When developing
the solution for the Internet of Things, they took the nodes’ ability to move around
in the network into account. FNR, FPR, and resource utilisation are all regarded as
performance criteria by the INTI. The biggest drawback of this technique is that it
requires more resources and computational power. As a result, this method shortens the
IoT ecosystem’s lifespan.
In [27] describes how the researchers developed four assault models and evaluated
how they affected important network metrics relating to the attacker’s topological loca-
tion. However, this study solely addressed static networks and did not offer any feasible
defenses against the rank assault. The effects of a ranked assault utilizing a fictitious IP
address were examined by the authors of [28]. However, the specific consequences of
the rank attack on the RPL procedure are not specifically examined in this study. The
enhanced rank assault was initially presented by Shukla et al. [29], who also proposed
an ego-based mitigation strategy. However, this study focuses on a topology with just
12 nodes rather than taking into account a realistic topology. Additionally, mobile nodes
were not included in this study.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 211

4 Proposed Security Approach

This section describes the proposed rank attack security solution for detecting rank
attacks in the IoT network. It consists of many IoT devices with various processing
and storage capacities. Our intended security approach employs the lightweight IDS.
Additionally, it utilizes minimum energy during the execution of the proposed technique
to detect rank and rank-related attacks.

4.1 Proposed Approach Assumption

We would want to consider the following assumptions before we discuss the lightweight
security mechanism and the methodologies.
We presume that every device with a resource limitation has a finite quantity of
available power.

– In the IoT ecosystem, every single device is different in terms of its resource limitations
(i.e. communication capacity, computation, memory, and battery power).
– Despite this, it is possible that billions or even trillions of IoT devices are connected
to the IoT ecosystem via wireless network. Nevertheless, we will provide our strategy
for a IoT ecosystem that is capable of holding a fixed quantity of the IoT devices like
8 to 128 IoT devices.

4.2 Security Model

The IoT network is made up of both genuine and malicious nodes (MN). The border
router (6BR) serves as the point of entry for the IoT. In addition, 6BR acts as a connection
point among the private and the public Internet network. The 6BR is a powerful hub
that may take the form of a portable computer. There is currently no computer or laptop
hardware that can be considered an equivalent to IEEE 802.15.4 devices. For this reason,
we run the Linux platform based communicate with the Cooja simulator [30]. The IoT
setup configuration for our investigation is shown in Fig. 4.

4.3 Proposed Rank Attack Detection Solution

The suggested lightweight security solution based on rank change threshold. We endorse
both static rank attack detection IDS module and lightweight nature is deploy them
through 6BR routers. Based on the features of IoT network characteristics like battery
power, resource-constrained, and lossy. Fig. 5 utilized to identify rank assault. Based on
residual energy, we calculate the lifetime of the network.

5 Experiments and Results Analysis

5.1 Setup and Execution of Experiments
Our proposed security solution is implemented in Contiki cooja [30] emulation settings
and presumes that the IoT ecosystem comprises various type of sensor like Tmotes Sky,
212 A. Das et al.

Fig. 4. IoT experiment setup and rank attack detection IDS module architecture

and sky mote which is CC2420 IEEE 802.15.4-compatible radio chips [31]. 6BR device
features an additional piece of hardware and increased processing power. To support the
intended solution to identifying Rank attacks. The IoT network is depicted in Fig. 4.
We conduct three different types of experiments in the experiment section. A thorough
justification of each experiment is provided below. Table 1 lists the simulation settings
with values for Contiki cooja.

Experiments 1: Non-rank Attack in IoT Ecosystem: IN this experiment, we take

into account various node counts (i.e., 4 to 128 node). The settings utilised to simu-
late the environment are listed in Table 1 and the simulation’s network architecture is
illustrated in Fig. 4. Fig. 6(a), 6(b), and 6(c), respectively, depict the topology, DODAG,
and power consumption profile. When there are no malicious nodes that are lowering
the rank and causing the network to function normally, we noticed that the nodes’ power
usage is lower.
At several nodes, the IoT network traffic is examined utilizing collect view tech-
niques. Fig. 7 displays the throughput achieved in this experiment. The figure depicts
that as the nodes number rises, the average throughput gradually declines. Additionally,
we can see the avg. energy utilization per node and the network’s overall energy utiliza-
tion in Figs. 8 and 9. RPL alone and RPL collect modules are used in the IoT paradigm
for power and energy analysis.

Experiments 2: Rank Attack in IoT Ecosystem: IN a situation involving a rank

attack, the same amount of nodes are used to perform an attack. At different nodes,
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 213

Fig. 5. Flowchart of rank attack detection process

214 A. Das et al.

Table 1. Cooja simulator settings with values [3]

Name of setting Value

OS Contiki NG, Contiki 3.0 OS
Simulator Cooja
Target area 100 m × 100 m
Radio environment UDGM with distance loss
Sensor node type Tmot Sky and sky mote
IoT routing protocol Contiki RPL
Adaptation layer 6LoWPAN
Transmitter power output (TPO) (dBm) 0–25
Receiver sensitivity threshold (dBm) - 94
Radio frequency (RF) IEEE 802.15.4 CC2420 2.4 GHz
Type of attack Rank attack
Simulation duration 60 min

collect view methods are used to gather and analyse IoT traffic in real time. We noticed
a significant decrease in throughput together with an increase in the typical amount of
power consumed by each node of 41–54%. Specifically, as seen in Fig. 10 and Fig. 11,
respectively. Figure 12 emphasizes a 37.8% increase in energy use for the whole network.

Experiment 3: Rank Attack with Security Solution: To identify a rank assault, we

execute a security solution based on the rank change threshold. Experiment 1 and Exper-
iment 2 are repeated numerous times along a varying number of connected IoT devices
in IoT networks and power usage overview with the help of Contiki Cooja Simulator’s
Collect-View application.

5.2 After Proposed Security Solution Implementation Performance Analysis

We provide a security solution to the widely-used Contiki OS [9] in order to safeguard

the Io network. The following are the measures that will be used to evaluate performance:

– Average energy usage: All IoT enabling device in the IoT ecosystem are battery-
powered device. As a result, there is a finite amount of energy. Here we compute
power depletion at node-level as well as system-level. For this purpose, we employ
powertrace on sky motes to estimate the power depletion. The powertrace output
shows system execution time with respective total time.

– Memory usage (RAM/ROM): It shows the percentage of memory utilization of the

IoT devices to run the proposed security solution throughout the experimentation.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 215

Fig. 6. Topology, DODAG, and power profiling are utilized to study non-rank assaults.

– TNR and TPR: The True Negative Rate (TNR) and True Positive Rate (TPR) are
estimated utilising sensitivity and specificity correspondingly. Sensitivity predicts
the number of legitimate nodes (LN) that are quickly and precisely recognised,
whereas specificity counts the quantity of rank attack generated nodes like (CN) that
are consistently discovered. Both sensitivity and specificity fit within the following
216 A. Das et al.

Fig. 7. Average throughput different IoT node with different run time

Fig. 8. Average power consumption per node

descriptions:
m
Senstivity = (1)
m+n
and
x
Specificity = (1)
x+y
where m = CN correctly recognise, n = Falsely labelled as authentic by CN, x = LN
correctly recognise. y = LN labelled wrongly as malicious.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 217

Fig. 9. Utilization of energy over the whole network in 60 min.

Fig. 10. Avg. Throughput different IoT node with different run time in rank attack

5.3 Comparison of the Suggested Security Solution to Similar Works

In this section, a comparison is made based on memory usage, TNR, TPR, Avg. power
usage, and IoT traffic. As depicted in Table 2, the comparative analysis of the suggested
technique and the most recent current solutions. The suggested technique uses additional
ROM and RAM for the memory utilization analysis. Every configuration has a different
baseline since it is based on a different Contiki system component. A 6BR router, for
instance, requires more memory usage than other IoT nodes. Nevertheless, 1760 and
365 bytes of total ROM and RAM are anticipated to be used to implement the suggested
218 A. Das et al.

Fig. 11. Avg. Power consumption per node in rank attack

Fig. 12. Energy usage for the entire network in rank attack

strategy on IoT nodes. In experiment 3, 64 IoT nodes were run for 60 min while 9834
and 54256 bytes of RAM/ROM consumption were observed, respectively.
The avg TNR and TPR attained using the proposed approach are 94.12% and 95.34%
correspondingly. Most comparable solutions demonstrated in the literature have lower
detection rates than the rank attack detection rate [23, 28, 29]. However, when the IoT
network grows (in terms of size and device count), the TPR declines as a result of the
massive network topologies. Additionally, it takes some time for the network to stabilize
and grow enough to get a more increased TNR and TPR. In experimentation work, the
rank attack case, it takes 128 nodes 60 min to obtain the same TPR and TNR as 64 nodes
do in 30 min.
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 219

The average energy consumption of the whole IoT ecosystem for 60 min, during
which our experiment was conducted with a certain number of devices with limited
resources. Energy usage increases as the number of IoT nodes in- creases from 8 to 128
nodes. Our proposed approach reduces energy consumption by up to 70000 mJ.

Table 2. Comparing proposed method to current works

Research paper Simulation Memory usage Avg. energy Rank attack

(in byte) utilization (mJ) detection (%)
ROM RAM TPR TNR
[17] Contiki Cooja 47137 15883 75522 78.98% 93.11%
[26] Contiki Cooja 51045 12894 76320 92.45% 93.93%
[28] Contiki Cooja 74914 14587 74567 93.34% 93.05%
[23] Contiki Cooja 70534 12544 71801 95.86% 94.31%
[22] Contiki Cooja 60476 11837 70480 96.11% 95.81%
[29] Contiki Cooja 56137 10383 62133 93.68% 92.22%
Proposed method Contiki Cooja 54256 9834 48391 95.34% 94.12%

6 Conclusion

The goal of the research is to provide rank change threshold-based security solutions
that monitor IoT network traffic and offer security against rank attacks. The execution of
the suggested security solution is assessed in contiki cooja with three types of network
scenarios, like non-rank attack, rank attack, and security solution with rank attacks. We
also included a comparison with recently developed solutions and closely related work.
Results from experiments demonstrate that the suggested system is capable of detecting
rank attacks. As a result, rank attacks are depreciated since the proposed security solution
uses the rank change threshold-based security solution to identify the rank attack. The
performance of accuracy and reaction time was superior to other security solutions. In
comparison to the closely linked work, the findings reveal that the average energy usage,
memory usage (RAM/ROM), TPR, and TNR are comparable 47540 mJ, 9834/54256 in
Byte, 95.34% and 94.12% respectively. In the future, the effort will incorporate the IoT
network’s lightweight security solution and evaluate the enhanced defence mechanism
by fending off multiple mixed attacks.

Acknowledgments. The network security and systems lab, IIT Guwahati, India; Utkal University,
Bhubaneswar, India; and BPUT, Bhubaneswar, India have all been the sites of the study activity.
IIT Guwahati, Utkal University, Bhubaneswar, and BPUT, Bhubaneswar are all thanked by the
authors for their assistance.
220 A. Das et al.

References
1. Ashton, K., et al.: That ‘internet of things’ thing. RFID J. 22(7), 97–114 (2009)
2. Sharma, N., Shamkuwar, M., Singh, I.: The history, present and future with IoT. In: Balas,
V.E., Solanki, V.K., Kumar, R., Khari, M. (eds.) Internet of Things and Big Data Analytics
for Smart Generation. ISRL, vol. 154, pp. 27–51. Springer, Cham (2019). https://doi.org/10.
1007/978-3-030-04203-5_3
3. Bhale, P., Dey, S., Biswas, S., Nandi, S.: Energy efficient approach to detect sinkhole attack
using roving ids in 6lowpan network. In: Rautaray, S.S., Eichler, G., Erfurth, C., Fahrnberger,
G. (eds.) I4CS 2020. CCIS, vol. 1139, pp. 187–207. Springer, Cham (2020). https://doi.org/
10.1007/978-3-030-37484-6_11
4. Kharrufa, H., AlKashoash, H.A., Kemp, A.H.: RPL-based routing protocols in IoT applica-
tions: a review. IEEE Sens. J. 19(15), 5952–5967 (2019)
5. Sheng, Z., Yang, S., Yu, Y., Vasilakos, A.V., McCann, J.A., Leung, K.K.: A survey on the
IETF protocol suite for the internet of things: standards, challenges, and opportunities. IEEE
Wirel. Commun. 20(6), 91–98 (2013)
6. Zhang, Z.K., Cho, M.C.Y., Wang, C.W., Hsu, C.W., Chen, C.K., Shieh, S.: IoT security:
ongoing challenges and research opportunities. In: 2014 IEEE 7th International Conference
on Service-Oriented Computing and Applications, pp. 230–234. IEEE (2014)
7. Bhale, P., Biswas, S., Nandi, S.: Liene: lifetime enhancement for 6lowpan network using
clustering approach use case: smart agriculture. In: Krieger, U.R., Eichler, G., Erfurth, C.,
Fahrnberger, G. (eds.) I4CS 2021. CCIS, vol. 1404, pp. 59–75. Springer, Cham (2021). https://
doi.org/10.1007/978-3-030-75004-6_5
8. Kuna, B., Fu, A., Susilo, W., Yu, S., Gao, Y.: A survey of remote attestation in internet of
things: attacks, countermeasures, and prospects. Comput. Secur. 112, 102498 (2022)
9. Bang, A.O., Rao, U.P., Kaliyar, P., Conti, M.: Assessment of routing attacks and mitigation
techniques with RPL control messages: A survey. ACM Comput. Surv. (CSUR) 55(2), 1–36
(2022)
10. Alsukayti, I.S., Singh, A.: A lightweight scheme for mitigating RPL version number attacks
in IoT networks. IEEE Access (2022)
11. Ray, D., Bhale, P., Biswas, S., Nandi, S., Mitra, P.: DAISS: design of an attacker identifica-
tion scheme in COAP request/response spoofing. In: TENCON 2021–2021 IEEE Region 10
Conference (TENCON), pp. 941–946. IEEE (2021)
12. Goyal, M., Dutta, M.: Intrusion detection of wormhole attack in IoT: a review. In: 2018 Inter-
national Conference on Circuits and Systems in Digital Enterprise Technology (ICCSDET),
pp. 1–5. IEEE (2018)
13. Krčo, S., Pokrić, B., Carrez, F.: Designing IoT architecture (s): a European perspective. In:
2014 IEEE World Forum on Internet of Things (WF-IoT), pp. 79–84. IEEE (2014)
14. Radanliev, P., De Roure, D., Nicolescu, R., Huth, M.: A reference architecture for integrating
the Industrial Internet of Things in the Industry 4.0. arXiv preprint arXiv:1903.04369 (2019)
15. Nandhini, P., Kuppuswami, S., Malliga, S.: Energy efficient thwarting rank attack from RPL
based IoT networks: a review. Materials Today: Proceedings (2021)
16. Raoof, A., Matrawy, A., Lung, C.H.: Routing attacks and mitigation methods for RPL-based
internet of things. IEEE Commun. Surv. Tutor. 21(2), 1582–1606 (2018)
17. Kamble, A., Malemath, V.S., Patil, D.: Security attacks and secure routing protocols in RPL-
based Internet of Things: survey. In: 2017 International Conference on Emerging Trends and
Innovation in ICT (ICEI), pp. 33–39. IEEE (2017)
18. Le, A., Loo, J., Chai, K.K., Aiash, M.: A specification-based IDS for detecting attacks on
RPL-based network topology. Information 7(2), 25 (2016)
 A Novel Approach to Detect Rank Attack in IoT Ecosystem 221

19. Perazzo, P., Vallati, C., Anastasi, G., Dini, G.: DIO suppression attack against routing in the
internet of things. IEEE Commun. Lett. 21(11), 2524–2527 (2017)
20. Wallgren, L., Raza, S., Voigt, T.: Routing attacks and countermeasures in the RPL- based
internet of things. Int. J. Distrib. Sens. Netw. 9(8), 794326 (2013)
21. Glissa, G., Rachedi, A., Meddeb, A.: A secure routing protocol based on RPL for Internet of
Things. In: 2016 IEEE Global Communications Conference (GLOBE-COM), pp. 1–7. IEEE
(2016)
22. Osman, M., He, J., Mokbal, F.M.M., Zhu, N., Qureshi, S.: Ml-LGBM: a machine learning
model based on light gradient boosting machine for the detection of version number attacks
in RPL-based networks. IEEE Access 9, 83654–83665 (2021)
23. Sahay, R., Geethakumari, G., Mitra, B.: A novel block chain based framework to secure
IoT-LLNS against routing attacks. Computing 102(11), 2445–2470 (2020)
24. Le, A., Loo, J., Luo, Y., Lasebae, A.: Specification-based IDS for securing RPL from topology
attacks. In: 2011 IFIP Wireless Days (WD), pp. 1–3. IEEE (2011)
25. Pongle, P., Chavan, G.: A survey: attacks on RPL and 6LoWPAN in IoT. In: 2015 International
Conference on Pervasive Computing (ICPC), pp. 1–6. IEEE (2015)
26. Cervantes, C., Poplade, D., Nogueira, M., Santos, A.: Detection of sinkhole attacks for sup-
porting secure routing on 6lowpan for internet of things. In: 2015 IFIP/IEEE International
Symposium on Integrated Network Management (IM), pp. 606–611. IEEE (2015)
27. Le, A., Loo, J., Lasebae, A., Vinel, A., Chen, Y., Chai, M.: The impact of rank attack on
network topology of routing protocol for low-power and lossy networks. IEEE Sens. J. 13(10),
3685–3692 (2013)
28. Rai, K.K., Asawa, K.: Impact analysis of rank attack with spoofed IP on routing in 6lowpan
network. In: 2017 Tenth International Conference on Contemporary Computing (IC3), pp. 1–
5. IEEE (2017)
29. Shukla, S., Singh, S., Kumar, A., Matam, R.: Defending against increased rank attack on
RPL in low-power wireless networks. In: 2018 Fifth International Conference on Parallel,
Distributed and Grid Computing (PDGC), pp. 246–251. IEEE (2018)
30. IoT-Simulator: Cooja. https://anrg.usc.edu/contiki/index.php/Cooja~Simulator. Accessed 29
July 2019
31. Bhale, P., Biswas, S., Nandi, S.: Ml for IEEE 802.15. 4e/TSCH: Energy efficient approach to
detect DDoS attack using machine learning. In: 2021 International Wireless Communications
and Mobile Computing (IWCMC), pp. 1477–1482. IEEE (2021)
 Energy Efficient Adaptive Mobile Wireless
Sensor Network in Smart Monitoring
Applications

Seli Mohapatra1,2(B) and Prafulla Kumar Behera1

1 Utkal University, Vani Bihar, Bhubaneswar, Odisha, India
[email protected]
2 C. V. Raman Global University, Mahura, Bhubaneswar, Odisha, India

Abstract. The Mobile Wireless Sensor Networks (MWSNs) owe its name due to
mobile sinks or mobile sensor nodes as one category of heterogeneous networks
with induced mobility of nodes making it autonomous and more suitable for smart
monitoring applications. The possible application in their implementation context
in home and disaster management involves the integration of mobile WSN in
ubiquitous computing. With the enormous growth of smart devices, the data shar-
ing among entities lead to hotspot problem nearby sink, that requires a powerful
computing, communicating and storage capable mobile devices, can benefit net-
work in terms of scalability, efficiency and data delivery speed. In this paper, we
proposed an adaptive heterogeneous multi-tiered architecture based mobile sen-
sor network and analyzed its performance and routing efficiency with respect to
the static wireless sensor network. The performance metrics used help in vali-
dating our proposed method suitable for smart data collecting and dissemination
for monitoring applications. Again, the low power wireless personal area network
(6LoWPAN) plays a vital role in interconnecting these IoT devices and imple-
menting mobility of nodes with appropriate data rate for robust communication
and achieving desired Quality of Service.

Keywords: Mobile wireless sensor network · Mobile hotspot · Heterogeneous

multi-tier model · Clustering · M-LEACH

1 Introduction
Smart Monitoring Applications is a great revolutionary aspect of the decade, made
more accountable and intelligent by embedding sensors and providing communication
among the network. The integration of IoT(Internet of Things) with wireless sensor
networks gives a vast range of application perspectives for monitoring like smart home,
field or environment monitoring, smart disaster management, or industrial automation.
Moreover, various companies have also supported the integration of IoT with wireless
sensor network. The crucial element of the Internet of Things is wireless communication
across different platforms and devices without human-to-human or human-to-device
interaction. It is important to think of a quick and flexible way to experiment with the

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 222–235, 2022.
https://doi.org/10.1007/978-3-031-23233-6_17
 Energy Efficient Adaptive Mobile Wireless Sensor Network 223

behaviour of an application, a protocol, and in a repeatable manner due to unforeseen

events and the physical features of wireless propagation and the internet of things. Thus,
it becomes a multi-objective optimization problem, resolved by constrained numerical
methods. To achieve the coverage as a major objective, connectivity and mobility are
considered as constraint, including energy optimization and improvement in network
lifetime. As a result, the Mobile Wireless Sensor Network coordinates the operation
of sensors to enable long-term monitoring of the environment over a vast surface area
[1]. The sinks and sensor nodes are presumed to be static in the conventional wireless
sensor network with heterogeneous nodes. These devices are energy efficient, reliable,
and commercial implementation is straightforward in their homogeneous counterpart.
However, in a large-scale network and real-time monitoring session, the static network
lacks data gathering and data dissemination. Thus, the self-organizing robots or randomly
deployed nodes that can be made mobile under certain circumstances become more
effective leads the implementation of mobile wireless sensor network [2]. Finally, the
integration of two distinct network entities, such as the data acquisition network and
a data dissemination network, seems to be how WSNs are best defined at a higher
level. Due to the numerous technologies involved, it might be challenging to choose the
components work for such complicated networks [3]. In order to create a consistent,
reliable, and strong overall system, these networks are monitored and controlled by the
control center. as shown in Fig. 1.

Fig. 1. Network diagram for wireless sensor networks

224 S. Mohapatra and P. K. Behera

The major challenges in integrating the MWSN to IoT is the data collection and data
sharing with energy management. Hence, various topology control and routing schemes
based on clustering has been incorporated. The issue we deal with here begins with a
coverage area’s random distribution of heterogeneous nodes. The first step is to organize
the networked in different cluster groups depending on energy distribution among nodes.
Assuming the nodes in a group contain at least one cluster head or routing node and
other sensing nodes, the network gets connected as a whole including the sink node.
However, some routing nodes undergoes controlled mobility pattern depending on the
data dissemination to the sink node, that aims to integrate the Internet of Things with
the Mobile Wireless Sensor Network.
The main contribution to our paper is to exploit the IoT based applications on MWSN
(mobile wireless sensor network) context which can be possible with the node hetero-
geneity in mobile sensors. The main objective of our paper proposals are summarized
as follows.

• A system model for level based heterogeneous mobile network is proposed that
signifies a better QoS.
• To initiate the criteria of energy and mobility model on heterogeneous network on the
basis of threshold parameter and received signal strength indicator.
• The performance parameters for energy utilization and lifetime maximization in het-
erogeneous network of different variants suitable for basic IoT applications with static
and mobile sensors are analyzed.
• Other routing protocols of tier based heterogeneous networks under stationary and
mobile condition of routing nodes with respect to multihop network are compared.

The rest of the paper is organised as follows. The background information and
related activities needed to construct the architecture of MWSN is provided in Sect. 2.
In Sect. 3, the proposed system model and algorithm are discussed. Section 4 then sets
up the simulation and analyses the results. The paper was concluded in Sect. 5.

2 Related Works
One of the promising areas in the field of IoT is monitoring applications in different real
time fields like home, agriculture, industry or forest [4]. A set of technologies and devices
has been surveyed and exposed to develop system model integrated for monitoring and
architectural applications [5]. Various papers have been surveyed to classify the types of
wireless sensor networks based on sensor type, deployment strategy, mobility pattern,
architectural design, sensing technique, coverage and connectivity pattern [6]. In our
paper, we have considered sensor network variant like static and mobile, clustered het-
erogeneous hierarchical architecture with different mobility, coverage and connectivity
pattern.
In [7], the authors have discussed on the requirements and features to offer diverse
and heterogeneous tools and services on testbeds. But there existed some gaps and
difficulties to choose a static or mobile sensor network depending on needs. However,
the authors from [8] focused on the deployment and working of static wireless sensor in
homogeneous and heterogeneous scenario.
 Energy Efficient Adaptive Mobile Wireless Sensor Network 225

The coverage and connectivity among the nodes are important perspective of sensor
networking and has been addressed by several authors in different methods adapted. In
[9], authors developed a decentralized solution for adjustment and improving the cover-
age and connectivity for prolonging the lifetime in sensor network. Snagwan et. al [10]
addresses the problem of supervision of at least one node at targeted region. The sensing
ability and energy consumption features are described to define the network lifetime.
However, based on the category of sensor network the mobility based on placement of
nodes in robots and its coverage pattern is observed in [11]. The author Tirandazi et.
al [12] created a dynamic network by robot path planning algorithm where the nodes
are placed on mobile robots in two steps. With local phase and global phases analysis
reveals the combined connected coverage performance that effect the overall network
operation.
The spatially distributed set of autonomous connected sensor nodes create hotspot
problems near the sink node. So, the dynamic heterogeneous network issues some rout-
ing strategy to mitigate this problem and enhance the energy efficiency and network
longevity. The authors Khalaf et. al in [13] explained the dynamic and unequal cluster-
ing technique by providing solution to hotspot problem from various parameters like
cluster head selection, number of clusters, zone formation and different routing param-
eters. Hence, clustering methods become an important and reliable method in static and
mobile wireless sensor network which various authors have focused with different rout-
ing strategies. The Cluster Head selection methods are addressed by many researchers
based on different criterion over LEACH (Low-energy adaptive clustering hierarchy).
The distance-based approach [14] optimizes energy, in Cluster Head (CH) Selection, by
equal distribution of network load. The residual energy distribution for CH selection is
another important criterion explained by Behera et.al in [15]. Similarly, the statistical
methods have also significantly improved the network lifetime and energy efficiency as
discussed in the paper Mohapatra et al. [16].
The major contributions of this paper are as follows.

• The network development is initiated with a random deployment of nodes which may
not be connected. Thus, the first objective is to ensure the full connectivity among the
nodes within appropriate coverage level.
• Based on communication range cluster groups are formed with selection of Cluster
Head based on statistical based residual energy of nodes.
• The data collected from remote node are disseminated at the sink node via Cluster
head. However, if the communication distance between CH and Sink node is large,
then in accordance to greedy algorithm, some group of nodes or the Cluster Head node
provide adaptive logic in greedy direction towards Sink Node for further processing.

The performance evaluation of our proposed method then evaluated with two basic
scenarios, with respect to static and mobile wireless sensor network with or without
clustering.
226 S. Mohapatra and P. K. Behera

3 Proposed Method
The design of low power and low delay mobile wireless sensor network is a challenging
task, as in such network’s topology is moderately dynamics and the mobile sensor nodes
are equipped with poor sensing and processing capabilities. So, it is required to develop a
hierarchical or tiered based heterogeneous model with controlled mobility to achieve the
requirement of energy consumption and data delay. In the paper, we proposed a mobile
wireless sensor network model with three tier architecture in which some sensor nodes
are mobile by adaptive method as discussed below. For this work we have considered
following assumptions.

• After the random deployment of nodes, the node positions are determined on initial-
ization. The heterogeneity is maintained based on non-uniform energy distribution.
• Nodes with higher energy levels are preferred for cluster head formation. However, it
is assumed that the distribution of nodes is such that in each group at least one CH is
possible. Also, these nodes are GPS enabled to track the location.
• Based on the distance of nodes to higher energy node Cluster Head or Sink Node the
node gets mobility associated with greedy algorithm to destination.
• The sink node is allocated at the center of the network and is considered as the base
station.

3.1 Mobile Sensor Node Architecture

Fig. 2. Architecture of mobile sensor node

The mobile sensor node architecture is almost identical to that of a regular sensor
node, with the inclusion of a few extra elements like locators, power sources, and mobi-
lizers. The sensing unit uses one or more sensors to provide basic sensing. Similar to
this, the processing unit contains memory and motes processors to execute signal pro-
cessing on digitized data with considering data aggregation into account. Once more, the
communication unit identifies the routing path and chooses the appropriate radio range.
 Energy Efficient Adaptive Mobile Wireless Sensor Network 227

To identify the position, generate more electricity for nodes like solar cells, and control
the mobility of sensor nodes, additional units are however employed. The architecture
of Mobile sensor node is shown in Fig. 2. The majority of MWSN considers large-scale
applications made up of numerous sensor and sink nodes, where the mobility of the
nodes depends on the application’s specific requirements. The key factor for the design
of a MWSN are scalability and topology, data routing, mobility and residual energy
utilization [17]. Our proposed model is based on three phases (a)Establishment Phase
(b)Routing and Clustering phase.

3.1.1 Establishment Phase

The topology of MWSN is considered dynamic as the connectivity among the CHs to
sensor nodes varies with time due to its mobility as and when desired[18]. The position
finder helps to place the nodes throughout the area is proper distribution. The sensor
nodes can be placed distributed randomly. So, if N nodes are put in an area of A = LxL
sq. m, so the sensor node density can be ρ = NA . The sensing range r, determines the
forwarding of data between nodes which must be at most the communication range 2r.
If the nodes are uniformly distributed then the probability that there are m nodes within
the space S is Poisson distributed as

(ρS)m −ρS
P(m) = e (1)
m!
where S = π r 2 for 2-D Space. Thus, the probability that the monitored space is covered
by at least one sensor node is given by

Pcov = 1 − P(0) = 1 − e−ρS (2)

. The sensor node computes the distance between its post deployment location and
the locations associated with the pre distributed units. The smaller distance having higher
priority over larger distance between adjacent nodes.

3.1.2 Routing and Clustering Phase

Another essential area for mobile WSN is routing and clustering. The network is driven
into a structure by the routing protocols to achieve stability, scalability, and energy effi-
ciency. The cluster head, nodes with high residual energy, coordinate cluster operations
and forward information among sensor nodes and sink nodes. This class of routing
protocol causes clustering. These cluster heads can be seen as dynamic, nevertheless,
depending on the situation. Mobility to the sensor nodes can be performed using mobi-
lizers to change places, wheels, or robots through the use of an suitable mobility pattern.
[19]. Mobile nodes’ dynamic topology depicts how routing, MAC, and physical charac-
teristic implementation are used in the network. Again, static WSN is inefficient owing
to energy exhaustion or device failure, but MWSN reorganizes the network. The ideal
routing protocol is one that never uses up too much network resources while still covering
all network states. The MWSN routing strategies can be both traditional and optimized.
When designing a routing protocol for MWSNs, consideration is given to the network
228 S. Mohapatra and P. K. Behera

topology, sensor node mobility, energy consumption, network coverage, data transmis-
sion techniques, quality of service (QoS), connectivity, data aggregation, sensor node
and communication link heterogeneity, scalability, and security. As a result, it’s crucial to
select a mobile node that can prolong network lifetime and attain better energy efficiency
[20].
Consider a network with N sensor nodes which is deployed over the sensor field
with remote/sensing nodes followed by advanced excluding the Sink Node. The levelled
based heterogeneous nodes are defined as

Nr = 0.7 ∗ N and Na = 0.3 ∗ N (3)

where Nr represents normal nodes with 20% being GPS enabled and rest 80% are without
GPS and Na are the advanced node.
After deployment the nodes configure their respective locations with coordinate
value (xi ,yi ). If rc is the communication range and rs is the sensing range of node N. then
Pt− Pth
rc = 10 10η (4)

rc ≤ 2r s (5)

with Pt being transmitter power and Pth being minimum threshold received power
required by nodes and η is the path loss factor.
The network nodes begin monitoring events that take place within sensing range
and report to CH using single-hop and multi-hop methods. However, nodes may cluster
in some areas and may generate coverage holes where nodes are not available when
sensors are randomly distributed in hostile target fields, resulting in network partitioning.
Therefore, it is necessary to find and fix the network’s gaps. In order to locate the holes, it
is reported to the base station or sink. In order to obtain a complete connectivity network,
the nodes are re-localized by introducing mobility into advanced nodes or sink nodes.
Either the sink node instructs the CH to travel a specific distance till the entire network
is joined to solve this issue.
When remote nodes and reference nodes are deployed at random, the nodes are
localized by calculating their distance from reference nodes using the total of their
shortest-path communication ranges. And utilizing the Angle of Arrival approach to
assess the sites. Nodes also identify coverage gaps and unconnected networks by exam-
ining the distribution of nearby neighbor nodes. To achieve the desired coverage, node
density should be properly considered as in Eq. (6).
N Nr + Na
ρ= = (6)
A π r0 2
rs
where r0 = 1.066 =Minimal sensing Radius of heterogeneous nodes.
Every initial GPS enabled node broadcasts location information (x,y) and rci to
neighbour nodes. The neighbour nodes store this information and add its own rcj . If
rci + rcj ≤ rcth , then nodes disseminate them together with reference node positions. By
attempting to segregate only the nearest reference nodes, the rcth minimizes the impact of
 Energy Efficient Adaptive Mobile Wireless Sensor Network 229

asymmetric bent routes on distance estimations. Following the submission of this data,
neighbour nodes check to see if the reference node information has already been saved
in their neighbor data. The received reference node location and Critical Range sum are
updated in the neighbor information if it corresponds to a new node. The received sum
of rc is compared to the stored value if the obtained node information has already been
saved, and the lowest of these two rc values is saved along with the location data. Each
node will have the locations of the reference nodes and the associated minimum value
of the sum of the rc from the reference node to itself at the completion of this stage [21].

4 Simulation and Result Analysis

In this section, we have simulated and analyzed our proposed method for various perfor-
mance metrics with its static sensor network counterpart. We have considered LEACH
and SEP under static scenario and with M-LEACH as mobile scenario and compared
them with our proposed implemented Mob M-LEACH.

4.1 Simulation Set Up

Fig. 3. A three tier heterogeneous network

The simulation work is carried out in MATLAB2020b with Intel core i5 processor,
64-bit operating system with 1.8 GHz frequency [22]. After the random deployment of
nodes, the three-tiered heterogeneous network is represented as in Fig. 3 and the required
simulation parameters is cited in Table 1.

4.2 Experimental Analysis

The paper focus of two basic simulation environment, static or dynamic sensor node
in small coverage area like smart home or office and adaptive mobile sensor network
230 S. Mohapatra and P. K. Behera

Table 1. Simulation parameters

Sl No. Parameters Font Size and Style

1 Number of nodes (N) 10, 50, 100, 200, 400, 600, 800, 1000
2 Topography 500 X 500 m2 ,
500 X 500 m2
3 Communication range (d0 ) 87 m
4 Simulation time 300 s
5 No. of rounds(r) 6000
6 Initial energy of normal node (E0 ) 1J
7 Packet Length 500 bits, 127 bytes
8 Mobility model Controlled with velocity 10 m/s
9 ETX (nJ/bit) 50
10 ERX (nJ/bit) 50
11 Eelec (nJ/bit) 50
12 EDA (nJ/bit/f rame) 5
13 εfs (pJ/bit/m2) 10
14 εmp (pJ/bit/m4) 0.0013

in large coverage area like smart farm monitoring applications. Thus, the performance
metrics [23] considered for these scenarios are the network lifetime with network con-
vergence factor, speed of coverage, end-to-end delay, energy efficiency. The analysis of
the mobility is based on above mentioned criteria with 500 m2 and 5000 m2 coverage
area with network size variations between 50 to 800 nodes.

4.2.1 Network Lifetime Analysis

The lifetime of a network is basically determined by its network convergence expressed
in Eq. 1 where LND is the last node dead factor, HND is the 50% dead factor and FND
is the first node dead factor.
LND − HND
NCI = (7)
HND − FND
The network lifetime indicates the longevity of sensor nodes of the network which
can be illustrated from the availability of nodes to carry out data processing task in the
network. As shown in Fig. 4(a) our proposed MSWN can sustain for large rounds of
operation for performing its task as the all-dead nodes are obtained considering maximum
rounds than other methods. However, the network lifetime is better analyzed from its
convergence factor as shown in Fig. 4(b). It is seen that the network sustainability is
better in our proposed MWSN in comparison to other protocols for any variation of
nodes. However, in a coverage area of 100 sqm, as the number of nodes increases
the convergence factor decreases due to inefficient data handling capability with more
 Energy Efficient Adaptive Mobile Wireless Sensor Network 231

interference levels. It also ensures that to implement an IoT in coverage area of 100 sqm.
we need to choose a proper scalable network suitable for smart applications.

Fig. 4. Analysis of (a) network lifetime (b) Network convergence indicator

232 S. Mohapatra and P. K. Behera

4.2.2 Residual Energy Analysis

The residual energy analysis indicates the energy remains in the entire network by
advanced node for different coverage area. Figure 5 represents the residual energy dis-
tribution over 100 sqm. From the graph it indicates the proposed mobile sensor network
with clustering have more energy and can sustain for more rounds while choosing the
cluster heads.

Fig. 5. Analysis of residual energy distribution for 100 nodes in the entire network with 100 m2

4.2.3 Network Latency Analysis

Fig. 6. Analysis of packets sent to sink node from CH with coverage area 100 m2

The network latency analysis directs the way in which robust data packets reach the
server as fast as possible. This factor signifies under the above-mentioned criteria, the
 Energy Efficient Adaptive Mobile Wireless Sensor Network 233

basic monitoring applications must utilize the method where data gathering and delivery
to sink node should be very fast with minimal loss [23]. In the paper, we compared
our proposed model with the its static and mobile sensor network Fig. 6 indicates the
packets sent to the sink node for the entire simulation duration. It is observed the adaptive
movement of sensor nodes in our proposed method gives the maximum data to the
sink node for any coverage area. Thus, it becomes suitable for monitoring based IoT
applications in different scenarios. Similarly, the time elapsed for network indicates the
computational complexity as shown in Fig. 7.

Fig. 7. Analysis of time elapsed for data delivery of network with coverage area 100 m2

5 Conclusion

With the advent of new technologies in the last few years, the integration of IoT and
Mobile Wireless Sensor Network has envisaged the ability of entities and environment
to recognize, communicate and share the data. The heterogeneity of nodes in an adaptive
mobile platform can be controlled via remote access. In this paper, we have proposed a
new adaptive method to develop mobile wireless senor network, specifically designed
for monitoring based IoT applications. The architectural implementation of layered
based clustered heterogeneous network with adaptive mobility of sensor node, followed
by analysis of performance related issues on the aspect of mobility is discussed. It is
observed that, our proposed hierarchical clustered mobile wireless sensor network is
more efficient than its static and mobile sensor network and it is also scalable to a
variety of network size. It reveals the fact that our proposed model is more useful in
different IoT based monitoring applications, from smart home to smart farm monitoring
scenario. Again, the energy efficiency and data dissemination over the network properly
address the effective data flow throughout the network to achieve the desired quality of
service (QoS).
234 S. Mohapatra and P. K. Behera

References
1. Reddy, V., Gayathri, P.: Integration of internet of things with wireless sensor network. Int. J.
Elec. Comput. Eng. (IJECE), 9, 439 (2019)
2. Tonneau, A.-S., Mitton, N., Vandaele, J.: How to choose an experimentation platform for wire-
less sensor networks? a survey on static and mobile wireless sensor network experimentation
facilities
3. Mhatre, V., Rosenberg, C.: Homogeneous vs heterogeneous clustered sensor networks: a
comparative study. In: 2004 IEEE International Conference on Communications (IEEE Cat.
No.04CH37577), vol. 6, pp. 3646–3651 (2004)
4. Etancelin, J.-M., Fabbri, A., Guinand, F., Rosalie, M.: DACYCLEM: A decentralized algo-
rithm for maximizing coverage and lifetime in a mobile wireless sensor network. Ad Hoc-
Networks, vol. 87, pp. 174–187 (2019). https://www.sciencedirect.com/science/article/pii/
S1570870518309430
5. Chen, X., Yu, P.: Research on hierarchical mobile wireless sensor network architecture with
mobile sensor nodes. In: 2010 3rd International Conference on Biomedical Engineering and
Informatics, vol. 7, pp. 2863–2867 (2010)
6. Hermanu, C., Maghfiroh, H., Santoso, H.P., Arifin, Z., Harsito, C.: Dual mode system of smart
home based on internet of things. J. Robot. Control (JRC) 3(1), 26–31 (2022)
7. Mohamed, S.M., Hamza, H.S., Saroit, I.A.: Coverage in mobile wireless sensor networks
(M-WSN): a survey. Comput. Commun. 110, 133–150 (2017)
8. Mohapatra, S., Mohapatra, R.K.: Comparative analysis of energy efficient mac protocol in
heterogeneous sensor network under dynamic scenario. In; 2017 2nd International Conference
on Man and Machine Interfacing (MAMI), pp. 1–5 (2017)
9. Liu, D., Ning, P.: Improving key predistribution with deployment knowledge in static sensor
networks. ACM Trans. Sen. Netw. 1(2), 204–239 (2005). https://doi.org/10.1145/1105688.
1105691
10. Huang, C.-F., Tseng, Y.-C., Wu, H.-L.: Distributed protocols for ensuring both coverage and
connectivity of a wireless sensor network. ACM Trans. Sen. Netw.3(1), p. 5–es (2007). https://
doi.org/10.1145/1210669.1210674
11. Sangwan, A., Singh, R.P.: Survey on coverage problems in wireless sensor networks. Wireless
Pers. Commun. 80(4), 1475–1500 (2015)
12. Tirandazi, P., Rahiminasab, A., Ebadi, M.: An efficient coverage and connectivity algorithm
based on mobile robots for wireless sensor networks. J. Ambient Intell. Humanized Comput.,
1–23 (2022)
13. Khalaf, O.I., Romero, C.A.T., Hassan, S., Iqbal, M.T.: Mitigating hotspot issues in heteroge-
neous wireless sensor networks. J. Sens. (2022)
14. Rajput, M., Sharma, S.K., Khatri, P.: Energy-Efficient multihop cluster routing protocol for
WSN. In: Poonia, R.C., Singh, V., Singh Jat, D., Diván, M.J., Khan, M.S. (eds.) Proceedings of
Third International Conference on Sustainable Computing. Advances in Intelligent Systems
and Computing, vol 1404, pp. 77–84. Springer, Singapore (2022). https://doi.org/10.1007/
978-981-16-4538-9_8
15. Behera, T.M., Mohapatra, S.K., Samal, U.C., Khan, M.S., Daneshmand, M., Gandomi, A.H.:
Residual energy-based cluster-head selection in WSNs for IoT application. IEEE Int. Things
J. 6(3), 5132–5139 (2019)
16. Mohapatra, S., Behera, P.K.: Statistical approach based cluster head selection in heterogeneous
networks for IoT applications. In: Behera, P.K., Sethi, P.C. (eds.) Digital Democracy – IT
for Change. CSI 2020. Communications in Computer and Information Science, vol 1372,
pp. 77–84. Springer, Singapore (2021). https://doi.org/10.1007/978-981-16-2723-1_4
 Energy Efficient Adaptive Mobile Wireless Sensor Network 235

17. Atay, N., Bayazit, B.: Mobile wireless sensor network connectivity repair with K-Redundancy.
In: Chirikjian, G.S., Choset, H., Morales, M., Murphey, T. (eds.) Algorithmic Foundation of
Robotics VIII. Springer Tracts in Advanced Robotics, vol 57, pp. 35–49. Springer, Heidelberg
(2009). https://doi.org/10.1007/978-3-642-00312-7_3
18. Sahoo, P.K., Hwang, I.-S.: Collaborative localization algorithms for wireless sensor networks
with reduced localization error. Sensors 11(10), 9989–10009 (2011). https://www.mdpi.com/
1424-8220/11/10/9989
19. Buehrer, R.M., Wymeersch, H., Vaghefi, R.M.: Collaborative sensor network localization:
algorithms and practical issues. Proc. IEEE 106(6), 1089–1114 (2018)
20. Sara, G.S., Sridharan, D.: Routing in mobile wireless sensor network: a survey. Telecommun.
Syst. 57(1), 51–79 (2014)
21. Tolba, F.D., Ajib, W., Obaid, A.: Distributed clustering algorithm for mobile wireless sensors
networks. SENSORS. IEEE 2013, 1–4 (2013)
22. Amine, D., Nassreddine, B., Bouabdellah, K.: Energy efficient and safe weighted clus-
tering algorithm for mobile wireless sensor networks. Proc. Comput. Sci. 34, 63–70
(2014). The 9th International Conference on Future Networks and Communications (FNC
2014)/The 11th International Conference on Mobile Systems and Pervasive Computing
(MobiSPC 2014)/Affiliated Workshops. https://www.sciencedirect.com/science/article/pii/
S1877050914008953
23. Mohapatra, S., Kanungo, P.: Performance analysis of AODV, DSR, OLSR and DSDV routing
protocols using NS2 simulator. In: Procedia Engineering, vol. 30, pp. 69–76 (2012). Interna-
tional Conference on Communication Technology and System Design (2011). https://www.
sciencedirect.com/science/article/pii/S1877705812008454
 Orthogonal Chirp Division Multiplexing:
An Emerging Multi Carrier Modulation Scheme

Mohit Kumar Singh(B) and Ashish Goel

ECE Department, Jaypee Institute of Information Technology, Noida, India

{19402019,ashish.goel}@mail.jiit.ac.in

Abstract. OCDM, a MCM scheme compatible to widespread OFDM, has been

proposed in the literature. It provides maximum spectral efficiency for chirp spread
spectrum system and outperforms over OFDM. In this paper, we have discussed
the generation of OCDM system and its implementation using OFDM system
with some minor correction. The computational complexity analysis of OCDM
system is also carried out in this paper and its comparison with OFDM system is
also done. It has been found that like OFDM, OCDM system also suffers from
the problem of high PAPR. Some of the popular schemes for PAPR reduction
applicable to OCDM system have been discussed and its performance analysis in
terms of PAPR and BER performance is also carried out in this paper. Further,
applications of OCDM for wireless communication, optical fiber communication,
UWA communication and MIMO system is also presented in detail.

Keywords: Chirp Signal · DFnT · OCDM · PAPR

1 Introduction
Chirp, a kind of signal whose phase changes with time, being of spread spectrum
nature guarantees secure and robust communication. X. Ouyang et al. [1] introduced an
Multi Carrier Modulation (MCM) technique for high-speed communication and named
it Orthogonal Chirp Division Multiplexing (OCDM). In OCDM, N orthogonal chirps
of same bandwidth are multiplexed. Fresnel transform is the fundamental mechanism
behind OCDM. Implementation of OCDM system in digital domain is obtained by Dis-
crete Fresnel Transform (DFnT). At the transmitter side, OCDM signal is generated
using Inverse Discrete Fresnel Transform (IDFnT) while at the receive side, OCDM sig-
nal is recovered using DFnT. Similar to Orthogonal Frequency Division Multiplexing
(OFDM) [2], OCDM system sends the modulated signals block-wise. Between two con-
secutive blocks, Guard Interval (GI) is inserted to avoid Inter Symbol Interference (ISI).
To fill the GI, Zero Padding (ZP) and Cyclic Prefix (CP) both can be used. For modula-
tion, amplitude and/or phase of each chirp are modulated therefore QAM and PSK can
be used as modulation scheme in OCDM system. Being a MCM scheme, OCDM system
also suffers from high Peak-to-Average Power Ratio (PAPR). Various PAPR reduction,
equalization and channel estimation schemes presented in the literature for OFDM sys-
tem are applicable for OCDM system also [1] and various precoding matrices can be
used to implement precoded OCDM [3].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 236–246, 2022.
https://doi.org/10.1007/978-3-031-23233-6_18
 Orthogonal Chirp Division Multiplexing 237

Partial Transmit Sequences (PTS) [4] is an attractive distortion less PAPR reduction
technique because there is no limitation on number of subcarriers but the transmitter
is required to send the optimum phase rotation factors, which is used to provide the
minimum PAPR, to the receiver. These phase factors are known as Side Information (SI).
Sending SI to receiver reduces the spectrum efficiency and it is very difficult to receive the
correct SI in multipath environment. Different PTS based schemes have been proposed to
eliminate the requirement of transmitting SI to the receiver. Multipoint square mapping
(MSM) combined with PTS (M-PTS) [5] and Concentric Circle Mapping-based PTS
(CCM-PTS) [6] are two of them.
In this paper, authors have presented the overview of OCDM along with its appli-
cations in various communication system. The main contribution of this paper is to
analyze the performance of OCDM system using PTS based PAPR reduction schemes
like M-PTS and CCM-PTS. Computational complexity of these PTS based systems are
also analyzed and their PAPR reduction capabilities along with their SER performance
over AWGN and Rayleigh fading channels are also given. OFDM is also simulated for
comparison. PAPR performance of OCDM and OFDM is same without any PAPR reduc-
tion scheme. PTS, M-PTS and CCM-PTS provide almost same reduction in PAPR for
OCDM and OFDM both. SER performance of OCDM with PTS, M-PTS and CCM-PTS
is identical to the performance of OFDM with PTS, M-PTS and CCM-PTS respectively
over AWGN channel. Over multipath Rayleigh fading channel, performance of OCDM
slightly degraded in the low SNR region whereas OCDM shows much better performance
compared to OFDM in high SNR region.
Rest of the paper is arranged as follows: Sect. 2 explains the compatibility of OCDM
with OFDM. Section 3 explains the computational complexity of OCDM. Section 4 pro-
vides various applications of OCDM. Section 5 provides results of computer simulation
and finally conclusion is drawn in Sect. 6.

2 Compatibility with OFDM

The discrete time OCDM signal [1] is given by

x = φH s (1)
 T
Here s = s0 , s1 , s2 . . . . . . ., sN −1 is information carrying symbol vector, x =
 T
x0 , x1 , x2 . . . .., xN −1 is OCDM symbol vector and φ is DFnT matrix. The (n, k)th
entry of N ×N DFnT matrix φ is given by
⎧
j π (n−k)2
1 −j π ⎨ e N for even N
2
φ(n, k) = √ e 4 X π (2)
N ⎩e Nj n−k+ 1
2
for odd N

The discrete time OFDM signal is given by

y = WN−nk s (3)
238 M. K. Singh and A. Goel

 T
Here y = y0 , y1 , y2 . . . .., yN −1 is OFDM symbol vector and WNnk is DFT matrix. The
(n, k)th entry of N ×N DFT matrix WNnk is given by

1 2π
WNnk = √ e−j N nk (4)
N
Inspecting Eq. (2) and (4), we can say that DFnT in Eq. (2) consists of DFT of Eq. (4)
with additional quadratic phases given by.
π 2
π ej N n even N
1 (n) = e−j 4 X π π 2 (5)
e ej N n +n odd N
j 4N

and
π 2
ej N k even N
2 (k) = π 2 (6)
ej N (k −k) odd N

So, one can find the DFnT using DFT in three steps:
Step 1: Multiple with quadratic phase 1 .
Step 2: Perform DFT by FFT algorithm.
Step 3: Multiple with 2nd quadratic phase 2 .
Here 1 and 2 are the diagonal matrices having mth diagonal entries 1 (m) and
2 (m) respectively.
Using above three steps, one can integrate OCDM into the widespread OFDM system
with minor modifications (colored boxes) as shown in Fig. 1. [1].

(i). Transmitter

(ii). Receiver

Fig. 1. OCDM Integration into OFDM System

 Orthogonal Chirp Division Multiplexing 239

3 Computational Complexity
Transmitter and receiver both are considered to compute the computational complex-
ity of OCDM and OFDM with PTS, M-PTS and CCM-PTS schemes. Computational
complexity is calculated by counting the total number of complex multiplications and
additions.

3.1 Computational Complexity of OCDM

In OCDM, with N chirps and L oversampling factor, one LN point IDFnT is calcu-
lated for each sub-block at transmitter which require 2LN + 21 LN log2 (LN ) complex
multiplications and LN log2 (LN ) complex additions [1]. Therefore, for M sub-blocks,
2MLN + 21 MLN log2 (LN ) complex multiplications and MLN log2 (LN ) complex addi-
tions are required. With K phase factors, K M −1 searches are performed to find the OCDM
signal with minimum possible PAPR and (M − 1)LN complex additions are required to
combine M sub-blocks per search [7]. Therefore, K M −1 (M − 1)LN complex additions
are required to combine the sub-blocks in K M −1 searches. At the receiver, one LN point
DFnT is calculated for OCDM symbol detection which requires 2LN + 21 LN log2 (LN )
complex multiplications and LN log2 (LN ) complex additions and for demapping of infor-
mation carrying symbols, DN complex multiplications and 2DN complex additions per
OCDM symbol are also required [7]. Here D is the total number of decision regions.
Therefore, total number of complex multiplications and additions required in OCDM
with PTS, M-PTS and CCM-PTS schemes are as follows:
Total complex multiplications: 2MLN + 21 MLN log2 (LN ) + 2LN + 21 LN log2 (LN ) +
DN .
Total complex additions: MLN log2 (LN ) + K M −1 (M − 1)LN + LN log2 (LN ) +
2DN .

3.2 Computational Complexity of OFDM

In OFDM, with N subcarriers and L oversampling factor, one LN point IDFT is calculated
for each sub-block at transmitter which require 21 LN log2 (LN ) complex multiplications
and LN log2 (LN ) complex additions [1]. Therefore, for M sub-blocks, 21 MLN log2 (LN )
complex multiplications and MLN log2 (LN ) complex additions are required. With K
phase factors, K M −1 searches are performed to find the OFDM signal with minimum
possible PAPR and (M −1)LN complex additions are required to combine M sub-blocks
per search [7]. Therefore, K M −1 (M − 1)LN complex additions are required to combine
the sub-blocks in K M −1 searches. At the receiver, one LN point DFT is calculated for
OFDM symbol detection which requires 21 LN log2 (LN ) complex multiplications and
LN log2 (LN ) complex additions and for demapping of information carrying symbols,
DN complex multiplications and 2DN complex additions per OFDM symbol are also
required [7].
Therefore, total number of complex multiplications and additions required in OFDM
with PTS, M-PTS and CCM-PTS schemes are as follows:
Total complex multiplications: 21 MLN log2 (LN ) + 21 LN log2 (LN ) + DN .
240 M. K. Singh and A. Goel

Total complex additions: MLN log2 (LN )+K M −1 (M − 1)LN +LN log2 (LN )+2DN .
Table 1 summarize the computational complexity of OCDM and OFDM with PTS,
M-PTS and CCM-PTS.

Table 1. Computational Complexity of OCDM and OFDM with PTS, M-PTS and CCM-PTS

System Total Complex Multiplications Total Complex Additions

OCDM (1 + M )LN 2 + 21 log2 (LN ) + DN LN {(1 + M )log2 (LN ) + K M −1 (M − 1)}+

2DN
OFDM (1 + M ) 21 LN log2 (LN ) + DN LN {(1 + M )log2 (LN ) + K M −1 (M − 1)}+
2DN

4 Applications of OCDM
This section presents the various applications of OCDM proposed in literature.

4.1 OCDM for Wireless Communication

In [1], OCDM is proposed for wireless multipath channel and simulated for PAPR per-
formance and BER performance with (a). ZF and MMSE equalizers (b). Insufficient GI
(c). Iterative block DFE (IB-DFE) and (d). Forward error coding (FEC). For comparison,
OFDM, DFT-P-OFDM and SC-FDE are also considered. DFT-P-OFDM provides the
best PAPR performance while the OFDM and OCDM systems have the identical PAPR
characteristics. As per simulation results, OCDM is superior to the traditional OFDM
and it also exhibits a higher anti-interference ability which occurs due to insufficient GI.

4.2 OCDM for Optical Fiber Communication

In [3], OCDM is proposed for coherent optical communication. Digital implementation
of coherent optical OCDM (CO-OCDM) system is realized using DFnT. The relation
between DFnT and DFT, given in Sect. 2, is utilized to integrate the CO-OCDM into
the conventional CO-OFDM with some additional operations. In [2], CO-OCDM is
simulated for PAPR performance, average Q factor and OSNR penalties with different
GI lengths at different transmission distance and BER performance. CO-OFDM is also
considered for comparison. OCDM and OFDM have the same PAPR performance under
the same conditions. In CO-OFDM, the Q factors of the subcarriers fluctuate, but in
OCDM, all chirps have same Q factors. However, for identical GI and transmission
distance, the averaged Q factors of CO-OCDM and CO-OFDM are same. Performance
of the CO-OCDM system is better as compared to CO-OFDM systems in terms of
transmission distance at the same GI length, OSNR penalty and BER. Performance of
CO-OCDM is better as compared to CO-OFDM because OCDM is less sensitive to the
channel impairments. As CO-OFDM does, CO-OCDM is also able to compensate the
chromatic dispersion. Similar to precoded OFDM, various precoding matrices can be
used to implement precoded OCDM.
 Orthogonal Chirp Division Multiplexing 241

4.3 OCDM for IM/DD Based Short Reach Systems

In [8, 9], DSB modulated OCDM is introduced for intensity modulation and direct
detection (IM/DD) based short reach systems. IM/DD system has one dimension only
for modulation therefor the OCDM signal of complex values is not feasible for it. Pass-
band signal is obtained from complex baseband signals using Digital up conversion
(DUC) technique. Real part of passband signal which is equivalent to DSB modulated
OCDM signal is reserved for IM and imaginary part is rejected. At receiving end, dig-
ital down conversion (DDC) converts back the passband signal to baseband signal for
demodulation. Due to presence of DUC and DDC techniques, complexity of IM/DD
OCDM system is increased.
In [8, 9], open loop IM/DD OCDM and in [9] closed-loop IM/DD OCDM are sim-
ulated for BER performance with different data rates at different transmission distance.
DMT-OFDM is also considered for performance comparison. For open loop system, the
IM/DD OCDM system consistently perform better than the DMT OFDM system in all
instances. For high modulation levels and at long transmission distances, the IM/DD-
OCDM signal has a clear benefit. Because subcarriers of high frequencies are more
sensitive to fading induced by chromatic dispersion, the SNR performance of OFDM
subcarriers of high frequencies begins to drop with the increase in transmission dis-
tance. In contrast, OCDM chirps are unaffected by these defects, with flat SNRs and
relatively slight deterioration. As a result, IM/DD OCDM system performance is supe-
rior to DMT OFDM. If channel state information (CSI) is available at the transmitter,
most of the imperfections may be addressed therefore the IM/DD OCDM system shows
only a slight improvement over DMT OFDM in closed loop scenario. In comparison
to open loop system, the closed loop system shows a significant improvement for both
systems.

4.4 OCDM for Underwater Acoustic Communication

In [10, 11,12] OCDM is proposed for underwater acoustic communication (UWAC).

In [10, 11], fully and underloaded OCDM are simulated for BER performance over
static and dynamic UWAC channels. For comparison, conventional OFDM and SC with
an iterative FDE (SC-FDE) are also considered. Over static UWAC channel, due to
enhanced diversity exploitation, the fully-loaded OCDM system gives an SNR gain
compared to fully-loaded OFDM. In underloaded configuration, both OFDM and OCDM
systems give an improved BER performance described by the spread spectrum effect.
However, underloaded OCDM outperformed the fully/underloaded OFDM significantly,
particularly in the low SNR region. Over dynamic UWAC channel, SC-FDE system
provides the best performance and fully-loaded OFDM outperform the fully-loaded
OCDM. Poor performance of OCDM is because MMSE equalizer enhances the noise in
the event of Doppler spread. Underloaded OCDM outperformed the underloaded OFDM
by a large margin for low values of SNR, however, the BER performance of SC-FDE
is comparable with underloaded OFDM. With reduction in load ratio, OCDM delivers
improved BER performance and converged to the AWGN boundary because underloaded
OCDM provides spreading gain as well as diversity gain while OFDM performance
increased marginally because underloaded OFDM provides enhanced robustness against
242 M. K. Singh and A. Goel

carrier frequency offset (CFO) only which is caused by increased inter carrier spacing.
By exploiting the multipaths diversity of the channel, underloaded OCDM provides good
robustness than OFDM but at the cost of reduced data rate.
Orthogonality of the subcarriers is destroyed by the Doppler effect therefore OCDM
reacts to Doppler spread and ICI. Hence, Doppler compensation algorithm is required.
In [12], Data Pick Rake OCDM, or DP-Rake OCDM, is proposed to remove ICI at short
GI and increase the data rate. The basic purpose of rake receiver is to determine the ideal
rake finger which minimize the ICI. The computational complexity of the receiver can
be considerably increased by using number of rake fingers. In [12], DP-Rake OCDM is
simulated over static and dynamic UWA channels. Static UWA channel is based on the
measured data of Xiamen Port Shallow Water and dynamic UWA channel is based on
the Water Mark Time-varying (WMT) channel. For comparison, SC-FDE, OFDM and
OCDM are also considered.
As FEC coding, Doppler estimation, and phase correction techniques remove the
Doppler shift introduced by channel variations, the BER performance of OFDM and
SC-FDE with these techniques is same over WMT channel. OCDM outperform OFDM
over WMT channel because it provides diversity and spreading gain both through its
chirp-based signal. The BER performance of DP Rake OCDM is the best, although the
performance improvement against OCDM is not considerable because WMT channel’s
delay spread is tiny in comparison to the CP, causing no severe ICI. Since the ICI caused
by delay spread has severe impact on MCM, BER performance of OCDM and OFDM
are poor over Xiamen Port SW channel, however, OCDM slightly outperforms OFDM.
With the addition of data pick-based rake receiver, which effectively eliminates ICI, BER
performance of DP Rake OFDM and DP Rake OCDM systems enhanced considerably,
however, the performance of DP Rake OFDM lies between DP Rake OCDM and OCDM.

4.5 OCDM for Baseband Data Communication

In [13], OCDM is proposed for baseband data communication. OCDM signal of complex
values is not feasible for baseband data communication. Therefore, four types of modified
OCDM schemes which permit baseband data communication, i.e., transmission and
reception of real-valued symbols, are proposed. In [13], four modified OCDM schemes
are simulated for PAPR performance and BER performance over AWGN, ABGN, PLC
with AWGN and PLC with ABGN environments. HS-OFDM along-with SC-CP is
considered for performance comparison. SC-CP has the best PAPR characteristics, while
the four OCDM schemes and HS-OFDM have the same PAPR performance. Proposed
OCDM schemes provide better BER performance as compared to HS-OFDM and SC-CP
but has greater computational complexity.

4.6 OCDM for MIMO Communication

In [14], OCDM in MIMO with different space time coding (STC) is proposed and sim-
ulated for URLLC. In addition to Alamouti and Cyclic Delay Diversity (CDD), authors
have used Alamouti with CDD (ACDD) also for transmit diversity while Maximum Ratio
Combining (MRC) is adopted for receive diversity. For comparison, OFDM is also sim-
ulated. Since OCDM spreads the symbols uniformly in frequency, FER performance of
 Orthogonal Chirp Division Multiplexing 243

OCDM is better than OFDM in all scenarios, for Nt = 1, 2, 4 or 8 and Nr = 1 or 2, with

CDD as well as ACDD. ACDD outperforms CDD for Nt ≥ 2. With Nr = 2, performance
improvement is very large, in all situations, compared with Nr = 1. With the increase
in N t in CDD, frequency selectivity of SISO channel improves and favors the spreading
feature of OCDM. Therefore, the performance difference between OCDM and OFDM
increases with increase in N t . In ACDD, the performance difference between OCDM
and OFDM begins to widen when Nt takes value greater than 2 because CDD begins to
show its influence for Nt > 2 only. Authors expect that, in more realistic situation, the
performance difference between CDD and ACDD will widen which will make ACDD
more attractive diversity technique compared to CDD.

5 Simulation Results
We have simulated OCDM and OFDM with N = 256 chirps/subcarriers, oversampling
factor L = 4 and 10,000 OCDM/OFDM symbols to investigate their PAPR performance.
QPSK modulation is used to modulate the chirps/subcarriers in original OCDM/OFDM.
PTS, M-PTS and CCM-PTS are applied for PAPR reduction. SER performance of PTS,
M-PTS and CCM-PTS over AWGN and Rayleigh fading channel is also investigated.
For PTS operation, we have taken K = 4 phase factors {1, -1, j, -j} and data symbols
are divided into M = 4 sub-blocks using adjacent partition methods. QPSK modulation
is used to modulate the chirps/subcarriers in OCDM/OFDM with PTS.

Fig. 2. PAPR Performance of OCDM and OFDM Signals

Figure 2 shows the PAPR performance of OCDM and OFDM with PTS, M-PTS and
CCM-PTS. It is clear from Fig. 2 that OCDM and OFDM have same PAPR performance.
Reduction in PAPR provided by PTS, M-PTS and CCM-PTS is almost same and is
approximately equal to 3 dB at CCDF = 10–3 .
Figure 3 shows the SER performance of OCDM and OFDM with PTS, M-PTS and
CCM-PTS over AWGN channel. In Fig. 3, analytical expression for SER of QPSK, M-
PTS [7] and CCM-PTS [7] have been used to validate the simulation results. As shown in
244 M. K. Singh and A. Goel

Fig. 3. SER Performance of OCDM and OFDM Signals over AWGN Channel

Fig. 3, SER performance of OCDM with PTS, M-PTS and CCM-PTS is same as that of
OFDM with PTS, M-PTS and CCM-PTS respectively over AWGN channel. CCM-PTS
and M-PTS require almost 3.5 dB and 4 dB more bit energy (Eb ) compared to PTS at
SER = 10–5 .

Fig. 4. SER Performance of OCDM and OFDM Signals over Rayleigh Fading Channel

Figure 4 shows the SER performance of OCDM and OFDM with PTS, M-PTS and
CCM-PTS over a 3-tap multipath Rayleigh fading channel having path delays [0 µs,
0.5 µs, 1 µs] and average path gains [0dB, -5dB, -10dB]. An OCDM/OFDM system
with N = 256 chirps/subcarriers, cyclic prefix length L CP = 4 and 10000 OCDM/OFDM
symbols are considered. FDE is used for equalization. As shown in Fig. 4, SER perfor-
mance of OCDM slightly degrades in low SNR region whereas OCDM shows much
 Orthogonal Chirp Division Multiplexing 245

better performance than OFDM in high SNR region for all the three schemes. OCDM
with PTS, M-PTS and CCM-PTS outperforms for SNR > 28 dB, 32 dB and 31 dB
respectively. M-PTS and CCM-PTS require almost 4 dB and 3.5 dB more bit energy
(Eb ) compared to PTS at SER = 10–5 .

6 Conclusion
OCDM is a MCM technique in which N orthogonal chirps having same bandwidth are
multiplexed. OCDM provides maximum spectral efficiency for chirp spread spectrum
system but suffers from high PAPR. Various PAPR reduction, equalization and channel
estimation schemes presented in the literature for OFDM are applicable to the OCDM
also. With minor additional operations, OCDM system can be obtained from OFDM.
Computational complexity of OCDM is higher than the OFDM. OCDM outperforms
OFDM for wireless communication, coherent OFC, IM/DD-based short reach system,
UWA communication, baseband data communication and MIMO communication. Thus,
for high-speed communication OCDM is an attractive solution.
PAPR performance of OCDM is same as that of OFDM. Reduction in PAPR offered
by PTS, M-PTS and CCM-PTS is almost same for OCDM and OFDM both. Over AWGN
channel, SER performance of OCDM with PTS, M-PTS and CCM-PTS is identical to
the performance of OFDM with PTS, M-PTS and CCM-PTS respectively but CCM-PTS
and M-PTS require more bit energy (Eb ) compared to PTS to achieve the same SER.
Performance of OCDM slightly degrades in low SNR region over multipath Rayleigh
fading channel but it shows much better performance as compared to OFDM in high
SNR region. M-PTS and CCM-PTS discard the requirement of sending any SI to the
receiver at the cost of increased SER compared to PTS.

References
1. Ouyang, X., Zhao, J.: Orthogonal chirp division multiplexing. IEEE Trans. Commun. 64(9),
3946–3957 (2016)
2. Wu, Y., Zou, W.Y.: Orthogonal frequency division multiplexing: a multi-carrier modulation
scheme. IEEE Trans. Consum. Elec. 41(3), 392–399 (1995)
3. Ouyang, X., Zhao, J.: Orthogonal chirp division multiplexing for coherent optical fiber
communications. J. Lightwave Technol. 34(18), 4376–4386 (2016)
4. Muller, S.H., Huber, J.B.: OFDM with reduced peak-to-average power ratio by optimum
combination of partial transmit sequences. Electron. Lett. 33(5), 368–369 (1997)
5. Zhou, Y., Jiang, T.: A novel multi-point square mapping combined with PTS to reduce PAPR
of OFDM signals without side information. IEEE Trans. Broadcast. 55(4), 831–835 (2009)
6. Goel, A., Gupta, P., Agrawal, M. : Concentric circle mapping based PTS for PAPR reduction
in OFDM without side information. In: Proceedings of 2010 Sixth International Conference
on Wireless Communication and Sensor Networks, Allahabad, India, pp. 1–4, December 2010
7. Goel, A., Gupta, P., Agrawal, M.: SER analysis of PTS based techniques for PAPR reduction
in OFDM systems. Digit. Signal Proc. 23(1), 302–313, January 2013
8. Ouyang, X., Tall, G., Power, M., Townsend, P.: Experimental demonstration of 112 Gbit/s
orthogonal chirp-division multiplexing based on digital up-conversion for IM/DD sys-
tems with improved resilience to system impairments. In: Proceedings of 2018 European
Conference on Optical Communication, Rome, Italy, pp. 1–3, September 2018
246 M. K. Singh and A. Goel

9. Ouyang, X., Talli, G., Power, M., Townsend, P.: Orthogonal chirp division multiplexing for
IM/DD-based short-reach systems. Opt. Expr. 27(16), 23620–23632 (2019)
10. Bouvet, P.J., Auffret, Y., Aubry, C.: On the analysis of orthogonal chirp division multiplexing
for shallow water underwater acoustic communication. In: Proceedings of OCEANS 2017,
Aberdeen, UK, pp. 1–5, June 2017
11. Bai, Y., Bouvet, P.J.: Orthogonal chirp division multiplexing for underwater acoustic
communication. Sensors 18(11), 3815 (2018)
12. Zhu, P., Xu, X., Tu, X., Chen, Y., Tao, Y.: Anti-multipath orthogonal chirp division
multiplexing for underwater acoustic communication. IEEE Access 8, 13305–13314 (2020)
13. de M.B.A. Dib, L., Colen, G.R., de L. Filomeno, M., Ribeiro, M.V.: Orthogonal chirp division
multiplexing for baseband data communication systems. IEEE Syst. J. 14(2), 2164–2174, June
2020
14. Bomfin, R., Chafii, M., Fettweis, G.: Performance assessment of orthogonal chirp division
multiplexing in MIMO space time coding. In: Proceedings of 2019 IEEE 2nd 5G World
Forum, Dresden, Germany, pp. 220–225, October 2019
Machine Learning and Data Analytics
 COVID-19 Outbreak Estimation Approach
Using Hybrid Time Series Modelling

Soham Chakraborty, Sushruta Mishra(B) , and Hrudaya Kumar Tripathy

School of Computer Engineering, Kalinga Institute of Industrial Technology, Deemed to be

University, Bhubaneswar, India
{sushruta.mishrafcs,hktripathyfcs}@kiit.ac.in

Abstract. In the beginning of March 2020, coronavirus was claimed to be a

worldwide pandemic by the World Health Organization (WHO). In Wuhan, a
region in China, around December 2019, the Corona virus, also known as the
novel COVID-19 was first to arise and spread throughout the world within weeks.
Depending upon publicly available data-sets, for the COVID-19 outbreak, we have
developed a forecasting model with the use of hybridization of sequential and
time series modelling. In our work, we assessed the main elements to forecasting
the potential of COVID-19 outbreak throughout the globe. Inside the work, we
have analyzed several relevant algorithms like Long short-term memory (LSTM)
model (which is a sequential deep learning model), used to predict the tendency
of the pandemic, Auto-Regressive Integrated Moving Average (ARIMA) method,
used for analyzing and forecasting time series data, Prophet model an algorithm to
construct forecasting/predictive models for time series data. Based on our analysis
outcome proposed hybrid LSTM and ARIMA model outperformed other models
in forecasting the trend of the Corona Virus Outbreak.

Keywords: Epidemic transmission · Time series forecasting · Machine

learning · Corona-virus · LSTM networks · ARIMA

1 Introduction
It is believed that every infectious disease shows some design or pattern which must be
identified to forecast the outbreak of such diseases. We can define a different category
for the spread of disease such as concerning time i.e., seasonal change [1]. Generally, it
is found that their forwarding from person to person can be represented with the help
of the non-linear system. We have found that in the previous studies using the data-
driven method either linear or non-linear components are captured. Since they depend
on both linear and non-linear, hence, they are not able to get hold of the situation of the
transmittance of these infective health risks completely. Here we aim to build a intelligent
statistical model that can capture both the linear and non-linear components in the trend.
Purely Statistical techniques like moving average methods primarily depend on use cases
and these techniques are tedious to forecast the actual rate of transmissions. There is a
large array of a statistical and mathematical model which have been proposed for the

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 249–260, 2022.
https://doi.org/10.1007/978-3-031-23233-6_19
250 S. Chakraborty et al.

broadcast of the covid-19 [2, 3]. In a general situation, the accuracy and prediction of
such a model are low because the model is unable to fit samples in an accurate manner
[4]. So, a deep learning approach is presented to address transmission in real-time,
which beats the barrier of the statistical approach [5]. Since much research has been
done on the same domain so after some intensive research using different methods like
ARIMA, SARIMAX we found that the demographic-based trend curve has some non-
linear characteristics which can’t be determined by the existing 3 models. Our research
work is based on the hypothesis that LSTM model and ARIMA can together capture the
non-linearity under the forecasting data. The main objective of the work is to build a
logical forecast for developing an intuition over the trend of coronavirus outbreaks [6].
We have also compared our algorithm with similar methods like seasonal ARIMA,
which can be claimed to learn the non-linear representation of the time-series data. But
again due to the unpredictable non-seasonality occuring in the curve it fails to represent
or learn the actual underlying distribution, rather it tries to map it to the nearest possible
seasonal cycles. Thus, we conclude to move forward with a combination of LSTM ()for
capturing non-linearity and non-seasonality) and ARIMA (for capturing the linear trend
components).
The rest of the work has arranged as follows. In Sect. 2 we have given a detailed
overview of the dataset. In Sect. 3 we have discussed the basics of time-Series Modelling.
In Sect. 4 the LSTM model for time series models their detailing, definitions, and related
terms have been discussed. In Sect. 5, the ARIMA model detailing, definitions, and
related terms have discussed. In Sect. 6 We have proposed a model for forecasting
the coronavirus trend using the LSTM model and the ARIMA model and formed a
hybrid model from both the model. In Section VII We have discussed the detailed
implementation and the results of our research work.

2 Background

It Time-series interpretation is a statistical procedure for processing time-series data or

trend study. There are three types of data.

• Time series data - Observation results of variables taking values at varying times [7].
• Cross-sectional data - When info of single or multiple parameters gathered at the
similar instance.
• Pooled data - A mixture of time stamped and hybrid information.

2.1 LSTM Network for Modelling Time Series

Most samples in the real world are temporary in nature. The data collected at regular
intervals are time series (TS) records where every record is evenly distributed. TS analysis
is to predict future patterns for predefined past data set with basic characteristics [8–
10]. TS samples may be cracked down to seasonal error. If it does not depend on time
components such as trend and seasonal influence, called a stationary series. If TS data
has a trend of seasonal influence and changes over time the TS data is considered non-
stationary.
 COVID-19 Outbreak Estimation Approach 251

2.2 ARIMA Model

ARIMA stands for Autoregressive Integrated Moving Average. Different ARIMA

metrics are:

• Auto-regressive constituent [11] - This metric is represented as p. If p = 0, no autocor-

relation in the sequence is noted and if p = 1, it signifies presence of auto-correlation
in sequence.
• Integration - In the ARIMA, the integration is represented by d. If d = 0, it signifies
that the set is stable, and we don’t require to make the difference. If d = 1, it signifies
the set not to be stationary.
• Moving average constituent - It is denoted by q. In ARIMA, the moving average q =
1, signifies an error and time auto correlation is there.
• Decomposition - It relates to dividing the time series into trends, seasonal impacts,
and residual variability.

2.3 Seasonal ARIMA Model

• Seasonal Autoregressive Integrated Moving Average, SARIMA or Seasonal ARIMA,

is an extension of ARIMA that explicitly supports univariate time series data with a
seasonal component.
• It adds three new hyperparameters to specify the autoregression (AR), differencing
(I) and moving average (MA) for the seasonal component of the series, as well as an
additional parameter for the period of the seasonality [12].

3 Proposed Model

It Time-series interpretation is a statistical procedure for processing time-series data or

trend study. There are three types of data.

• Time series data - Observation results of variables taking values at varying times [13].
• Cross-sectional data - When info of single or multiple parameters gathered at the
similar instance.
• Pooled data - A mixture of time.

Because of the automatically extracting significant feature nature of the Deep learn-
ing method like recurrent neural network (RNN) it is observed to be a powerful technique
for the forecast [14]. In this the previous step’s activation is taken and is given is input
to the current time step and network self-connection. To eliminate the shortcoming long
short-term memory RNN structure was designed that regulates the data pass and mem-
ory cell in the current hidden layer [15–17]. The framework of the LSTM constitutes
4 gates i.e., input gate, forget gate, control gate, and output gate. The main job of the
252 S. Chakraborty et al.

input gate is to make a decision depending on which details can be transported to the
cell. The forget gate decides which details from the input of the previous memory are to
neglected. The work of the control gate is to control the updating of the cell. The output
layer is accountable for updating the hidden layer(ht-1) and also updating the output.
The LSM model is shown in Fig. 1.

Fig. 1. A basic representation of the LSTM model

In the proposed model we consider the dataset from Kaggle which contains a list of
daily Corona affected people, the daily number of deaths, in different provinces. These
data points are considered as observational points in multidimensional vector space.
These observational points are then fed into the ARIMA model and LSTM model sepa-
rately. The ARIMA model is fitted to the data to learn the moving average and the trends
in the data. Then the multidimensional vectors are fed into the LSTM algorithm, which
generates an embedding. The ARIMA output and its corresponding residual embedding
vector are concatenated in the final stage and down-projected by passing through a linear
function to generate the next data points sequentially. The proposed model flow diagram
has been shown in the following Fig. 2.
 COVID-19 Outbreak Estimation Approach 253

Fig. 2. A simple representation of the proposed model

4 Implementation and Results Discussion

In this section we will give a detailed exploratory data analysis overview of the complete
dataset. A sample of the used COVID-19 dataset has given in Table 1. We have provided
3 samples from the available data sets. The methods utilized here are predicated on data-
supervised methods which vary from precedent research work [18–21]. This forecasting
model can be used to pre-assume the outbreak trend of COVID-19 outbreak which can

Table 1. A Sample of India COVID-19 Dataset

S no. Period Time State Confirm Confirm Cure Life lost Confirm
cases abroad cases
1 Jan 30, Evening Kerala 1 0 0 0 1
2020
2 Jan 31, Evening Kerala 1 0 0 0 1
2020
3 Fev 1st , Evening Kerala 2 0 0 0 2
2020
254 S. Chakraborty et al.

help in controlling the outbreak. In the research, it is determined that norms and rules
formed by the regime is most likely to impact present outbreak.

Fig. 3. Distribution of COVID patients according to age in India

From the Fig. 3, age-wise distribution of the India population, we infer that out of
100% the most affected group is between the age of 20 to 29 with having approximately
25% of the total cases while the least cases have been found in the group of old people
having age above 80. The people between the ages group 30 to 39 occupy the second
position to easily get infected with 21% of the total cases. Next, comes the age group of
the 40 to 49 people having 16% of the total cases.
From the Fig. 4(a), we found that the total number of confirmed cases along with
the recovery is increasing exponentially, while the death rate is increasing slowly with
respect to time. From the graph it can be refer that the total number of cases by the end
of August has reached 3500000 which is very threatening. But one good thing can we
gather from the graph that the recovery rate is also increasing so we can hope for the
best. In Fig. 4(b). we compared the situation of few countries facing global pandemics
like China, the US, Italy, Spain, France, And India [22–24]. From the graph, we found
that the total number of cases in India and the US are increasing rapidly with time,
while countries like China, Italy and France, and Spain have somewhat controlled their
condition so the total number of cases is increasing but a bit slow.
 COVID-19 Outbreak Estimation Approach 255

Fig. 4. a. Distribution of COVID patients according to age in India b. Graph for comparison with
other countries.

4.1 Prediction Using LSTM Model

LSTM forms a recurrent network for storing previous outcomes into its storage units
and in training phase, it discovers to utilize its memory [25, 26]. From the Fig. 6(a) we
can observed that the proposed model performed well. It is able to accurately follow the
inclination. Also, our model predicted the upcoming next month’s trend. Hence, by the
end of September 2020, the cases in India could touch up to 5000000. Figure 5 shows
the growth rate of 15 states of India.
256 S. Chakraborty et al.

Fig. 5. Growth rate for 15 states

4.2 Prediction Using ARIMA Model

From the Fig. 6(b), it is clear that by using ARIMA model it performs well. It is
able to accurately follow the inclination. Also, our model predicted the upcoming next
month’s trend. Hence by the end of September 2020, the cases in India could touch up
to 57,00,000.

4.3 Prediction Using Hybrid Model

Figure 7 is the hybrid model by using ARIMA and LSTM network. By using the com-
bination of both models i.e. ARIMA and LSTM we produced a forecast of the total
cases for next 15 days. According to the model, the number of total cases will increase
exponentially, until or unless some strict step will not be taken. Of course, it is not the
exact prediction but based upon the trend.
 COVID-19 Outbreak Estimation Approach 257

Fig. 6. a. Prediction using LSTML model b. Prediction using ARIMA model

258 S. Chakraborty et al.

Fig. 7. Prediction using HYBRID model

5 Conclusion
In this research paper we have predicted the trend of coronavirus for next 15 days by
using the LSTM model and ARIMA model and their combined model to form a hybrid
model which worked better than the two. From the result of the hybrid model, it can be
seen that by the end of the month September the number of the cases in India will cross
5000000, which is a close approximation of the original estimate value.

References
1. World Health Organization. Naming the Coronavirus Disease (COVID-19) and the Virus
that Causes it. World Health Organization (2020). https://www.who.int/emergencies/dis
eases/novelcoronavirus-2019/technical-guidance/naming-the-coronavirus-disease-(covid-
209)-and-the-virus-thatcauses-it
2. Coronaviridae Study Group: The species Severe acute respiratory syndrome-related coro-
navirus: classifying 2019-nCoV and naming it SARS-CoV-2. Nat. Microbiol. 5, 536
(2020)
3. Lu, H., Stratton, C.W., Tang, Y.W.: Outbreak of pneumonia of unknown etiology in Wuhan
China: the mystery and the miracle. J. Med Virol. 92, 401–402 (2020)
4. Klompas, M.: Coronavirus disease 2019 (COVID-19): protecting hospitals from the invisible.
Ann. Intern. Med. 172, 619–620 (2020)
5. Roser, M., Ritchie, H., Ortiz-Ospina, E.: Coronavirus Disease (COVID-9)–Statistics and
Research. Our World Data (2020)
6. Rath, M., Mishra, S.: Security approaches in machine learning for satellite communication.
In: Hassanien, A.E., Darwish, A., El-Askary, H. (eds.) Machine Learning and Data Mining in
Aerospace Technology. SCI, vol. 836, pp. 189–204. Springer, Cham (2020). https://doi.org/
10.1007/978-3-030-20212-5_10
7. Dutta, A., Misra, C., Barik, R.K., Mishra, S.: Enhancing mist assisted cloud computing toward
secure and scalable architecture for smart healthcare. In: Advances in Communication and
Computational Technology, pp. 1515–1526s. Springer, Singapore (2021)
 COVID-19 Outbreak Estimation Approach 259

8. Mishra, S., Tripathy, H.K., Panda, A.R.: An improved and adaptive attribute selection
technique to optimize dengue fever prediction. Int. J. Eng. Technol. 7, 480–486 (2018)
9. Roy, S.N., Mishra, S., Yusof, S.M.: Emergence of drug discovery in machine learning.
In: Technical Advancements of Machine Learning in Healthcare, pp. 119–138). Springer,
Singapore (2021)
10. Salinas, D., Flunkert, V., Gasthaus, J., Januschowski, T.: DeepAR: Probabilistic forecasting
with autoregressive recurrent networks. Int. J. Forecast. (2019). https://doi.org/10.1016/j.ijf
orecast.2019.07.001
11. Oreshkin, B.N., Carpov, D., Chapados, N., Bengio, Y.: N-BEATS: neural basis expansion
analysis for interpretable time series forecasting. arXiv 2019, arXiv:1905.10437
12. Adhikari, R., Agrawal, R.K.: An introductory study on time series modeling and forecasting.
arXiv 2013, arXiv:1302.6613
13. Tripathy, H.K., Mishra, S., Thakkar, H.K., Rai, D.: CARE: a collision-aware mobile robot
navigation in grid environment using improved breadth first search. Comput. Electr. Eng. 94,
107327 (2021)
14. Sahoo, S., Mishra, S., Mishra, B.K.K., Mishra, M.: Analysis and implementation of artificial
bee colony optimization in constrained optimization problems. In: Handbook of Research on
Modeling, Analysis, and Application of Nature-Inspired Metaheuristic Algorithms, pp. 413–
432. IGI Global (2018)
15. Anastassopoulou, C., Russo, L., Tsakris, A., Siettos, C.: Data-based analysis, modelling and
forecasting of the COVID-19 outbreak. PLoS ONE 15, e0230405 (2020)
16. Kane, M.J., Price, N., Scotch, M., Rabinowitz, P.: Comparison of ARIMA and Random Forest
time series models for prediction of avian inflfluenza H5N1 outbreaks
17. Gelper, S., Fried, R., Croux, C.: Robust forecasting with exponential and Holt-Winters
smoothing. J. Forecast. 29, 285–300 (2010)
18. Mishra, S., Mallick, P.K., Tripathy, H.K., Jena, L., Chae, G.-S.: Stacked KNN with hard
voting predictive approach to assist hiring process in IT organizations. Int. J. Electr. Eng.
Educ., February 2021. https://doi.org/10.1177/0020720921989015
19. Harvey, A.C., Peters, S.: Estimation procedures for structural time series models. J. Forecast.
9, 89–108 (1990)
20. Mishra, S., Tadesse, Y., Dash, A., Jena, L., Ranjan, P.: Thyroid disorder analysis using random
forest classifier. In: Mishra, D., Buyya, R., Mohapatra, P., Patnaik, S. (eds.) Intelligent and
Cloud Computing. SIST, vol. 153, pp. 385–390. Springer, Singapore (2021). https://doi.org/
10.1007/978-981-15-6202-0_39
21. Chaudhury, P., Mishra, S., Tripathy, H.K., Kishore, B.: Enhancing the capabilities of stu-
dent result prediction system. In: Proceedings of the Second International Conference on
Information and Communication Technology for Competitive Strategies pp. 1–6, March 2016
22. Jena, L., Mishra, S., Nayak, S., Ranjan, P., Mishra, M.K.: Variable optimization in cervical
cancer data using particle swarm optimization. In: Mallick, P.K., Bhoi, A.K., Chae, G.-S.,
Kalita, K. (eds.) Advances in Electronics, Communication and Computing. LNEE, vol. 709,
pp. 147–153. Springer, Singapore (2021). https://doi.org/10.1007/978-981-15-8752-8_15
23. Mishra, S., Dash, A., Jena, L.: Use of deep learning for disease detection and diagnosis. In:
Bio-inspired Neurocomputing, pp. 181–201. Springer, Singapore (2021)
24. Madhu, G., et al.: Imperative dynamic routing between capsules network for malaria
classification. CMC-Comput. Mater. Continua 68(1), 903–919 (2021)
260 S. Chakraborty et al.

25. Chakraborty, S., Sahoo, K.S., Mishra, S., Islam, S.M.: AI Driven cough voice-based COVID
detection framework using spectrographic imaging: an improved technology. In: 2022 IEEE
7th International conference for Convergence in Technology (I2CT), pp. 1–7. IEEE, April
2022
26. Mishra, S., Thakkar, H.K., Singh, P., Sharma, G.: A decisive metaheuristic attribute selec-
tor enabled combined unsupervised-supervised model for chronic disease risk assessment.
Comput. Intell. Neurosci. (2022)
 Analysis of Depression, Anxiety, and Stress
Chaos Among Children and Adolescents Using
Machine Learning Algorithms

Satyananda Swain(B) and Manas Ranjan Patra

Department of Computer Science, Berhampur University, Berhampur, Odisha, India

[email protected], [email protected]

Abstract. In recent times, the pandemic seems to have a serious impact on the
mental health of people around the world across all age groups. This has been man-
ifested in the form of unstable mental conditions, depression, anxiety, stress, and
many other similar mental illnesses among individuals. In this study, we explore
the use of machine learning classification algorithms to detect and classify children
and adolescents with unstable mental conditions such as depression, stress, and
anxiety through the Covid-19 period based on demographic information and char-
acteristics using the DASS-21 Scale. Using a dataset of 2050 Chinese participants,
an attempt has been made to classify their depression, stress, and anxiety behavior
into different levels (Normal, Moderate, and Severe). The classification algorithms
considered are Support Vector Machines, KNN, Naive Bayes, and Decision Trees.
It is observed that the Support Vector Machine is the most effective method for
the classification of mental depression, anxiety, and stress conditions. The goal of
the study is to build a classification model for accurate categorization of unknown
samples into appropriate psychological chaos levels.

Keywords: Anxiety · DASS-21 · Decision tree · Depression · KNN · Naïve

bayes · Stress · Support vector machine

1 Introduction
The COVID-19 pandemic and frequent lockdowns led to fear and anxiety around the
world. This phenomenon has given rise to short-term and long-term psychological and
mental health implications like depression, anxiety, and stress both in children and
adolescents. The nature and scale of impact on individuals depending on factors such as
age, educational position, pre-existing mental health conditions, etc. This necessitates a
scientific study to analyze the impact and develop a strategy to deal with the psychosocial
and intellectual health issues of susceptible children and adolescents during the pandemic
and post-pandemic period.
The COVID-19 outbreak provided a scenario to study the correlation between stress-
ful life, their resulting psychological responses, and addictive behaviors. Through our
study, we demonstrate that an increased level of depression, anxiety, and stress caused
due to COVID-19 which can be established using the Depression, Anxiety, and Stress
Scale popularly called the DASS-21 Scale.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 261–272, 2022.
https://doi.org/10.1007/978-3-031-23233-6_20
262 S. Swain and M. R. Patra

1.1 Background

Depression (D)
Depression is a mood-based temperament illness that causes a sense of sadness and loss
of attention. In medical terms, depression can be specified as a major depressive disorder
that can affect how an individual feels, anticipates, and acts which can lead to a variety
of emotional and bodily hitches. Depression is not a common illness, and it should not
be taken casually.

Anxiety (A)
Anxiety is a feeling of strong, undue, tenacious apprehension and fear about unremark-
able situations. In a worrying situation, anxiety is usually obvious. When feelings become
extreme, all-consuming, and hinder daily living then a check on anxiety issues is a must
which indicates an underlying disease.

Stress (S)
Stress is a sensation of emotive or bodily strain. It’s a reaction of our body to a challenge or
claim. In a medical sense, post-traumatic stress chaos is a psychological health disorder
that is caused by a petrifying incident either undergoing it or perceiving it.
Depression, anxiety, and stress may require long-term treatment but most people
with these mental disorders feel better with medication, psychotherapy, or both. Early
detection is usually required for a proper diagnostic evaluation so that treatment can be
provided before severe mental distortions occur.

DASS-21 Details
DASS-21 has been an established psychometric screening tool with acceptable valid-
ity and reliability. DASS-21 is a set of three self-report scales designed to measure the
emotional states of depression, anxiety, and stress. Each of the three DASS-21 scales
contains seven items, divided into subscales with similar content. The depression scale
assesses dysphoria, hopelessness, devaluation of life, self-deprecation, lack of inter-
est/involvement, anhedonia, and inertia. The anxiety scale assesses autonomic arousal,
skeletal muscle effects, situational anxiety, and subjective experience of anxious affect.
The stress scale is sensitive to levels of chronic non-specific arousal. It assesses diffi-
culty in relaxing, nervous arousal, and being easily upset/agitated, irritable/over-reactive,
and impatient. Scores for depression, anxiety, and stress are calculated by summing the
scores for the relevant items [1].
The questions used for the DASS-21 test are described in Table 1 [1]. Table 2 presents
the cut-off scores for all three conventional severity labels as normal, moderate, and
severe which are considered to evaluate a person with these mental disorders[1].
 Analysis of Depression, Anxiety, and Stress Chaos Using ML Algorithms 263

Table 1. DASS-21 questıon set with question type specification (D-Depression, A-Anxiety, and
S-Stress Chaos). [Adopted from Reference-1]

DASS-21 Questions Type

I found it hard to wind down S
I was aware of the dryness of my mouth A
I couldn’t seem to experience any positive feelings at all D
I experienced breathing difficulty (e.g. Excessive rapid breathing, breathlessness in the A
absence of physical exertion)
I found it difficult to work up the initiative to do things D
I tended to overreact to situations S
I experienced trembling (e.g., in my hands) A
I felt that I was using a lot of nervous energy S
I was worried about situations in which I might panic and make a fool of myself A
I felt that I had nothing to look forward to D
I found myself getting agitated S
I found it difficult to relax S
I felt downhearted and blue D
I was intolerant of anything that kept me from getting on with what I was doing S
I felt I was close to panic A
I was unable to become enthusiastic about anything D
I felt I wasn’t worth much as a person D
I felt that I was rather touchy S
I was aware of the action of my heart in the absence of physical exertion (e.g., sense of A
heart rate increase, heart missing a beat)
I felt scared without any good reason A
I felt that life was meaningless D

Table 2. DASS-21 cut-off scores for severıty labels [Adopted from Reference-1]

Depression Anxiety Stress

Normal 0–9 0–7 0–14
Mild 10–13 8–9 15–18
Moderate 14–20 10–14 19–25
Severe 21–27 15–19 26–33
Extremely Severe 28 + 20 + 34 +
264 S. Swain and M. R. Patra

1.2 Motivation and Objective of the Work

The goal of this research is to analyze the level of mental chaos among Chinese children
and adolescents with unstable mental conditions during the pandemic situation and
develop a classification model to categorize them into normal, moderate, or severe mental
illness levels. It is intended to address some of the following research questions:

• How machine learning algorithms can be adopted for the classification of mental
chaos or illness based on psychological features?
• How DASS-21 features can be used to ensure the categorization of Chinese chil-
dren and adolescents with unstable mental conditions as per the Machine Learning
algorithm?
• Which Machine Learning algorithm can be employed that guarantees optimal
classification results?
• How to develop a classification framework for the accurate categorization of unknown
samples into appropriate psychological chaos levels?

2 Literature Review
S. Aleem, N. U. Huda, et al. discussed the idea of using machine learning methodologies
in the mental health domain to predict the probabilities of mental disorders and tried to
adopt potential treatment outcomes from the prediction. They proposed a general model
for depression detection using machine learning procedures [2].
S. A. H. M. Alim, M. G. Rabbani, et al. discussed a cross-sectional and descriptive
study to access depression, anxiety, and stress among medical students. As per their
studies, almost eighty-one percent of students either had depression, anxiety, or stress
alone or in combination. The combination of depression, anxiety, and stress was highest
among students [3].
L. Evans, K. Haeberlein, and A. Chang suggests that DASS-A and DASS-D provide
a good convergence validity and may be suitable for identifying adolescents with a
significant amount of anxiousness and depression [4].
V. Farnia, D. Afshari, et al. presented a study on the effect of substance abuse on
anxiety, depression, and stress in epileptic patients. They also showed that psychological
symptoms have an important role in the development of addiction among epileptic
patients [5].
S. H. Hamaideh, H. Al-Modallal, M.A. Tanash, and A. Hamdan Mansour prescribe
the prevalence and predictors of depression, anxiety, and stress among university students
during home quarantine due to COVID-19. They found a strong correlation between
depression, anxiety, and stress with demographic, health-related lifestyle variables [6].
E. Kakemam, E. Navvabi, and A.H. Albelbeisi proposed a study for examining the
psychometric properties of the Persian version of DASS-21 for nurses. They properly
evaluated and confirmed that DASS-21 is considered to be a valid and reliable tool for
evaluating depression, anxiety, and stress among Iranian nurses [7].
R. B. Khalil, R. Dagher, and M. Zarzour demonstrate the psychological impact of
the lockdown in Lebanon. Lockdown and other stressful life events are major factors in
developing depression, anxiety, and stress in individuals. The DASS-21 score was found
 Analysis of Depression, Anxiety, and Stress Chaos Using ML Algorithms 265

to be correlated with the impact of lockdown to develop such drastic mental disorders
[8].
I. Marijanović, M. Kraljević, T. Buhovac, et al. showed that the COVID-19 pandemic
affected depression, anxiety, and stress levels in oncological staff in BiH. They monitored
and provided support to staff members for their wellbeing and retention at a time of global
crisis in the healthcare sector [9].
Oli Ahmed, Rajib Ahmed Faisal, Sheikh M. A. H. Mostafa Alim, Tanima Sharker,
and Fatema Akhter Hiramoni demonstrate their research work to access mental health
status in various situations among Bangladeshi adults. They used the Bangla version of
the DASS-21 to access psychometric properties by utilizing both classical test theory
and item response theory. Their main idea is to propose a psychometrically sound tool
to access depression, anxiety, and stress in the non-clinical sample of Bangladesh [10].
P. Sharif-Esfahani, R. Hoteit, C. El Morr, and H. Tamim, provide the relationship
between fear of COVID-19 and depression, anxiety stress, and PTSD among Syrian
refugee parents in Canada. They found severe levels of depression, anxiety, and stress
in the participants [11].
R. S. Vaughan, E. J. Edwards, and T. E. MacIntyre provide initial support for use of
the DASS-21 as an operationalization of mental health symptomology in athletes. They
examine the internal consistency, factor structure, invariance, and convergent validity of
the DASS-21 scale in two athlete samples [12].

3 Methodology
3.1 Data Set Description

A self-reported, cross-sectional survey was conducted among Chinese children and ado-
lescents aged between 6 to 18 years. Participants responded to questionnaires contain-
ing DASS-21 information and internet usage characteristics like internet use frequency
before COVID and after COVID, degree of internet usage, etc. A total of 2050 par-
ticipants participated in the survey. The demographic measures include gender, age,
DOB, frequency of internet use, depression, anxiety, stress, and many other features.
The DASS-21 was used to screen mental issues from three different mental perspectives.
Accordingly, the DASS-21 scale defines the categorization principle of categorizing per-
sons into normal, moderate, and severely affected individuals with depression, anxiety,
or stress (Figs. 1, 2, 3).

3.2 Implementation

Step-1: The problem is described as a classification problem that attempts to classify

persons with depression, anxiety, or stress into normal, moderate, or severely affected
individuals based on the demographic data gathered through the DASS-21 test (Fig. 4).
Step-2: The original data sample pertains to school-going children and youngsters
from three different zones (central, east, and north) of China. In addition to details like
age, gender, education, parent details, income, etc., the DASS-21 scale questionnaires
were used to observe different symptoms of depression, anxiety, or stress. DASS-21 is
266 S. Swain and M. R. Patra

Fig. 1. Sample data set for depression

Fig. 2. Sample data set for anxiety

Fig. 3. Sample data set for stress

 Analysis of Depression, Anxiety, and Stress Chaos Using ML Algorithms 267

a scale with 21 basic questions with three subgroupings. Each response is recorded on
a 0–3 point scale. According to DASS-21, participants were classified into Normal (1),
Moderate(2), and Severe (3) predictable label groups.

Fig. 4. Schematic diagram for depression, anxiety, and stress behavior classification.

Step-3: Based on the available raw data, a comprehensive analysis has been done
to evaluate each feature with its respective values and an attempt has been made to
determine parameters that are used as classification criteria. The features that appeared
prominent, promising, and selected for learning were the 21 items of the DASS-21 scale.
Based on these features, classification labels were defined to categorize individuals into
different class labels.
Step-4: After feature selection, the data set is divided into the training set and testing
set in the ratio of 70:30.
Step-5: Various machine learning classification algorithms, viz., SVM, NB, KNN,
and DT are employed for training and testing.
268 S. Swain and M. R. Patra

Step-6: The accuracy, precision, recall, and F1-score are computed as performance
measures for the chosen classification algorithms.
Step-7: A comparative study of the classifiers is done to judge their relative
performance on the dataset.
Step-8: Subsequently, the K-Fold cross-validation procedure is also employed to
assess the performance of the machine learning algorithms on the dataset with K = 10.
Step-9: The classification model is then applied to unlabelled data samples to
categorize the samples into normal, moderate, or severely affected user categories.

4 Results and Discussion

Experiments were conducted using python programming language and a web-based

interactive computing platform called Jupyter notebook to implement different super-
vised machine learning algorithms for studying mental illness in children and adoles-
cents. The algorithms considered were SVM, NB, KNN, and DT. Classification matrices
were defined in terms of true and false positives and true and false negatives. Obtained
results depict the outcomes of ML algorithms in terms of accuracy, precision, recall, F1
scores, and support where:

Accuracy = (TP + TN)/(TP + FP + FN + TN) (1)

Precision = TP/(TP + FP) (2)

Recall = TP/(TP + FN) (3)

F1 Score = 2 ∗ (Recall ∗ Precision)/(Recall + Precision) (4)

Further, the support factor defines the number of samples with the true responses in
each class of target values.

4.1 Classification Results for Depression

Overall, the SVM algorithm demonstrates the highest accuracy of 0.997 as compared
to other classification algorithms. KNN and NB yield an accuracy of 0.964 and 0.938
respectively. The lowest performance was observed in the case of DT with a minimal
accuracy score of 0.896. The classification result for depression behavior is depicted in
Table 3. Figure 5 indicates the supremacy of SVM over other classifiers with an accuracy
score of 0.997.
 Analysis of Depression, Anxiety, and Stress Chaos Using ML Algorithms 269

Table 3. Depression behavior classification results

Algo. Name Accuracy Precision Recall F1Score Support

SVM 0.997 0.997 0.997 0.997 615
KNN 0.964 0.964 0.964 0.963 615
NB 0.938 0.95 0.938 0.941 615
DT 0.896 0.865 0.896 0.879 615

Comparision of Result for AC,PR,RC AND

F1 SCORE FOR SVM,KNN,NB AND DT
1
0.8
0.6
0.4
0.2
0
Accuracy Precision Recall F1 Score

SVM KNN NB DT

Fig. 5. Comparative analysis of classification results for depression

4.2 Classification Results for Anxiety

For anxiety also, the SVM algorithm demonstrates the highest accuracy of 1.0 as com-
pared to other classification algorithms. KNN and NB yield an accuracy of 0.94 and
0.912 respectively. The lowest performance was observed in the case of DT with a min-
imal accuracy score of 0.893. The classification result for anxiety behavior is depicted
in Table 4. Figure 6 also indicates the efficacy of SVM over other classifiers with an
accuracy score of 1.0 for anxiety chaos.

Table 4. Anxiety behavior classification results

Algo. Name Accuracy Precision Recall F1Score Support

SVM 1 1 1 1 615
KNN 0.94 0.937 0.94 0.935 615
NB 0.912 0.943 0.912 0.912 615
DT 0.893 0.846 0.893 0.868 615
270 S. Swain and M. R. Patra

Comparision of Result for AC,PR,RC AND F1

SCORE FOR SVM,KNN,NB AND DT

0.8

0.6

0.4

0.2

0
Accuracy Precision Recall F1 Score

SVM KNN NB DT

Fig. 6. Comparative analysis of classification results for anxiety

4.3 Classification Results for Stress

The SVM algorithm again demonstrates the highest accuracy of 0.997 as compared
to other classification algorithms. KNN and NB yield an accuracy of 0.959 and 0.938
respectively. The lowest performance was observed in the case of DT with a minimal
accuracy score of 0.925. The classification result for stress behavior is presented in
Table 5. Figure 7 indicates the supremacy of SVM yet again over other classifiers with
an accuracy score of 0.997 for stress chaos.

Table 5. Stress behavior classification results

Algo. Name Accuracy Precision Recall F1Score Support

SVM 0.997 0.997 0.997 0.997 615
KNN 0.959 0.954 0.959 0.953 615
NB 0.938 0.971 0.938 0.947 615
DT 0.925 0.921 0.925 0.923 615

4.4 Cross-Validation Results for Depression, Anxiety, and Stress Behaviour

The performance of the classifiers SVM, KNN, NB, and DT using 10-fold cross-
validation is presented in Table 6. Once again SVM emerges as the best performer con-
cerning accuracy in classification for all three aspects of our study, namely, depression,
anxiety, and stress stages.
 Analysis of Depression, Anxiety, and Stress Chaos Using ML Algorithms 271

Comparision of Result for AC,PR,RC AND F1

SCORE FOR SVM,KNN,NB AND DT
1
0.8
0.6
0.4
0.2
0
Accuracy Precision Recall F1 Score

SVM KNN NB DT

Fig. 7. Comparative analysis of classification results for Stress

Table 6. 10-Fold cross-validation results for depression, anxiety, and stress behaviour.

Depression Anxiety Stress

Algo Name Acc. Std. Dev. Acc. Std. Dev. Acc. Std. Dev.
(Mean) (Mean) (Mean)
SVM 0.99 0.004 0.986 0.009 0.987 0.005
KNN 0.943 0.02 0.901 0.016 0.915 0.025
NB 0.968 0.011 0.953 0.14 0.969 0.007
DT 0.959 0.017 0.967 0.01 0.955 0.014

5 Future Work
In future, we will investigate the impact of deep learning techniques like CNN and
RNN on depression, anxiety, and stress issues which are expected to provide better
classification results and for DASS-21 scaled features. Also, suitable ensemble models
will be looked at, to build effective and robust classification models.

6 Conclusions
The COVID-19 epidemic situation had a visible impact on the behavior of children and
adolescents because of the digital shift from offline to online modes of working. Our study
investigated Chinese children and adolescents for depression, anxiety, and stress issues
during the lockdown period. We have classified the individuals into normal, moderate,
and severely affected categories based on the DASS-21 scale which comprises of 21
questions related to the behavioral aspects of children and adolescents. The classification
272 S. Swain and M. R. Patra

algorithms considered were SVM, KNN, NB, and DT which provide interesting results
with classification accuracies of 0.997, 1.0, and 0.997 in the case of SVM for depression,
anxiety, and stress chaos respectively. Thus, one can use these classification models to
classify new behavioral data samples accurately for predicting depression, anxiety, or
stress disorders.

References
1. Lovibond, S.H., Lovibond, P.F.: Manual for the Depression Anxiety & Stress Scales. (2nd
Ed.) Psychology Foundation, Sydney (1995)
2. Aleem, S., Huda, N.U., Amin, R., Khalid, S., Alshamrani, S.S., Alshehri, A.: Machine learning
algorithms for depression: diagnosis, ınsights, and research directions. Electronics 11(7), 1111
(2022). https://doi.org/10.3390/electronics11071111
3. Alim, S.A.H.M., et al.: Assessment of depression, anxiety, and stress among first-year MBBS
students of a public medical college, Bangladesh. Bangladesh J. Psychiatry 29(1), 23–29
(2017). https://doi.org/10.3329/bjpsy.v29i1.32748
4. Evans, L., Haeberlein, K., Chang, A., Handal, P.: Convergent validity and preliminary cut-off
scores for the anxiety and depression subscales of the DASS-21 in US adolescents. Child
Psychiatry Hum. Dev. 52(4), 579–585 (2020). https://doi.org/10.1007/s10578-020-01050-0
5. Farnia, V., et al.: The effect of substance abuse on depression, anxiety, and stress (DASS-21)
in epileptic patients. Clinical Epidemiology and Global Health 9, 128–131 (2021)
6. Hamaideh, S.H., Al-Modallal, H., Tanash, M. A., Hamdan-Mansour, A.: Depression, anxiety,
and stress among undergraduate students during COVID-19 outbreak and “home quarantine”.
Nursing Open 9(2), 1423-1431 (2022)
7. Kakemam, E., Navvabi, E., Albelbeisi, A.H.: Psychometric properties of the persian version
of depression anxiety stress scale-21 Items (DASS-21) in a sample of health professionals: a
cross-sectional study. BMC Health Serv Res 22, 111 (2022). https://doi.org/10.1186/s12913-
022-07514-4
8. Khalil, R.B., et al.: The impact of lockdown and other stressors during the COVID-19 pan-
demic on depression and anxiety in a Lebanese opportunistic sample: an online cross-sectional
survey. Current Psychology, 1-11 (2022)
9. Marijanović, I., et al.: Use of the depression, anxiety, and stress scale (DASS-21) questionnaire
to assess levels of depression, anxiety, and stress in healthcare and administrative staff in
5 oncology institutions in Bosnia and Herzegovina during the 2020 COVID-19 pandemic.
Medical Science Monitor: Int. Medical J. Experimental Clinical Res. 27, e930812–e930821
(2021)
10. Ahmed, O., Faisal, R.A., Alim, S.M.A.H.M., Sharker, T., Hiramoni, F.A.: The psychome-
tric properties of the depression anxiety stress scale-21 (DASS-21) Bangla version. Acta
Psychologica 223, 103509 (2022). ISSN 0001-6918, https://doi.org/10.1016/j.actpsy.2022.
103509
11. Sharif-Esfahani, P., Hoteit, R., El Morr, C., & Tamim, H. (2022). Fear of COVID-19 and
Depression, Anxiety, Stress, and PTSD among Syrian Refugee Parents in Canada. Journal of
Migration and Health, 100081
12. Vaughan, R.S., Edwards, E.J., MacIntyre, T.E.: Mental health measurement in a post Covid-19
World: psychometric properties and invariance of the DASS-21 in athletes and non-athletes.
Front. Psychol. 11, 590559 (2020). https://doi.org/10.3389/fpsyg.2020.590559
 Sentiment Analysis from Student Feedbacks
Using Supervised Machine Learning Approaches

Preeti Routray1 , Chinmaya Kumar Swain1(B) ,

and Rakesh Chandra Balabantaray2
1 Department of CSE, SRM University, Amaravati, Andhra Pradesh, India
[email protected]
2 Department of CSE, IIIT Bhubaneswar, Bhubaneswar, Odisha, India

[email protected]

Abstract. Human life experiences a radical change due to the availability of all
kinds of information on the World Wide Web. This phenomenon makes the life
of each individual easier and better than before. This flooding of information also
contains huge amount of people’s opinions about their used products or services.
The massive flow of information in terms of opinions and views need to be analyzed
to know the efficacy of the products and/or services. The subjective expressions,
which are known as opinions describes the peoples’ feelings or sentiments towards
the service or product. The opinion can be either positive or negative, depending
on the evocation that the product or service has created. Similarly, in the realms
of education, initiatives are being taken by Govt. Organizations as well as private
organizations to know the effectiveness of the contemporary teaching-learning
process. The present undertaking of the investigation is to evaluate the academic
transaction by eliciting the students’ subjective expressions and analyzing the same
as their feedback. The scope of this paper is to analyze the student’s feedback,
which covers the teaching quality, style of delivery by the faculty, the compo-
nents of the course, and the overall ambiance in which the academic activities are
conducted. The present study investigates the applicability of different supervised
machine learning approaches for sentiment analysis from students’ subjective
feedback. The chosen supervised machine learning approaches are Naive Bayes
(NB), N-gram, Support Vector Machine (SVM), and Maximum Entropy (ME).
These four approaches are applied only for binary sentiment classifications and
to obtain a comparative analysis that would help to build a superior teaching and
learning process.

Keywords: Sentiment analysis · Supervised learning · Student feedback

1 Introduction
The sharing of textual information on the Internet has ushered in substantial changes
in human life [1]. The change is so pervasive that it is doubtful if any aspect of life is
untouched by the web. The paradigm shift in Information and Communication Technol-
ogy (ICT) has become a regular phenomenon nowadays [5, 7]. The World Wide Web is

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 273–288, 2022.
https://doi.org/10.1007/978-3-031-23233-6_21
274 P. Routray et al.

flooded with opinions or reviews shared in many forms. The huge collections of reviews
or opinions in textual and other forms make person’s life simple in many ways. One
of the fundamental benefit is to know about the quality of a product or service, which
is already being used by others. The opinion of other users play an important role for
drawing any conclusion about the quality of the product. Opinions about a product or
service is the subjective expression which describes the person’s sentiment towards that
product or service. The subjective expression can be positive or negative, that depends
on the quality of the product or service [17].
In most of the research on natural language processing aim at processing of textual
information, that is retrieval or mining of information [2–4]. Incidentally, the quantum
of research has overlooked the emotional response towards a product or service, which
is more spontaneous than rational thinking. Consequently, scant attention and research
have been directed towards sentiment analysis to improve the teaching-learning process.
Now the need is being felt that ICT recognizes the usefulness of the sentiment analysis,
also known as opinion mining, to make decisions based on the opinions of others [8].
The World Wide Web evolved the way people share their opinions or views. Now, a
person can post his perspective on web sites and express anything in the Internet forums,
groups, and blogs. These opinions or user-generated content can be used in many practi-
cal applications. If a person wants to buy a product, he is going to ask his friends to know
about the product, but due to the user- generated content on the web, he can get access to
those opinions for buying the products. This is quite helpful not only for the individuals
but also for the organizations. To obtain opinions about a product, organizations besides
relying on the surveys, they are extracting the feedbacks from the opinion polls and use
those feedbacks to improve the quality of the product or service [6].
The monitoring of opinions about a product or service on the web is a challenging
task as large amount of textual information available on the web and that are from diverse
sources. It is not effortless for a person to collect the opinions from all the sources and
extract the relevant opinions, summarize them and organize them into usable forms [9].
So there is a need for automated opinion discovery and summarization system [10],
which will help the user to make decisions faster [11]. This would immensely benefit
the user and the organizations from the inevitable chaos of information overload.
Sentiments can be described as opinions, judgments, emotions, or ideas prompted
[13]. The word sentiment and opinion are used interchangeably in technical writings.
But in the context of this paper, sentiments refer to the complexities of very subtle
and refined human feelings like tenderness, admiration and subdued resentment, etc.
Such subjective conditions of fellow beings hold immense value in the academic arena.
The trainer and the trainees, the teacher and the students form a sustained face-to-face
relationship and interaction between them based on the course held in a controlled
environment. Receiving only opinions from the students may not do optimal justice to
the feedback system. That is why the history of feedback is growing by the day and yet
remains chequered.
The unitary body of research on opinion mining or sentiment analysis is to extract
opinions (positive/negative) from the text documents. Against this backdrop, there is
little research on scattered sentiments of students in the teaching- learning process.
Possibly this could be due to the variations in the perceptive- cognitive levels of the
 Sentiment Analysis from Student Feedbacks 275

students in a class. The trainer or the instructor faces this problem of addressing the
collective student group with a standardized content delivery without any discrimination.
Today’s education system is a choice-based credit system that hardly leaves any room for
negotiation in the academic calendar. The present study has future promises of benefiting
the educational system in India. At this juncture, it is not elementary to ignore the element
of student’s subjective feedback as the students and teachers are groomed in a varied
cultural and linguistic matrix.
In general, an opinion is a message expressing a belief or skepticism or cynicism
or bias or exaltation about a wide range of products and services- from a sports car to
an android OS. This subjectivity, often deeply rooted and held with confidence but may
not be validated or substantiated by knowledge or proof [11]. In many cases, opinions
are shrouded in emotional articulations and embedded in long forum posts and blogs.
Obviously, it is not very easy for a human reader to retrieve relevant sources, extract the
desired information from the voluminous opinions. Consequently, it becomes a tiring
and hesitant struggle to process them and bring them in to usable form. Thus it is imper-
ative today that automated opinion discovery and summarization systems are urgently
developed [15], which would address the sentiment analysis of students hailing from
diverse cultural complexes and varied native languages in a country like India [12, 14].
The rest of the paper organized as follows. Section 2 discusses basic back- ground
and related works in this area of research. Section 3 discusses four machine learning
approaches used for sentiment analysis. The experimentation and results are discussed
in Sect. 4. Finally, Sect. 5 discusses about the conclusion and future direction of this
research.

2 Background
In the past few years, extensive studies were being carried out about sentiment analysis.
The reported studies classified the approaches in three categories like (a) lexicon-based
approach, (b) machine learning approach, and (c) hybrid approach.
The lexicon-based approach uses lexicon containing sentiment polarity of the words
to determine the sentiment of a given textual document. The sentiment lexicon can be
constructed either automatically or manually with list words and their associated senti-
ment polarity. In this context Hu and Liu [18] utilized WordNet to predict the semantic
orientation of a word. There is several domain-specific and general-purpose lexicons
(MPQA subjectivity lexicon, Word Counts database, Linguistic Inquiry, Harvard Gen-
eral Inquirer, etc.) have been constructed for the extraction of the semantic orientation
of words. However, this type of lexicon misses the domain-specific and contextual ori-
entation of a word. Rajput and Haider [19] used a lexicon-based approach to determine
the semantic polarity of the student’s feedback. They have used an academic domain-
specific lexicon for this purpose and also suggested that the domain-specific lexicon
gives the better result as compared to the general-purpose lexicon. Fernández et al. [21]
used the dependency parsing technique for reporting the semantic orientation of the
unstructured text. The proposed method used sentiment lexicon, which was constructed
using a semi-automatic polarity expansion algorithm.
Sentiment analysis using machine learning approaches relies on building the model
using training data set and evaluation of those models using test data set. The machine
276 P. Routray et al.

learning approaches can be classified as unsupervised and supervised methods based

on the availability of the labeling of data set. Qiang Ye et al. [29] used supervised
machine learning approaches for sentiment classification from the travel destinations
reviews. The same way Turney [20] proposed an unsupervised approach for sentiment
analysis. Altrabsheh et al. [38] analyzed the sentiment from student’s feedback using
three supervised learning approaches.
However, the hybrid approach is the combination of a lexicon-based approach and a
machine learning-based approach. Zhang et al. [10] used a hybrid approach for sentiment
analysis of Twitter data. They have used an opinion lexicon for labeling the data set and
used those to train the prediction model. Appel et al. [24] used a hybrid approach for the
sentiment analysis at the sentence level. They used SentiWordNet a sentiment lexicon
and fuzzy set theory to know the sentiment polarity of a sentence.
This study is intended to apply four supervised machine learning approaches for
sentiment analysis from students’ feedback collected during an academic session.

3 Machine Learning Approaches

Machine learning approaches proved to be very promising approaches for natural lan-
guage processing, specifically for sentiment analysis [29]. In this work we focus on
the four machine learning approaches (Naive Bayes, Support Vector Machine (SVM),
N-gram model, and Maximum Entropy (ME)) to classify the sentiment of the student’s
feedback. The overall process is represented in Fig. 1.

Fig. 1. Overview of supervised machine learning approaches

3.1 Naive Bayes Model

The popular Naive Bayes model is used for text classification and performs well in various
domains of study [25]. The Naive Bayes model uses stochastic model of document
generation during its operation. In order to predict the class of a new document Bayes
rules are used. In this paper, multinomial event model [26] is used to generate the
documents. Each document (d i ) is represented as a vector vi = (f i1 , f i2 ,….., f ik ), where
f ik represents the frequency of the k th word of the vocabulary set in ith document. The
 Sentiment Analysis from Student Feedbacks 277

vocabulary set V consists of all the words present in the documents under study, i. e. V =
w1 , W 2 ,…., wk . Given with the model parameter P (wk cj ) and class prior probabilities
P (cj ) with the assumption that the words are independent, the most likely class of the
document d i is computed as
|V |
c∗ (di ) = argmax P(cj )P(d |cj ) = argmax P(cj ) P(wk |cj )n(wk ,di ) (1)
j j k=1

where n(wk , d i ) represents the frequency of the word wk in document d i . The terms P
(wk cj ) and P (cj ) are estimated from the documents used in training, whose class labels
are known apriori. The terms are estimated as follows:

1 + di ∈cj n(wk , di )
P(wk |cj ) = |V | 
|V | + k=1 di ∈cj n(wk , di )
|cj |
P(cj ) = |C|
l=1 |cl |

3.2 Support Vector Machine (SVM)

It has been observed that SVM is one of the highly effective models for text categoriza-
tion, and it performs better than other traditional approaches [27]. SVM segregates the
positive and negative documents which are represented through vectors (υ).  The hyper-
planes represented by vectors segregated with maximum margin, as shown in Fig. 2. The
approach is to find the desired hyperplane, which can be made by solving a constraint
optimization problem. Let x i takes the value +1(−1), if the document d i belongs to
positive or negative class. The representative solution can be written as:


n
−
→
−
→
υ = λ∗i xi d i (2)
i=1

where λi ≥ 0 and the values of λ∗i can be obtained by solving a dual optimization problem.
The weight vector of the hyperplane can be constructed with a linear combination of di
using Eq. 2. The document vectors whose corresponding λi ≥ 0 are call support vectors
as those document vectors contribute to the calculation of υ . The performance of the
SVM mostly affected by the kernel it uses. The well-known kernel methods used in
SVM are linear, polynomial, and radial basis functions. In our experimentation, we used
a radial basis kernel (SVM RB) as it is mostly used and flexible also [36].

3.3 N-gram Based Character Language Model

There is another popular model called the N-gram character language model used in
natural language processing [30]. This model is derived from the popular N- gram
language models. This model considers character as the basic units instead of considering
words as the basic units. The probability distribution of a string s ∈  ∗ over a fixed
278 P. Routray et al.

Fig. 2. SVM as maximum margin classifier [29]

alphabets Σ can be presented as P (s). The chain rule of the N-gram character language
model can be defined as P(sc) = P(s).P(c|s) for a character c and string s. Based on
N-gram Markovian assumption that considers previous n − 1 characters, the chain rule
can be modified as

P(cn |sc1...cn−1 ) = P(cn |c1 . . . cn−1 )

Thus the maximum likelihood estimator for N-gram can be defined as,
C(sc)
PML (c|s) =  (3)
c C(sc)

where C(sc) represents the frequency of the sequence sc observed in the training data
and c C(sc) represents the number of single character extension of sc. In this work, the
used N-gram character language model is a generalized form of Witten-Bell smoothing
[30]. For the experiment purpose the default value of N = 8 is being used.

3.4 Maximum Entropy

There is an alternative technique that was proved to be effective for sentiment classifi-
cation is maximum entropy [31]. This technique sometimes performs better than Naive
Bayes approach for text classification [32]. The estimate of P (c d) takes the following
exponential form:
 
1 
PME (c|d ) = exp λi,c Fi,c (d , c) (4)
Z(d )
i

where Z(d) is a normalization function. The binary function F i,c is a feature/class func-
tion for the feature f i and class c. The term λi,c represents the feature-weight parameter
for the feature f i and class c. Based on the expected values of the feature/class func-
tions, the parameters are set to maximize the entropy. Due to the missing relationships
between features, ME model performs better for the case where conditional independent
assumptions are not met.
 Sentiment Analysis from Student Feedbacks 279

4 Experimentation and Results

This section discusses different steps being followed for finding the efficacy of the ML
approaches. These steps include the corpus description, which was used for the evaluation
purpose, pre-processing of the raw data, feature extraction from those data, evaluation
metrics used and finally the result analysis.

4.1 Data Corpus Description

For this work, we had collected the feedback of 1344 students from our institute web
portal. The feedbacks received consists of two parts, (a) unstructured textual feedback
and (b) a numerical scale of 1–5 about the quality of the teaching and learning process.
Table 1 presents the sample feedbacks of the students, and Table 2 reports the distribution
of differently labeled feedbacks. The textual data set is manually labeled {positive,
negative, neutral} by two experts with the linguistic background. The majority rule is
being used to label the data, and no majority case, the neutral labels are assigned. To verify
the reliability of the labels given by the experts and the students (based on numerical
values), we match their labels and found high percentage of agreement (81.09%). The
other measures like Krippendorff’s alpha [34] was 0.631 and Fleiss kappa [33] was
0.629 are seems to be more conservative [35]. The quantitative labels of 1–5 interpreted
as follows, 1–2 are considered as negative, 3 - as neutral, and 4–5 are considered as
positive feedbacks.

4.2 Preprocessing
As the students’ feedbacks contain unstructured text, so to extract the meaningful infor-
mation from those texts need preprocessing of the data set. The well-known preprocess-
ing steps which are applied to clean the data set may be the removal of spelling errors,
stop word removal, removal of punctuations, etc. In this work, we used the following
preprocessing steps to remove the noise from the data set.

Table 1. Sample of students feedbacks collected for our study

Sl. Feedbacks Sentiment Feedback

No. labels Score
1 This class is very boring and we are not feeling sleepy subject Negative 1
matter is not explained properly
2 faculty is very confusing while explaining the concept. He need to Negative 1
1st read properly and then teach the course
3 The subject is very interesting and she made the class lively. She Positive 4
explains the subject nicely
4 Excellent subject, taught very nicely. Lab experiment are very Positive 5
well organized to understand the subject well
(continued)
280 P. Routray et al.

Table 1. (continued)

Sl. Feedbacks Sentiment Feedback

No. labels Score
5 We have work out many assignments in theory and lab. That takes Neutral 3
so much time for this subject. We get less time to study other
subjects
6 Subject is toooo good. Taught in a wow manner. It was really Positive 5
interesting in the laboratory
7 Subject is really boring. Thumb down for the faculty I was so Negative 1
happppy in skipping the classes

Table 2. Distribution of students feedbacks collected for our study

Positive Negative Neutral

No. of instances 654 573 117

– Tokenization - It is the process of splitting the sentence into a list of words.

– Case conversion - The case conversion is used to convert the letters in to lower or
upper case. The step is followed the tokenization, to match the words in the training
data.
– Stop word removal - In NLP, the frequently used words in sentences like a preposition,
helping verbs, articles are termed as stop-words. The removal of stop-words will
reduce the number of words for further processing, and that reduces the space used
for their storage and improve the response time.
– Punctuation removal - The punctuation which would not carry any useful information
for the sentiment analysis is removed from the text.
– Removal of numbers - The chat like language contains numbers like 2 to represent ‘to’
or ‘too’, 4 to represent ‘for’, etc. So those numbers to be converted to their respective
English words for better understanding and clarity.
– Normalization - Here, we are using standard normalization techniques of NLP for
canonicalizing tokens so that matches occur despite superficial differences in the
character sequences of the tokens i.e. converting happy to happy.

4.3 Feature Extraction

After preprocessing of the dataset, features are extracted for the analysis of the sen-
timents. During this phase, the preprocessed text of the training and testing dataset is
converted into numerical feature vectors which are used in different machine learning
models. For the N-gram model, the sequence of letters, syllables, or words is extracted
for the experimentation.
 Sentiment Analysis from Student Feedbacks 281

Table 3. Useful information for evaluation metrics

Predicted Class
Positive Class Negative Class
Actual Class Positive Class True Positive (TP) False Negative (FN)
Negative Class False Positive (FP) True Negative (TN)

Table 4. Evaluation metrics for this work

Evaluation metrics
TP+TN
Accuracy = TP+FP+TN TP
+FN Precision = TP+FP
TP F − measure = 2 ∗ Precision∗Recall
Recall = TP+FN Precision+Recall

Table 5. Performance of different approaches with considering neutral class

Naive Bayes ME N-gram SVM RB

Accuracy 0.54 0.62 0.69 0.92
Precision 0.33 0.34 0.36 0.91
Recall 0.32 0.33 0.38 0.92
F-measure 0.32 0.33 0.36 0.91

4.4 Evaluation Metrics

In this paper, we used the following evaluation metrics to evaluate the performance of
different supervised machine learning approaches using the test data set. The evaluation
metrics are accuracy, precision, recall and F-measure [37, 38] and those are defined with
the help of Table 3 and Table 4.
From the reported results, the following observations can be inferred:

1. Support Vector Machine with radial basis kernel performed well in terms of accuracy,
precision, and recall.
2. The values of precision and recall are high in SVM RB as compared to the other
three models (Naive Bayes, Maximum Entropy, and N-gram).
3. Consideration of neutral classes for model building decreases the performance of
the classifiers.
4. In terms of accuracy Naive Bayes model performed worst among all the four models
under discussion.
282 P. Routray et al.

Table 6. Performance of different approaches without considering neutral class

Naive Bayes ME N-gram SVM RB

Accuracy 0.55 0.64 0.71 0.94
Precision 0.48 0.35 0.38 0.93
Recall 0.49 0.34 0.41 0.93
F-measure 0.48 0.34 0.39 0.93

Table 7. Number of feedbacks used in the training of the models

Round of experiments No. of feedbacks in each round of training

Positive Negative Total
1 25 25 50
2 50 50 100
3 75 75 150
4 100 100 200
5 150 150 300
6 200 200 400
7 250 250 500
8 300 300 600
9 350 350 700
10 426 382 808

4.5 Result Analysis

There are two parts of the experiment. In the first part to know the performance of the
four models, we used 10-fold cross-validation. The results are reported in Table 5 and
Table 6. Table 5 reports the performance by considering the neutral classes, and Table 6
reports the performance without considering the neutral classes.
The second part of the experiment conducted with 3-fold cross-validation [39]. In
this case, we are not considering the data points of the neutral class. The 654 positive and
573 negative feedbacks are considered to make 3-fold cross validation. The total number
of positive and negative data points are partitioned randomly into three folds, each with
218 positive and 191 negative feedbacks. During the experiment at each round, two-fold
was used as a training data set, and the remaining one fold was used as the testing data
set (shown in Table 7). This set of experiments was conducted to know the performance
of the models with an increasing number of training data set. All the cases, the number
of testing data set was 409 (218 positive and 191 negative).
The accuracy of the four approaches are reported in Table 8, and in Fig. 3. From
the result, it is confirmed that the SVM with radial basis kernel performed best out of
 Sentiment Analysis from Student Feedbacks 283

Table 8. Accuracy of different models

Round of experiments Training dataset size Accuracy of the models

Naive ME N-gram SVM p-value
Bayes RB
1 50 0.52 0.57 0.65 0.71 0.0000
2 100 0.54 0.58 0.71 0.76 0.0000
3 150 0.69 0.71 0.76 0.81 0.4352
4 200 0.72 0.73 0.79 0.84 0.2159
5 300 0.74 0.74 0.81 0.85 0.3178
6 400 0.76 0.77 0.82 0.86 0.2362
7 500 0.77 0.79 0.83 0.88 0.1994
8 600 0.80 0.80 0.84 0.89 0.4126
9 700 0.80 0.81 0.85 0.91 0.3024
10 808 0.80 0.81 0.85 0.93 0.4561

1.2
NB N-gram
ME SVM RB
1
Accuracy

0.8

0.6

0.4

0.2
0 100 200 300 400 500 600 700 800 900
Number of training data
Fig. 3. Accuracy of four machine learning models

four approaches. The difference between accuracy among the four approaches was very
significant (p < 0.01) when the number of training instances was less than equal to 100.
We found from the experiment that the increased size of the training data set improved
the accuracy of all the models. The accuracy of all the models reaches more than 80%
when the number of training data set contains more than 700 reviews. The precision of
all the models is presented in Table 9, and Fig. 4. Table 10 and Fig. 5 reports the recall
of all the four models.
284 P. Routray et al.

Table 9. Precision of different models

Round of experiments Training data set size Precision of the models

Naive ME N-gram SVM p-value
Bayes RB
1 50 0.50 0.53 0.67 0.73 0.0000
2 100 0.53 0.56 0.70 0.77 0.0021
3 150 0.54 0.61 0.74 0.79 0.3525
4 200 0.66 0.63 0.76 0.81 0.3159
5 300 0.71 0.67 0.79 0.83 0.4171
6 400 0.73 0.72 0.81 0.87 0.3252
7 500 0.74 0.74 0.83 0.86 0.2594
8 600 0.78 0.78 0.85 0.89 0.3912
9 700 0.81 0.80 0.84 0.90 0.2246
10 808 0.83 0.84 0.85 0.92 0.1987

1.2
NB N-gram
ME SVM RB
1
Precision

0.8

0.6

0.4

0.2
0 100 200 300 400 500 600 700 800 900
Number of training data
Fig. 4. Precision of four machine learning models
 Sentiment Analysis from Student Feedbacks 285

Table 10. Recall of different models

Round of exper- iments Training data set size Recall of the models
Naive ME N-gram SVM p-value
Bayes RB
1 50 0.49 0.52 0.64 0.72 0.0000
2 100 0.52 0.55 0.67 0.76 0.0030
3 150 0.53 0.60 0.70 0.79 0.2577
4 200 0.65 0.61 0.73 0.80 0.1582
5 300 0.69 0.66 0.77 0.82 0.3271
6 400 0.72 0.70 0.79 0.86 0.4242
7 500 0.75 0.73 0.81 0.87 0.2371
8 600 0.77 0.77 0.82 0.88 0.3253
9 700 0.80 0.79 0.83 0.91 0.2361
10 808 0.82 0.82 0.83 0.92 0.1307

1.2
NB N-gram
ME SVM RB
1

0.8
Recall

0.6

0.4

0.2
0 100 200 300 400 500 600 700 800 900
Number of training data
Fig. 5. Recall of four machine learning models

5 Conclusion and Future Work

The machine learning approaches have proved to be quite successful in identifying

sentiments from the unstructured text documents. In the current investigation, we have
used four supervised machine learning approaches of Naive Bayes, Maximum Entropy,
N-gram, and SVM models to analyze the sentiments from the student’s feedback. Here
all the four models report more than 80% of classification accuracy. As the number of
286 P. Routray et al.

entities of the training set progressively increases, the accuracy also increases. For the
fewer entities of the training dataset (50 or 100), the accuracy of the four algorithms is
significant, whereas, for many entities, there is no significant change in the accuracy.
Pre-processing of the data is indispensable for the optimization of results. Such
pre-processing would include named entity recognition, better tokenization, and pars-
ing. Named entity recognition would remove the possibility of names (Alex, Bob etc.)
and place names such as Delhi and Mumbai occurring in the output lexicon, which
would certainly not bear any sentiment. Tokenizing could be improved to remove noisy
expressions, such as numbers, dates, and words with unreadable/special characters and
symbols.
In this work, there is a need to add more data in the corpus of subjective feedback
from the students. The more the training data set, the more accurate is the result, which
improves the performance of the model. It is noteworthy to mention that the sentiment
analysis could also be done by implementing other machine learning algorithms like
Decision Tree, Artificial Neural Network, etc. There may be some more extensions
of this work, like combining the machine learning approach with a dictionary-based
approach. With growing computer literacy and connectivity, the present investigation
would gain momentum in the future.

References
1. Appel, G., Grewal, L., Hadi, R., Stephen, A.T.: The future of social media in marketing. J.
Acad. Mark. Sci. 48(1), 79–95 (2019). https://doi.org/10.1007/s11747-019-00695-1
2. Brody, S., Elhadad, N.: An unsupervised aspect-sentiment model for online reviews. In:
Human Language Technologies: The 2010 Annual Conference of the North American Chapter
of the Association for Computational Linguistics (HLT 2010). Association for Computational
Linguistics, Stroudsburg, PA, USA (2010)
3. Das, S.R., Chen, M.Y.: Yahoo! For Amazon: sentiment extraction from small talk on the web.
Manage. Sci. 53(9), 1375–1388 (2007)
4. Routray, P., Swain, C.K., Mishra, S.P.: A survey on sentiment analysis. Int. J. Comput. Appl.
76(10), 1–8 (2013)
5. Chinmaya Kumar Swain and Deepak B. Phatak “Evaluation model for the teacher’s training
program under National mission on education through ICT ”, International Conference on
e-Education, e-Business, e-Management and E-Learning, 7–9th Jan. 2011,
6. Drus, Z., Khalid, H.: Sentiment analysis in social media and its application: systematic
literature review. Procedia Comput. Sci. 161, 707–714 (2019)
7. Swain, C.K., Patel, A., Routray, P.: Model to evaluate the teaching enhancement of teacher’s
under the effect of teacher’s training program through ICT. In: International Conference on
e-Education, e-Business, e-Management and E- Learning, 7–9th January 2011
8. Boiy, E., Hens, P., Deschacht, K., Moens, M.-F.: Automatic sentiment analysis of on-line text.
In: Proceedings of the 11th International Conference on Electronic Publishing, pp. 349–360,
Vienna (2007)
9. Pang, B., Lee, L.: Opinion mining and sentiment analysis. Found. Trends Inf. Retr. 2(1–2),
1–135 (2008)
10. Pang, B., Lee, L.: A sentimental education: sentiment analysis using subjectivity summariza-
tion based on minimum cuts. In: Proceeding of ACL 2004, pp. 271–278 (2004)
 Sentiment Analysis from Student Feedbacks 287

11. Jurafsky, D., Martin, J.H.: Speech and Language Processing: An Introduction to Natural Lan-
guage Processing, Computational Linguistics, and Speech Recognition. Pearson Education
Inc, New Jersey (2009)
12. Santi, P.K., Mohanty, S., DasAdhikary, K.P., Swain, K.C.: Design and implementation of
nouns in OriNet: based on the semantic word concept. Arch. Control Sci. 15(3), 429–436
(2005)
13. Liu, B.: Sentiment Analysis and Opinion Mining, Morgan & Claypool Publishers, May 2012
14. Swain, K.C., Santi, P.K., Mohanty, S.: Morphological analyser based on finite state transducer:
a case study for Oriya language. Archives Control Sci. 15(3), 451–460 (2005)
15. Prabowo, R., Thelwall, M.: Sentiment analysis: a combined ap- proach. J. Informet. 3(2),
143–157 (2009)
16. Liu, B.: Sentiment Analysis: A Multi-Faceted Problem. IEEE Intell. Syst. 34(4) (2010)
17. Panigrahi, C.R., Panda, B., Pati, B.: Exploratory data analysis and sentiment analysis of drug
reviews. Computación y Sistemas, 26(3) (2022)
18. Hu, M., Liu, B.: Mining and summarizing customer reviews. In: Proceedings of the Tenth
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 168–
177. ACM (2004)
19. Rajput, Q., Haider, S., Ghani, S.: Lexicon-based sentiment analysis of teachers evaluation.
Appl. Comput. Intell. Soft Comput. 2016, 1 (2016)
20. Turney, P.D.: Thumbs up or thumbs down?: semantic orientation applied to unsupervised
classification of reviews. In: Proceedings of the 40th Annual Meeting on Association for
Computational Linguistics. Association for Computational Linguistics, pp. 417–424 (2002)
21. Fernández-Gavilanes, M., Álvarez-López, T., Juncal-Martınez, J., Costa- Montenegro, E.,
González-Castaño, F.J.: Unsupervised method for sentiment analysis in online texts. Expert
Syst. Appl. 58, 57–75 (2016)
22. Altrabsheh, N., Cocea, M., Fallahkhair, S.: Learning sentiment from students feedback for
real-time interventions in classrooms. In: Bouchachia, A. (ed.) Adaptive and Intelligent Sys-
tems, vol. 8779, pp. 40–49. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-112
98-5_5
23. Zhang, L., Ghosh, R., Dekhil, M., Hsu, M., Liu, B.: Combining lexicon-based and learning-
based methods for twitter sentiment analysis. Technical Report (2011)
24. Appel, O., Chiclana, F., Carter, J., Fujita, H.: A hybrid approach to the sentiment analysis
problem at the sentence level. Knowl.-Based Syst. 108, 110–124 (2016)
25. Domingos, P., Pazzani, M.J.: On the optimality of the simple bayesian classifier under zero-one
loss. Mach. Learn. 29(2–3), 103–130 (1997)
26. McCallum, A., and Nigam, K.: A comparison of event models for Naive Bayes text classifi-
cation. In: AAAI-98 Workshop on Learning for Text Categorization, pp. 41–48. AAAI Press
(1998)
27. Joachims, T.: Text categorization with Support Vector Machines: learning with many relevant
features. In: Nédellec, C., Rouveirol, C. (eds.) ECML 1998. LNCS, vol. 1398, pp. 137–142.
Springer, Heidelberg (1998). https://doi.org/10.1007/BFb0026683
28. RushdiSaleh, M., MartnValdivia, M.T., MontejoRez, A., UreaLpez, L.A.: Experiments with
SVM to classify opinions in different domains. Expert Syst. Appl. 38, 14799–14804 (2011)
29. Ye, Q., Zhang, Z., Law, R.: Sentiment classification of online reviews to travel destinations
by supervised machine learning approaches. Expert Syst. Appl. 36, 6527–6535 (2009)
30. Carpenter, B.: Scaling high-order character language models to gigabytes. In: Proceedings of
the Workshop on Software (Software 2005), Stroudsburg, PA, USA, 86–99. Association for
Computational Linguistics
31. Berger, A.L., Della Pietra, S.A., Della Pietra, V.J.: A maximum entropy approach to natural
language processing. Comput. Linguist. 22(1), 39–71 (1996)
288 P. Routray et al.

32. Nigam, K., Lafferty, J., McCallum, A.: Using maximum entropy for text classification. In:
Proceedings of the IJCAI 1999, pp. 61–67 (1999)
33. Fleiss, J.L.: Measuring nominal scale agreement among many raters. Psychol. Bull. 76(5),
378 (1971)
34. Hayes, A.F., Krippendorff, K.: Answering the call for a standard reliability measure for coding
data. Commun. Methods Meas. 1(1), 77–89 (2007)
35. Lombard, M., Snyder-Duch, J., Bracken, C.C.: Practical resources for assessing and reporting
intercoder reliability in content analysis research projects (2004). http://astro.temple.edu/lom
bard/reliability/
36. Chang, Y.W., Hsieh, C.J., Chang, K.W., Ringgaard, M., Lin, C.J.: Training and testing low-
degree polynomial data mappings via linear svm. J. Mach. Learn. Res. 11, 1471–1490 (2010)
37. Nasim, Z., Rajput, Q., Haider, S.: Sentiment analysis of student feedback using machine
learning and lexicon based approaches. In: 2017 International Conference on Research and
Innovation in Information Systems (ICRIIS), Langkawi, pp. 1–6 (2017)
38. Bouchachia, A. (ed.): ICAIS 2014. LNCS (LNAI), vol. 8779. Springer, Cham (2014). https://
doi.org/10.1007/978-3-319-11298-5
39. Kohavi, R.: A study of cross-validation and bootstrap for accuracy estimation and model
selection. In: Proceedings of the 14th International Joint Conference on Artificial Intelligence
- Volume 2 (IJCAI 1995), vol. 2, pp. 1137–1143. Morgan Kaufmann Publishers Inc., San
Francisco (1995)
 Machine Learning Algorithms for Diabetes
Prediction

Sudipta Priyadarshinee and Madhumita Panda(B)

Department of Computer Science, G. M. University, Sambalpur, Odisha, India

[email protected]

Abstract. Diabetes has emerged as one of the most deadly and prevalent illnesses
in the modern society, not just in India but also everywhere else. Diabetes now
impacts individuals of every age and is associated with lifestyle, genetics, stress,
and ageing. Different types of machine learning approaches are now applied to
forecast diabetes and also the disorders brought on by this disease. In this study
we have used five machine learning classifiers such that Extra Tree (ET),Decision
Tree (DT),Random Forest (RF), K-Nearest Neighbour (KNN) and Passive Aggres-
sive Classifier (PAC) for diabetes mellitus prediction. The experimental findings
demonstrate that Random Forest and Extra Tree have the lowest error rates with
the highest accuracy (81.16%).

Keywords: Diabetes · Machine learning · Accuracy · Classifiers

1 Introduction
A subfield of Artificial Intelligence (AI) known as Machine Learning (ML)enables
programmes to forecast outcomes more accurately even when they weren’t explicitly
intended to do so. In order to forecast future output values, algorithms of machine learn-
ing utilize the past data as its input. These algorithms employ mathematical approaches
that are highly helpful in assessing a lot of data and making recommendations for actions
based on these data. Machine learning is now being utilized in many facets of medical
health. Numerous researchers [1, 2] are now using algorithms of ML to predict and
manage a variety of diseases. In order to take the required steps to prevent diabetes,
machine learning algorithms are being applied to investigate their potential for diabetes
prediction. These algorithms may be grouped basically into 3 types: Supervised, Unsu-
pervised and Reinforcement learning [3].Commonly referred to as Diabetes Mellitus
(DM) by medical experts, diabetes disease is a collection of metabolic illnesses where
an individual has excessive blood sugar due to insufficient insulin secretion, improper
insulin cell response, or a combination of both. [4]. Diabetes illness is separated into 2
groups namely type l and type 2. The main distinction between these two is that person
with type 1 diabetes can not make insulin while those with type 2 diabetes produce a
small amount of insulin that is not effective enough. Urinating frequently, experienc-
ing frequent hunger and thirst, feeling exhausted, and having impaired vision are the
main symptoms of people with type 1 or type 2 diabetes [5]. The early this disease is

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 289–298, 2022.
https://doi.org/10.1007/978-3-031-23233-6_22
290 S. Priyadarshinee and M. Panda

detected, better chances are there to cure the patient. As forecast by International Dia-
betes Federation in 2017 [6] that there have been 425 million diabetics worldwide, and
that population will rise to 625 million by 2045 [7].
There are three more distinct parts to this work, and they are as follows: Sect. 2 of this
paper, outlines the research and studies that back up the findings on diabetes. Section 3,
outlines the suggested framework and the several different machine learning classifiers
utilized in the forecast of diabetes. The outcomes of the experiments are displayed in
Sect. 4, while a summary and analysis of experiments are presented in Sect. 5.

2 Related Works
This section presents the many methods that has been used to foretell diabetes issues.
In this study [8], a Decision Support System that employed the Decision Stump
base classifier and the Ada Boost algorithm for classification has been proposed. As
foundation classifiers for the SVM, Naive Bayes,Ada Boost and Decision Tree were also
used to verify accuracy. The experimental results showed that when using a Decision
Stump as the base classifier, Ada Boost algorithm achieved an accuracy of 80.72%.
The following work [9] was to build a system which can more precisely judge a
condition of diabetic patient. Naive Bayes, Decision Tree, ANN, and SVM methods
were employed as the foundation for categorization approaches in model construction.
ANN, Decision Tree, Naive Bayes, and SVM approaches offered precisions of 70%,
85%, 77%, and 77.3% respectively.
For the 2-way categorization of diabetes, the authors [10] utilized Convolutional
Neural Network (CNN). With a 75/25 training/testing configuration and normalisation
set at 550 for input data, the authors achieved an accuracy of 77.98% and obtained a
kappa coefficient of 0.549.
Random Forest, Nave Bayes and J48 decision tree were the three ML models that
were used in this study [11]. The experimental findings concluded that Nave Bayes
surpassed both the J48 decision tree and random forest for accuracy in the 3-factor
and 5-factor feature-selected data subsets. With an accuracy of 79.13% compared to
79.57%, the Naive Bayes on the 3-factor data subset outperformed the Random Forest
on the entire dataset.
In order to identify diabetes in its early stages, the author of this study [12] utilized
three algorithms of machine learning:SVM, Decision Tree, and Naive Bayes. Observed
outcomes indicated that Naive Bayes provided the best results with the maximum
accuracy (76.30%).
The authors [13] have used six well-known algorithms of machine learning such
that SVM, KNN, LR, DT, RF, and NB for predicting diabetes disease. According to the
experimental findings, SVM and KNN provided the greatest accuracy of 77%.

3 Proposed Methodology
The proposed framework is shown in Fig. 1.In our proposed methodology, data are
first pre-processed for missing values, then after that the standard scalar method is used
to standardize them. For prediction of diabetes illness, we have used five classification
 Machine Learning Algorithms for Diabetes Prediction 291

Start

Load the Diabetes

Dataset

Prepossessing Data

Split the Dataset in to

Training and Testing part

Applying Machine Learning

Algorithms
Random Forest, Extra Tree,Passive
Aggressive Classifier,K-Nearest
Neighbour,Decision Tree

Performance Evaluation on Various

Measures

Comparative analysis based on

accuracy

Stop

Fig. 1. Proposed framework

292 S. Priyadarshinee and M. Panda

approaches like Extra Tree, Random Forest, Decision Tree, KNN and Passive Aggressive
Classifier.For this study, 20% of the data are utilised for testing, while 80% of the data
are used for training. After applying machine learning algorithms on the dataset, the
accuracy rate was computed. The main objective was to find an algorithm that could
classify the given dataset most accurately.

3.1 Description of Dataset

The Kaggle [14] diabetes dataset has been utilized for the experiment. This dataset
consists of 768 records, with 500 negative classes referring to non-diabetes and 268
positive classes referring to diabetes patients, making up 34.9% and 65.1%, respectively,
of the total dataset. It has one result class and eight vital characteristics, which are shown
in the Table 1 and also graphically represented in Fig. 2.

Table 1. Dataset specification

Sr No Attributes Description
1 Pregnancies Count of pregnancies
2 Glucose 2 h plasma glucose concentration in an oral glucose
tolerance test
3 Blood Pressure Diastolic blood pressure (mm Hg)
4 Skin Thickness Thickness of the triceps skin fold (mm)
5 Insulin 2-h serum insulin (mu U/ml)
6 BMI Body mass index (weight in kg/(height in m)^2)
7 Diabetes Pedigree Function Diabetes pedigree function
8 Age Age (years)
9 Outcome category variable (1 or 0)

3.2 Classification Algorithms

A. Random Forest (RF)

It’s a form of machine learning called supervised learning, and it’s used to
do things like classifying data or making predictions [15]. The decision trees are
built using data samples, and predictions are obtained from each one. The best
answer is then chosen utilising this algorithm’s voting mechanism. Given that it is
the most common algorithm, it has simplicity and diversity. This method creates
several decision trees, which it then blends. It provides more reliable and accurate
prediction (Fig. 3).
 Machine Learning Algorithms for Diabetes Prediction 293

Fig. 2. Distribution of the dataset

Fig. 3. Random forest

B. Extra Tree (ET)

Extra Tree is a method for enhancing accuracy and computing efficiency by
combining bagging classifiers with traditional tree-based approaches [16]. The pri-
mary distinctions from other tree-based algorithms are able to separate the node
by randomly selecting cut-points and building the trees utilising all of the learning
samples (Fig. 4).

C. Decision Tree
It works well for both regression and classification. It functions as a classifier
with a framework of tree structure having internal nodes signifying elements of
dataset, branches signifying classification rules, and every leaf node signifying the
classification outcomes. Two nodes,the Decision Node which are used to create
decision and the Leaf Node which are the results of the decision make up a decision
tree.[17] (Fig. 5).
294 S. Priyadarshinee and M. Panda

Fig. 4. Extra tree

Fig. 5. Decision tree

D. Passive-Aggressive Classifier
Large-scale learning usually uses passive-aggressive algorithms. As opposed to
batch learning, when the entire training dataset is applied at once, online machine
learning techniques employ sequential input data and update the model of machine
learning one step at a time. When there is a huge amount of data available, this is
really helpful [18].
E. K-Nearest Neighbors (KNN)
A supervised machine approach called K-nearest neighbours is used to cat-
egorise label datasets. This algorithm finds neighbours for a specific data point.
Additionally, predictions for the labels of unknown data points can be generated
using information from these neighbours [19] (Fig. 6).
 Machine Learning Algorithms for Diabetes Prediction 295

Fig. 6. k-Nearest neighbors

4 Results and Discussion

The programming language Python was employed for doing the comparative study in
finding the performance matrices of the five algorithms as mentioned namely Decision
Tree (DT), Random Forest (RF), Extra Tree (ET), K-Nearest Neighbour (KNN) and
Passive Aggressive Classifier (PAC)on the data set of Kaggle [14]. Table 2 displays the
hyper parameter values utilised in this investigation of algorithms.

Table 2. Hyper parameter values of machine learning algorithms

Algorithms Hyperparameter Values

Random Forest n_estimators = 500, max_depth = 5, random_state = 123
Extra Trees n_estimators = 270,max_depth = 9, random_state = 42
Decision Tree random_state = 123,criterion = ‘entropy’,max_depth = 7
Passive Aggressive Classifier max_iter = 300,random_state = 1,fit_intercept = False
K-Nearest Neighbour n_neighbors = 17,weights = ‘distance’,leaf_size = 20

4.1 Performance Measure

The performance of five algorithms were estimated on four factors as described
below.These measurements rely on classification labels like TP (True Positive), TN
(True Negative), FP (False Positive), and FN (False Negative) [9]. Table 3 displays the
performance of various algorithms.

A. Accuracy: Divide the total population by the sum of TP and TN.

TP + TN
Accuracy =
TP + TN + FN + FP
296 S. Priyadarshinee and M. Panda

B. Precision: It is computed as the proportion of the number of correctly identified

positive results to the number of positive outcomes that the classifier forecasted.
TP
Precison =
(TP + FP)
C. Recall: It is determined by dividing the overall number of pertinent samples by the
quantity of accurate positive results. Recall = (TP+FN
TP
)
D. F1-Score: It is applied to determine whether or not a test is correct.

precision ∗ recall
F1 − Score = 2 ∗
precision + recall

Table 3. Performance of classifier for diabetes prediction

Algorithms Accuracy Precision Recall F1-Score

Random Forest 81.16 0.75 0.57 0.65
Extra Tree 81.16 0.80 0.51 0.62
Passive Aggressive Classifier 69.48 0.00 0.00 0.00
Decision Tree 70.77 0.52 0.51 0.51
K-Nearest Neighbour 78.57 0.68 0.55 0.61

As seen from Table 3, Random Forest is giving the highest accuracy of 81.16%,
precision of 75%, recall of 57%, f1-score of 65% and Extra Tree is also giving the highest
accuracy of 81.16%, precision of 80%, recall of 51%, f1-score of 62% for prediction
in diabetes among all classifiers. The accuracies obtained from the analysis is shown
graphically for better understanding in Fig. 7.

81.16 81.16
78.57

70.77
69.48

RF DT ET KNN PAC

Series1

Fig. 7. Comparison of classifier’s accuracy

 Machine Learning Algorithms for Diabetes Prediction 297

From the above Table 3 and Fig. 7, the Random Forest and Extra Tree clearly demon-
strate that they have performed the best, with an accuracy of 81.16 percent and KNN is
giving the second highest accuracy of 78.57%.

5 Conclusion
One of the deadliest illnesses in the real world is diabetes mellitus, and it might be
difficult to diagnose this illness in its early stages. In this study, five different machine
learning algorithms were used to develop a model that successfully addresses all issues
and aids in the early detection of diabetes disease. After comparing these algorithms, it
was experimentally seen that Random Forest and Extra Tree have the lowest error rates
and have the highest accuracy (81.16%).

References
1. Vats, V., et al.: A comparative analysis of unsupervised machine techniques for liver disease
prediction. In: 2018 IEEE International Symposium on Signal Processing and Information
Technology (ISSPIT), Louisville, KY, USA, 486-489 (2018)
2. Gavhane, A., Kokkula, G., Pandya, I., Devadkar, K.: Prediction of heart disease using machine
learning. In: 2018 Second International Conference on Electronics, Communication and
Aerospace Technology (ICECA), Coimbatore, pp. 1275–1278 (2018)
3. https://www.javatpoint.com/types-of-machine-learning
4. Joshi, R., Alehegn, M.: Analysis and prediction of diabetes diseases using machine learning
algorithm: ensemble approach. Int. Res. J. Eng. Technol. 4(10), 426–435 (2017)
5. Sivaranjani, S., et al.: Diabetes prediction using machine learning algorithms with feature
selection and dimensionality reduction. In: 2021 7th International Conference on Advanced
Computing and Communication Systems (ICACCS). Vol. 1. IEEE (2021)
6. International Diabetes Federation, IDF Diabetes Atlas, 8th edn. (2017)
7. Li, G., Peng, S., Wang, C., Niu, J., Yuan, Y.: An energy-efficient data collection scheme using
denoising autoencoder in wireless sensor networks. Tsinghua Sci. Technol. 24(1), 86–96
(2019)
8. Vijayan, V.V., Anjali, C.: Prediction and diagnosis of diabetes mellitus—a machine learning
approach. In: 2015 IEEE Recent Advances in Intelligent Computational Systems (RAICS).
IEEE (2015)
9. Sonar, P., JayaMalini, K.. Diabetes prediction using different machine learning approaches.
In: 2019 3rd International Conference on Computing Methodologies and Communication
(ICCMC). IEEE (2019)
10. Nagabushanam, P., et al.: CNN architecture for diabetes classification. In: 2021 3rd
International Conference on Signal Processing and Communication (ICPSC). IEEE (2021)
11. Chang, V., et al.: Pima Indians diabetes mellitus classification based on machine learning
(ML) algorithms. Neural Computing and Applications, pp. 1-17 (2022)
12. Sisodia, D., Sisodia, D.S.: Prediction of diabetes using classification algorithms. Procedia
Computer Science 132, 1578–1585 (2018)
13. Sarwar, M.A., et al.: Prediction of diabetes using machine learning algorithms in healthcare.
In: 2018 24th international conference on automation and computing (ICAC), pp. 16. IEEE
(2018)
14. https://www.kaggle.com/datasets/kandij/diabetes-dataset
298 S. Priyadarshinee and M. Panda

15. Breiman, L.: Random forests. Mach. Learn. 45(1), 5–32 (2001)
16. Sharaff, A., Gupta, H.: Extra-tree classifier with metaheuristics approach for email classifica-
tion. In: Proceedings Advances Computer Communication Computational Science. Springer,
Singapore, pp. 189197 (2019). https://doi.org/10.1007/978-981-13-6861-5_17
17. https://towardsdatascience.com/decision-trees-in-machine-learning-641b9c4e8052
18. https://thecleverprogrammer.com/2021/02/10/passive-aggressive-classifier-in-machine-lea
rning/
19. https://www.javatpoint.com/k-nearest-neighbor-algorithm-for-machine-learning
 CRODNM: Chemical Reaction Optimization
of Dendritic Neuron Models for Forecasting Net
Asset Values of Mutual Funds

Sarat Chandra Nayak1(B) , Satchidananda Dehuri2 , and Sung-Bae Cho3

1 Yonsei University, Seoul, South Korea
[email protected]
2 Fakir Mohan University, Balasore, Odisha, India
[email protected]
3 Yonsei University, Seoul, South Korea

[email protected]

Abstract. Learning algorithm and aggregation function used in the neurons have
imperative influence on the approximation of an artificial neural network. Den-
dritic neuron model (DNM) using additive and multiplicative-based aggregation
functions has been emerging as a machine learning approach and found successful
in many engineering applications. This study attempts to advance the predictive
accuracy of DNM through maintaining a decent steadiness between exploration
and exploitation of its search space with chemical reaction optimization (CRO)
algorithm, termed as CRODNM. The CRO, being a parameter free and powerful
global search optimization method synergies with better approximation capability
of DNM thus, able to overcome the limitations of conventional back propagation
learning based DNM. In addition to this, to start the search operation with a better-
quality initial population, we propose a new initial population generation method
for CRODNM by incorporating several methods. The proposed CRODNM is eval-
uated on forecasting net asset values of four mutual funds in terms of convergence
and prediction accuracy. The learning paradigm formed due to reasonable com-
bination of CRO and DNM (i.e., CRODNM) found competitive and outperforms
over DNM, multilayer perceptron (MLP), and genetic algorithm trained DNM
prediction models.

Keywords: Dendritic neuron model · Chemical reaction optimization · Financial

time series forecasting · Net asset value prediction · Mutual fund · MLP · GA

1 Introduction

Please Financial market data such as stock closing prices, commodity prices, energy
prices, currency exchange rates, and net asset value of mutual fund companies, etc.
exhibit high dynamism, chaotic in nature, following irregular movements, thus it is hard
to predict their future values [1]. The inherent uncertainties and associated volatilities
make the financial market data analysis a complex task [2]. Further, financial market is

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 299–312, 2022.
https://doi.org/10.1007/978-3-031-23233-6_23
300 S. C. Nayak et al.

highly fluctuating due to several macroeconomic factors. Regulatory bodies, political

scenarios and international policies and so on. Accurate prediction of such financial
data movement helps naïve investor and managers of financial institutions in decision
and policy making process. During past two decades many machine learning algorithms
such as artificial neural networks (ANN) have been emerging as sophisticated tools for
forecasting financial data. Several applications of neuron models found effective tools in
predicting financial data through modelling the underlaying nonlinearities coupled with
such data [3–7]. Investing in mutual funds have been growing interest in last decades.
Net asset value (NAV) is a decisive quantity to judge the financial health and stability
of a mutual fund thus, it is important to predict the NAV by the fund administrators
in addition to investors. The NAVs of mutual funds for a given day reflects assets and
liabilities, which depend on few components that are difficult to anticipate and frequently
changes. Accomplished brokers and fund managers attempt to estimate NAV in order
to maximize the return from the available opportunities accessible from the dynamic
international financial market with minimum risk. Therefore, an accurate and robust
NAV forecasting method is essential for the smooth management of a mutual funds.
A comparative analysis and review of machine learning models for NAV prediction is
carried out in [8]. Few machine learning models are applied in predicting bond risk
premiums [9] and asset pricing model [10, 11].
Though substantial number of ANN application exist in the literature, identification
of a best model over all financial market data is difficult. Each ANN has its unique archi-
tecture, distinct characteristics and limitations as well. The multilayer perceptron (MLP)
is the most frequently used model in financial forecasting. However, complex network
and back propagation learning in MLP influenced its performance a lot. Also, the aggre-
gation function used at the neurons plays a major role in the performance of a network.
Single neuron models with dendritic computations are proposed in the literature [12, 13].
Lately, dendritic neuron models (DNM) using multiplicative and additive aggregations
at its neurons are emerging as approximation tools [14]. Using a sigmoid function to
model the synaptic nonlinearity with a single neuron, it is found capable of computing
nonlinearly separable functions and approximating any multifaceted continuous func-
tion. The model uses Boolean logic to represent the nonlinear interactions in a dendrite
tree instead of calculating complex functions. DNM is a model of single neuron with
plastic dendritic morphology. It has a number of distinctive features which distinguish it
from conventional ANNs. It eliminates the redundant synapses and dendritic branches
through a neural pruning approach and thus possess a simple structure. Its simplified
structure can easily transform into logic circuit classifiers through logic approximation
scheme which achieves faster computation, smaller computational rate, and can resolve
problems with high-speed data streams compared to other neural models. A survey on
mechanism, algorithm, and real-world applications of DNM is carried out by authors in
[15].
The back propagation learning of conventional DNM makes it trapping into local
minima, fails to achieve the global best weights and thresholds for DNM and hence,
largely limiting its performance [16]. To avoid such problems, several evolutionary
algorithms are used for DNM training [16, 17]. Artificial immune system is used to train
DNM and evaluated on eight classification and eight prediction datasets [17]. The study
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 301

used Taguchi method to find the suitable user-defined parameters. The method found
better to compared methods and able to reduce the redundant synapses and dendritic
layers. In a short period of time, few evolutionary algorithms for DNM parameter tuning
are found in the literature such as GA [18], biogeography-based optimization [19], social
learning particle swarm optimization [20], differential evolution [21, 22], cuckoo search
[23], near population size reduction [24], whale optimization [25], and particle swarm
optimization [26, 27], etc. These hybrid DNMs are used for classification and forecasting
problems.
In the above cited works, evolutionary algorithms are used to decide the suitable
weights and thresholds of the DNM whereas, the other DNM controlling parameters
such as k, k soma and θsoma are decided experimentally. Though such procedures have
been combined with DNM and used for cracking several complex problems, their effi-
ciency is determined by well adjustable learning parameters. In the quest of searching
global optima, selection of such parameters makes the use of these technique problem-
atic. Selection of such control parameters while solving a particular problem necessi-
tates numbers of trial-and-error methods. Improper selection of such algorithm-specific
parameters may land the search operation at local optima or erroneous solution(s). Also,
these methods start their search process from a random initial population that may
affect the quality of final solution. Hence, for better accuracy, techniques with higher
approximation capability and fewer control parameters will be of interest. CRO is an
evolutionary optimization method inspired by the process of natural chemical reactions
[28, 29]. It has no algorithmic control parameter. Only the population size needs to be
declared. Hence, it is easy to implement without human interventions. ANN with CRO
methods is found successful applications in financial forecasting [5–7, 30].
This study is an attempt to enhance the predictability of DNM through searching its
weights, thresholds and three user-defined parameters by CRO. The optimal CRODNM
is obtained on fly during the evolution process rather fixing its structure at the beginning.
Therefore, a self-adjustable CRODNM is formed in an automatic fashion. The CRO is
parameter-free and does not necessitate any algorithmic control parameter. In addition
to this, to start the search operation with a better-quality initial population, we propose a
new initial population generation method for CRODNM incorporating several methods.
Finally, the proposed CRODNM is evaluated on predicting NAV of four mutual funds.
The methodologies used in this study are explained by Sect. 2. Outcomes from
simulation studies, comparative analysis are done in Sect. 3. Concluding remarks are
mentioned in Sect. 4 followed by a list of references.

2 Methodologies
The computations of basic DNM and the mechanism of the proposed CRODNM are
discussed in this section. The readers are suggested to refer the base articles as suggested
in Sect. 1 for DNM.

2.1 DNM
The computations in DNM are carried out through four layers such as synaptic layer, den-
dritic layer, membrane layer and soma layer. For an input vector X = {x1 , x2 , · · · , xn },
302 S. C. Nayak et al.

where xi is in [0, 1], weight wij , and threshold θij of DNM, the output Yij corresponds to
ith synaptic input in j th synaptic layer is computed as:

Yij = 1/1 + e−k (wij ∗xi −θij ) (1)

Here, k is a positive user-defined constant. Based on values of wij , and θij , six different
situations are arising as follows. If 0 < θij < wij , then Yij is proportional to xi and it
is called excitatory state. If wij < θij < 0, then Yij is inversely proportional to xi and it
is called inhibitory state. If 0 < wij < θij or wij < 0 < θij , then Yij value is always 0
and it is called constant-0 state. Finally, for conditions θij < wij < 0 or θij < 0 < wij , the
output Yij is always tends to 1, called as constant-1 state.
The nonlinear interaction of dendritic layer is computed by a multiplication unit.
The output of dendritic layer is given as:
n
Zj = Yij (2)
i=1

Outputs from m numbers of dendritic layers are linearly summed in the membrane
layer and the result is shown in Eq. 3.
m
V = Zj (3)
j=1

The final DNM output is computed at the soma layer using a positive constant ksoma
and self-defined parameter θsoma as in Eq. 4.
1
O= (4)
1 + e−ksoma (V −θsoma )
The total error signal E from N training samples at the output neuron is measured
as the deviation of from target values is computed as in Eq. 5.

1   
N
E= target i − Oi  (5)
N
i=1

During the training process, the learning algorithm continuously adjust the DNM
parameters, i.e., threshold θij and weight wij values in each synapse to optimize the DNM
performance.

2.2 CRODNM Based Forecasting

The goodness of initial population directly impacts the convergence rate and quality of
final solution of an evolutionary algorithm. Random initialization method is the mostly
used process which may not guarantee the quality of initial population. Therefore, we
propose a new initial population generation method for CRODNM by fusing random
initialization, opposition, quasi-opposition and generalized opposition learning based
methods to avoid its premature convergence and performance enhancement. The method
is described by Algorithm 1.
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 303

Algorithm 1: Initial Population Generation

1. Begin
2. Set and design variable .
3. Set counter
4. Randomly initialize each molecule in of size .
5. for do
for do

end for
6. end for
7. Set counter
8. for do
for do

end for
9. end for
10. Set counter
11. for do
for do

end for
12. end for

13. Evaluate fitness of each molecule of .

14. Select best molecules from
15. End

The CRODNM starts with the initial population as formed by Algorithm 1 and
different chemical reactions are applied as search operator to explore the search space
as presented in Algorithm 2. Here, the possible set of DNM parameters are considered
as the search space for CRO. An individual of CRODNM in t th iteration that represents
a potential parameter set of DNM is depicted in Fig. 3. There are 2NM + 3 parametersin
total need to be adjusted. Prospective readers are suggested to refer literature [5–7] for
detail implementation of different chemical reactions.
304 S. C. Nayak et al.

Algorithm 2: Chemical reactions search operation for DNM

Initialize population by Algorithm 1.
Set
While ( ) do
for do
if
if
Apply reaction on molecule
else
Apply reaction on
end if
else
Select another molecule
if
Apply reaction on
else if
Apply reaction on
else
Apply reaction on
end if
end if
Apply Reaction for increased fitness and update reactant
end for

end While
return the best molecule

Inputs are sequentially fed to DNM along with the molecules of the population.
The model computes an output as in Eq. 4. This output is compared with the target to
obtain an error value. The error signal so generated is considered as the fitness of the
corresponding molecule. Chemical reaction operators are applied to explore the search
space. Based on the fitness values, selection process is carried out and population is
updated in successive iterations. The process stops at the maximum iteration (similar to
equilibrium condition in chemical reaction). The best-fit molecule is chosen from the
converged population and supplied to the DNM along with the test data. The error values
calculated from the test data are considered as the CRODNM prediction performance
and preserved for comparative studies. Lower the error value, better is the predictability
of the model. The CRODNM process is shown in Fig. 1.
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 305

Fig. 1. CRODNM process

3 Experiments, Outcomes and Analysis

Collection of experimental data, input preparation for model and normalization, and
analysis of experimental outcomes are discussed in this section.

3.1 Collection and Statistical Summary of Data

The experimental data comprising stock prices of four emerging companies such as
Apple Inc., Cisco Systems, Microsoft, and Starbucks Inc. Listed in NASDAQ are down-
loaded from the Website: https://www.amfiindia.com/net-asset-value/nav-history.The
duration of data collection is 8th September 2021 to 7th September 2022. Statistics
from the datasets are summarized in Table 1. All the series showing stable and smaller
kurtosis value that indicates investment opportunity. All the datasets are platykurtic. The
Birla Sunlife and SBI data are deviating highly while ICICI and IDBI are stable.
306 S. C. Nayak et al.

Table 1. Numerical dataset and statistics.

Dataset Total no. of Mean Standard Kurtosis Skewness Generated

samples deviation Input
patterns
ICICI 243 11.48286 0.074645 2.694837 −0.23598 238
Prudential
Birla Sunlife 248 686.447 24.69773 2.294563 −0.14455 243
Mutual Fund
SBI Large & 248 360.8158 14.77311 2.423099 −0.37398 243
Midcap Fund
IDBI B&F 248 12.97681 0.826513 2.108236 −0.3046 243
Service Fund

3.2 Model Input Preparation and Normalization

Since the NAV prediction problem is a sequence prediction task, we adopted a sliding
window mechanism for generating training and test patterns from the original datasets
[5, 6]. For example, input patterns generated using a window size of 3 along with
corresponding target values are depicted in Fig. 2.

Fig. 2. Train and test patterns generated by sliding window

The model inputs are and normalized using sigmoid method as in (6) and finally after
prediction process, a de-normalization method is carried out to get the real-time data. In
(6), xnorm is the normalized value of xi , xmin and xmax are the minimum and maximum
data value of a pattern generated by sliding window respectively.
1
xnorm = 
xi −xmin
 (6)
− xmax
1+e −xmin

3.3 Experimental Design

As stated earlier, CRO has no algorithmic control parameter that need to be tuned. Only
the population size needs to be defined. In this work, the population size, i.e., ReacNum
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 307

was set to 50. The convergence criterion was set to 100 maximum number of iterations.
The DNM parameters such as weight wij , threshold θij , k, k soma and θsoma are decided
by CRO in an evolutionary manner. Figure 3 depicts an individual representation for
CRODNM. The population size, cross-over and mutation probability of GA was chosen
as 80, 0.6 and 0.003 respectively. The MLP used has one hidden layer and the learning
rate and momentum factor was set to 0.002 and 0.01 respectively. For a fair comparison,
the same input pattern is fed to all predictors and all methods iterated for 100 times. In
order to compensate the stochastic nature of predictors, for each input pattern a model
was executed 30 times and average prediction errors from 30 runs are considered as
the performance of a model. Only one neuron is at output layer because there is only
one target variable in each input pattern to be predicted. The mean absolute percentage
of error (MAPE) as in (7) is chosen as the model performance measure. For the sake
of space, the error convergence graph for ICICI prudential series only presented as in
Fig. 4. It can be seen that the CRODNM converged faster than others.
 
1 No.ofpattern Target i − Predicted i 
MAPE = ∗ 100% (7)
No.ofpattern i=1 Target i

Fig. 3. A molecule representation of DNM

3.4 Analysis of Experimental Results

The MAPE statistics from four datasets using four predictors are recorded and listed
in Table 2. The average MAPEs generated by CRODNM are 0.012816, 0.015368,
0.018815, and 0.021825 for ICICI prudential, Birla Sunlife mutual fund, SBI large
and midcap, and IDBI mutual fund respectively which are lowest when compared with
that from others. The GADNM model is the second good model. Both CRODNM and
GADNM found efficient compared to conventional DNM. It seems that, they easily fine-
tuned the DNM parameters, converged faster and thus, capable to escape the local optima.
The MLP predictions are found inferior. Further, the box-whisker plots of MAPEs from
ICICI dataset are depicted in Fig. 5 established the better performance of CRODNM. The
predicted NAVs for ICICI data series by four predictors are plotted in Fig. 6. It shows
that CRODNM predictions are much closer to actual NAVs compared to others. The
error histogram in Fig. 7 shows that for most of the input patterns of ICICI, the proposed
CRODNM produced error values closer to zero. These evidences from experimental
work highly suggested the goodness of CRODNM predictions over others.
308 S. C. Nayak et al.

Fig. 4. Error convergence of predictors from ICICI data series

Table 2. MAPE statistics for four datasets and four predictors.

Dataset Statistic CRODNM GADNM DNM MLP

ICICI Prudential Minimum 2.22E-05 0.000129 4.17E-05 0.000645
Average 0.012816 0.014045 0.013301 0.031826
Maximum 0.126585 0.122585 0.132585 0.097585
Median 0.00806 0.010362 0.007406 0.031674
Standard dev 0.016763 0.015977 0.017816 0.014524
IQR 0.012362 0.011174 0.013579 0.014113
Birla Sunlife Mutual Fund Minimum 2.04E-05 3.29E-05 0.00349 0.009415
Average 0.015368 0.017624 0.061509 0.134923
Maximum 0.168562 0.195585 0.238457 0.383457
Median 0.00806 0.014627 0.062924 0.137835
Standard dev 0.013760 0.015223 0.015856 0.02109
IQR 0.012325 0.01219 0.015216 0.015495
SBI Large & Midcap Fund Minimum 2.26E-05 3.25E-05 0.00366 0.006191
(continued)
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 309

Table 2. (continued)

Dataset Statistic CRODNM GADNM DNM MLP

Average 0.018815 0.017624 0.061509 0.095336
Maximum 0.106500 0.115585 0.108457 0.143457
Median 0.007930 0.014627 0.062924 0.097835
Standard dev 0.016337 0.015223 0.015856 0.019128
IQR 0.011625 0.01432 0.012164 0.013458
IDBI B&F Service Fund Minimum 1.98E-04 3.29E-05 0.004240 0.009455
Average 0.021825 0.027465 0.060526 0.124925
Maximum 0.100658 0.135544 0.117453 0.163457
Median 0.007935 0.014627 0.069263 0.133572
Standard dev 0.013703 0.017249 0.013562 0.021109
IQR 0.011566 0.013115 0.015224 0.015485

Fig. 5. Box plots of MAPEs generated by four predictors from ICICI data series
310 S. C. Nayak et al.

Fig. 6. Model predictions against actual NAV values from ICICI data series

Fig. 7. Error histogram of CRODNM from ICICI data series

4 Conclusions and Further Scope

In order to improve the prediction accuracy of DNM, this study developed a hybrid
forecast termed as CRODNM. The powerful global search ability of CRO hybridized
with better approximation capability of DNM thus, able to overcome the limitations
of conventional back propagation learning based DNM. To initiate the search opera-
tion with a better-quality initial population, a new initial population generation method
for CRODNM incorporating several methods is suggested. The proposed CRODNM is
 CRODNM: Chemical Reaction Optimization of Dendritic Neuron 311

applied to predict the NAVs of four real-world mutual funds exploring one-year his-
torical data. It observed that CRODNM is quite able to model NAV historical data
and capture the underlaying dynamism coupled with such data efficiently. The pre-
dictability of CRODNM is compared with three others similarly developed forecasts
such as GADNM, DNM, and MLP and found superior. Experimental studies suggest
that CRODNM clearly improves the NAV prediction accuracy with good convergence
speed and effectively reaching the global optima. The study can be extended further
by incorporating advance learning mechanisms and application to other financial time
series as well.

Acknowledgement. Dr. Sarat Chandra Nayak was supported by BK21 grant funded by Korean
government at Yonsei University. Dr. Satchidananda Dehuri would like to thank SERB, Govt. of
India for financial support under Teachers’ Associateship for Research Excellence (TARE) fel-
lowship vide File No. TAR/2021/000065 for the period 2021–2024. Dr. Sung-Bae Cho was sup-
ported by an IITP grant funded by the Korean government (MSIT) (No. 2020-0-01361, Artificial
Intelligence Graduate School Program, Yonsei University.

References
1. Fama, E.: Efficient Capital Markets: A Review of Theory and Empirical Work. R. Lowbridge
(Module Leader), New York (1970)
2. Kara, Y., Boyacioglu, M.A., Baykan, O.K.: Predicting direction of stock price index movement
using artificial neural networks and support vector machines: the sample of the Istanbul stock
exchange. Expert Syst. Appl. 38(5), 5311–5319 (2011)
3. Hafezi, R., Shahrabi, J., Hadavandi, E.: A bat-neural network multi-agent system (bnnmas)
for stock price prediction: case study of dax stock price. Appl. Soft Comput. 29, 196–210
(2015)
4. Wang, J., Wang, J.: forecasting stock market indexes using principal component analysis and
stochastic time effective neural networks. Neurocomputing 156, 68–78 (2015)
5. Araujo, R., Oliveira, A., Meira, S.: A hybrid model for high-frequency stock market
forecasting. Expert Syst. Appl. 42(8), 4081–4096 (2015)
6. Nayak, S.C., Misra, B.B., Behera, H.S.: ACFLN: artificial chemical functional link network
for prediction of stock market index. Evol. Syst. 10(4), 567–592 (2018)
7. Nayak, S.C., Misra, B.B., Behera, H.S.: Artificial chemical reaction optimization of neural
networks for efficient prediction of stock market indices. Ain Shams Eng. J. 8(3), 371–390
(2017)
8. Yang, Y.J., Chen, B., Zhang, L.L.: Asset price prediction via machine-learning method: a
review. In: 2021 17th International Conference on Computational Intelligence and Security
(CIS), pp. 168–172. IEEE (2021)
9. Bianchi, D., Büchner, M., Tamoni, A.: Bond risk premiums with machine learning. Rev.
Financial Studies 34(2), 1046–1089 (2021)
10. Fama, E.F., French, K.R.: A five-factor asset pricing model. J. Financ. Econ. 116(1), 1–22
(2015)
11. Gu, S., Kelly, B., Xiu, D.: Empirical asset pricing via machine learning. Rev. Financial Studies
33(5), 2223–2273 (2020)
12. Tang, Z., Tamura, H., Ishizuka, O., Tanno, K.: A neuron model with interaction among
synapses, IEEJ Trans. Electron., Inform. Syst. 120(7), 1012–1019 (2000)
312 S. C. Nayak et al.

13. Tang, Z., Kuratu, M., Tamura, H., Ishizuka, O., Tanno, K.: A neuron model based on dendritic
mechanism. IEICE 83, 486–498 (2000)
14. Todo, Y., Tamura, H., Yamashita, K., Tang, Z.: Unsupervised learnable neuron model with
nonlinear interaction on dendrites. Neural Netw. 60, 96–103 (2014)
15. Ji, J., Tang, C., Zhao, J., Tang, Z., Todo, Y.:. A survey on dendritic neuron model: mechanisms,
algorithms and practical applications. Neurocomputing (2022)
16. Gao, S., Zhou, M., Wang, Y., Cheng, J., Yachi, H., Wang, J.: Dendritic neuron model with
effective learning algorithms for classification, approximation, and prediction. IEEE Trans.
Neural Networks Learning Syst. 30(2), 601–614 (2018)
17. Tang, C., Todo, Y., Ji, J., Lin, Q., Tang, Z.: Artificial immune system training algorithm for
a dendritic neuron model. Knowl.-Based Syst. 233, 107509 (2021)
18. Ji, J., Song, Z., Tang, Y., Jiang, T., Gao, S.: Training a dendritic neural model with genetic algo-
rithm for classification problems. In: 2016 International Conference on Progress in Informatics
and Computing (PIC), pp. 47–50. IEEE (2016)
19. Wang, S., Sugiyama, D., Sun, J., Yang, L., Gao, S.: Dendritic neuron model trained by
biogeography-based optimization for crude oil price forecasting. In: 2018 10th Interna-
tional Conference on Intelligent Human-Machine Systems and Cybernetics (IHMSC), Vol.
1, pp. 36–40. IEEE (2018)
20. Song, S., Chen, X., Tang, C., Song, S., Tang, Z., Todo, Y.: Training an approximate logic
dendritic neuron model using social learning particle swarm optimization algorithm. IEEE
Access 7, 141947–141959 (2019)
21. Wang, S., et al.: A novel median dendritic neuron model for prediction. IEEE Access 8,
192339–192351 (2020)
22. Xu, Z., Wang, Z., Li, J., Jin, T., Meng, X., Gao, S.: Dendritic neuron model trained by
information feedback-enhanced differential evolution algorithm for classification. Knowl.-
Based Syst. 233, 107536 (2021)
23. Qian, X., Tang, C., Todo, Y., Lin, Q., Ji, J.:. Evolutionary dendritic neural model for
classification problems. Complexity (2020)
24. Song, Z., Tang, Y., Ji, J., Todo, Y.: Evaluating a dendritic neuron model for wind speed
forecasting. Knowl.-Based Syst. 201, 106052 (2020)
25. Han, Z., Shi, J., Todo, Y., Gao, S.: Training dendritic neuron model with whale optimiza-
tion algorithm for classification. In: 2020 IEEE International Conference on Progress in
Informatics and Computing (PIC), pp. 11–15. IEEE (2020)
26. Yilmaz, A., Yolcu, U.: Dendritic neuron model neural network trained by modified particle
swarm optimization for time-series forecasting. J. Forecast. 41(4), 793–809 (2022)
27. Egrioglu, E., Bas, E., Karahasan, O.: Winsorized dendritic neuron model artificial neural
network and a robust training algorithm with Tukey’s biweight loss function based on particle
swarm optimization. Granular Computing, 1–11 (2022)
28. Lam, A.Y., Li, V.O.: Chemical-reaction-inspired metaheuristic for optimization. IEEE Trans.
Evol. Comput. 14(3), 381–399 (2009)
29. Alatas, B.: A novel chemistry-based metaheuristic optimization method for mining of
classification rules. Expert Syst. Appl. 39(12), 11080–11088 (2012)
30. Nayak, S.C., Misra, B.B.: Extreme learning with chemical reaction optimization for stock
volatility prediction. Financial Innovation 6(1), 1–23 (2020)
 Machine Learning Approaches and Particle
Swarm Optimization Based Clustering
for the Human Monkeypox Viruses: A Study

Akshaya Kumar Mandal1 , Pankaj Kumar Deva Sarma1(B) ,

and Satchidananda Dehuri2
1 Department of Computer Science, Assam University, A Central University of India, Silchar,
Assam 788011, India
[email protected]
2 Department of Computer Science, Fakir Mohan University, Vyasa Vihar, Balasore,

Odisha 756019, India

Abstract. The world is currently battling with epidemic fear because of the preva-
lence of monkeypox cases in various corners of the globe, even as the threat of
the COVID-19 pandemic declines. The majority of the 2022 outbreak’s cases of
monkeypox virus are found in Africa now found in nations throughout Europe and
the Western Hemisphere. With ten instances of the monkeypox virus recorded in
August 2022 in India, along with one fatality, this endemic has now spread to more
than 90 nations globally. The ‘Orthopoxvirus genus of the Poxviridae family’ con-
tains a large number of zoonotic viruses, including the monkeypox virus. The term
“monkeypox” describes the disease because the first viral infection detected by
this virus was in macaque monkeys. This article presents a summary of the most
current, cutting-edge applications of “Machine Learning (ML) and Particle Swarm
Optimization (PSO)” clustering to the Monkeypox instances. The literature has
highlighted many large data elements where ML has promised to play a significant
role. Advanced analytical techniques have proved useful in forecasting the onset
of the disease as well as recognizing its symptoms and signs. Furthermore, ML-
based bio-inspired big data analytics has greatly aided contact tracing, molecular
analysis, and drug development. Even though it is exceedingly rare, it spreads
when direct contact with the “mucocutaneous lesions” of an infected patient or
respiratory droplets has been associated with the spread of the monkeypox virus.

Keywords: Monkeypox · PSO · Clustering · Zoonotic viruses · ML ·

Bio-inspired techniques

1 Introduction

A viral illness known as monkeypox can affect both humans and other animals [4].
‘Fever, enlarged lymph nodes, and a blistering and crusting rash’ are all symptoms of
monkeypox virus [11]. Between 5–21 days may pass between exposure and the develop-
ment of symptoms [7, 55]. ‘Symptoms of the monkeypox virus generally last between

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 313–332, 2022.
https://doi.org/10.1007/978-3-031-23233-6_24
314 A. K. Mandal et al.

2 to 4 weeks [7]. It can happen without any symptoms as well as with minor signs
[54]. There has not been any evidence that the distinctive appearance of swelling of the
glands and lesions, because of fever and muscle ache are common throughout differ-
ent outbreaks [11, 54]. As shown in Table 1, Particularly in young infants, expecting
women, and individuals who have weak immune systems, cases can become severe [6,
56]. The ‘zoonotic virus’ belonging to the ‘Orthodox virus’ genus is responsible for
the illness. This genus includes the smallpox-causing variola virus [14, 35]. Compared
to the ‘Central African’ (Congo Basin) strain, the ‘West African’ strain infects people
less severely [19, 30]. Body fluids, contaminated products, handling of dumpsters, and
other close contacts with an infected person are ways to spread the virus [55]. Spread-
ing can be aided by tiny droplets and, presumably, airborne routes [53, 54]. When the
same symptoms last longer than a week, people are still contagious from the moment
the symptoms appear until all lesions have crusted over and fallen off [55]. The virus
reportedly originated in Africa among certain rodents [30] in the first instance. A lesion
of viral DNA can have retrieved to confirm the diagnosis [42].
The Democratic Republic of the Congo (DRC) (Zaire), in the city of ‘Basankusu of
Equateur Province’, was where the first case of human monkeypox identified in 1970
[19]. Confirmed cases in total, 338 and 33 deaths (9.8% CFR) has found between 1981
and 1986 in the ‘DRC/Zaire’ under WHO surveillance. From 1996–1997, the ‘DRC and
Zaire saw the second wave of human sickness. Between 1991 and 1999, 511 occurrences
in total are founds in the ‘DRC/Zaire’ [19, 30]. In the DRC, the Congo Basin Clade of
illness is still present and has a high CFR (case fatality rate) [19]. It was once restricted
to tropical jungle regions. The pattern was interrupted in 2005 when 49 instances were
‘reported in Sudan’ (regions that are now known as South Sudan), but there were no
deaths [19]. The genomic study indicates that the virus ware imported, most likely from
the DRC [30] (Fig. 1).

Table 1. Symptoms of Monkeypox Virus

Monkeypox virus category Symptoms

Asymptotic pre-symptom No Symptoms [11, 35]
Mild The rash appeared, Fever, Mucocutaneous lesions, and Fatigue
[35, 53]
Moderate Mucocutaneous lesions, Muscle aches [35, 53]
Severe High fever, Mucocutaneous lesions, Difficulty [53, 54]
Critical Mucocutaneous lesions, Weak pulse, Respiratory failure, and
Organ failure [46, 53, 54]
 Machine Learning Approaches and Particle Swarm Optimization 315

Fig. 1. Monkeypox lesion on human body and DNA of Monkeypox Virus [58]

Following that, many additional cases of monkeypox are founds in ‘Central and West
Africa’, mostly in the DRC: 2000 cases each year between 2011 and 2014. Due to the
current data’s frequent fragmentation and lack of support, it is difficult to make accurate
estimates of the number of cases of monkeypox over time. However, historical data
showing both the quantity and geographic distribution of cases recorded in 2018 had
increased [11]. By January 1, 2022, reports of monkeypox cases had come in from 42
member states across five distinct WHO regions. Table 2 displays the number of mon-
keypox cases and fatalities globally from January to August 2022. The geographical
regions of the ‘Americas are composed of Africa, Europe, the Eastern Mediterranean,
and the Western Pacific. Guys who have had intercourse with other males and who
have admitted to having sex lately with several or new partners are more susceptible
to contracting the monkeypox outbreak. The majority of current epidemic cases were
reported through “sexual health” or other health services at primary or secondary health
care institutions, with a history of travel to “Europe, North America, or other countries”
rather than previously unknown regions, and increasingly with recent local travel or no
travel at all. Epidemiological investigations are conducted if a single case of this virus
is detected then the outbreak of monkeypox in that nation is considering an epidemic.
The rapid development of monkeypox in many areas raises the potential that transmis-
sion has been going on for some time without being observe given the initial absence
of epidemiological linkages to previously reported monkeypox locations. This article
presents a summary of the monkeypox virus based on ‘machine learning and particle
swarm optimization clustering’. The use of bio-inspired approaches such as machine
learning may speed up diagnosis on a far wider scale than physical testing by improving
screening and using the massive volumes of data generated by numerous ‘laboratory
tests, clinical symptoms, and radiological imaging [27]. Using current quantitative and
qualitative data, machine-learning approaches can predict ‘infected patients’ [9, 25]. The
general characteristics of monkeypox occurrences are illustrate in Fig. 2, which provides
a mechanism for medical practitioners to make judgments during treatment and diag-
nosis. There is currently no feasible treatment available for those who have infected.
In severe situations, supportive treatments such as ‘tecovirimat’ may be used. Because
there are no clear guidelines for symptom relief as these treatments are random.
316 A. K. Mandal et al.

2 Machine Learning for Monkey Pox Virus

The present global monkeypox virus endemic has infected thousands of people, infras-
tructure stability, and global economies [53]. The ongoing COVID-19 epidemic has
placed a tremendous strain on the healthcare sector due to a lack of suitable facilities to
respond to the disease. The sickness has imposed a tremendous load on the healthcare
system due to a lack of capacity to tackle the challenges caused by the COVID-19 pan-
demic [10]. Additionally, it has been demonstrate that the employment of ML enables
some essential functions inside the healthcare system, saving many lives while enhanc-
ing the quality of healthcare service [25, 38]. Furthermore, like many “chronic diseases”,
the monkeypox virus is a global endemic that poses a threat to the health of the whole
world’s population. These occurrences include death, the need for hospital facilities,
and the arising of these requirements. If a person has diabetes, hypertension, immuno-
suppressive diseases, cystic fibrosis, or cancer, they may be more susceptible to these
harmful effects [31, 44]. Furthermore, children may be more vulnerable to the severity
of this monkeypox virus than adults. Using the available data, machine learning algo-
rithms may be incredibly effective in determining the risk of exposure or potential side
effects in individuals infected with the monkeypox virus [59, 60]. Due to its data-driven
nature, machine learning (ML) may be useful in determining the optimal method to
treat individuals infected with the monkeypox virus based on recent data or features of
contagious viruses [57]. It might also aid in the development of treatment regimens and
even the prediction of treatment results. For individuals with persistent monkeypox virus
infection, for example, an artificial intelligence model can forecast what would happen if
a mechanical ventilator was delay. The medical staff may apply machine learning (ML)
approaches to anticipate the probability of infection, mortality rate, and other bad event
consequences with effective private-public sector cooperation using data from ‘airlines,
social media, traffic regions, and other monkeypox virus susceptible locations’ [9, 25].
To successfully developed and enhance smallpox vaccines, find the ideal clinical tri-
als, evaluate infected patients at different time points, and track disease transmission
by the specialists in the medical profession, they are looking for innovative technology.
Because of its quicker processing speed, better effectiveness, and enhanced scale-up,

Fig. 2. Different characteristics of the monkeypox incidents

 Machine Learning Approaches and Particle Swarm Optimization 317

recent research has shown that using ML with bio-inspired methodologies can mod-
ernize the health industry [12, 25]. Figure 4 illustrates a variety of machine learning
applications for the monkeypox virus (Fig. 3).

2.1 Screening and Treatment

The endemic monkeypox virus is control with early detection and treatment. The “nu-
cleic acid amplification testing” (NAAT) technique, which uses real-time or conventional
“polymerase chain reaction,"(PCR) is the industry standard for identifying unique viral
DNA sequences. Sequencing can do independently or in conjunction with PCR [12].
However, research focuses on other ways to improve this [27]. These methods, in gen-
eral, are often less costly, time-consuming, have a low true-positive rate, and require the
use of sophisticated equipment. Because of this, timely identification and monitoring can-
not accomplish using the traditional approach. Utilizing smart devices and bio-inspired
frameworks is a quick and affordable way to identify the monkeypox virus [55, 59].
Using ‘skin lesion samples, such as swabs of the lesion, and plasma antibody testing’,
the monkeypox virus can detect. On the other hand, monkeypox is not be detected by the
use of ‘plasma antibody detection. Finding LgM (Laboratory and Genomic Medicine)
in recently acutely unwell individuals or lgG (Low-Grade Glioma) in matched blood
samples taken at least 5 to 21 days apart, with the first sample taken during the first week
of sickness, may help with the diagnosis [56], if the tested samples produce different
results. Recent vaccinations may affected by blood screening. Jabeen et al. [13] provide
a comprehensive overview of Monkey Pox virus replication and the requirement for
antiviral treatment in humans. According to his studies, neither humans nor animals can
get the monkeypox virus through the respiratory route. Human-to-human transmission is
possible because of genetic alterations that may have led to the emergence of monkeypox
in humans. A powerful antiviral drug that works against the monkeypox virus is urgently
required since it may use as a ‘bioweapon’. The efficacy of ST 246 on monkeypox [37] or
Orthopox-infected people is unknown; however, it demonstrated to be effective in vivo,
in vitro, and in trials on infected animals as well as non-infected humans. Suspected
human transmission mechanisms for the monkeypox virus, are depicts in Fig. 5.

Table 2. The global monkeypox virus cases with death cases between January to august 2022
[57].

Region of WHO Member States Confirmed cases Deaths

Region of the Americas 20 438 2
African Region 404 7
Western Pacific Region 121 0
Eastern Mediterranean Region 35 0
Region of South-East Asia 14 1
Region of European 20652 2
Total 41664 12
318 A. K. Mandal et al.

Fig. 3. The global monkeypox virus cases with death between Januarys to august 2022

2.2 Contact Tracing

To prevent the disease from spreading, it is critical to identify the person’s contacts
after infected with the monkeypox virus. Contact tracing is the procedure of finding
and tracing people, who have recently met a monkeypox virus patient. The approach
is often uses to identify the affected person for a follow-up 5 to 21 days following
exposure [7, 55]. If correctly implemented, this technique has the potential to disrupt
the new Monkey Pox Virus transmission cycle and lessen the epidemic by boosting
the possibility of appropriate control. Several affected countries have developed various
digital contact tracing methods that make use of mobile apps that use techniques such as
“Bluetooth, GPS,” “system physical address,” and others [5]. A digital contact tracing
system may operate significantly quicker than a non-digital system and is very similar
to real-time. The majority of these digital tools are designs to acquire user information
through machine learning algorithms evaluate to identify who is most infects with a
particular virus based on recent contact chains [55]. Petersen et al. [36] meticulously
prepared a list of ML-based contact tracing and diagnosis applications. The first viremia
that occurs from this virus’s intracellular replication spreads to neighboring lymph nodes.
Infection with monkeypox can result in ‘bronchopneumonia, dehydration, respiratory
distress, encephalitis’, and other consequences. There is no treatment for monkeypox,
as cases are still report across the world. Supportive therapy, such as ‘Tecovirimat’,
can used to treat illness. Because there are no accepted recommendations for symptom
reduction, as all treatments are entirely individual.

2.3 Prediction and Forecasting

The government, public health institutions, and policymakers will be able to deal with
the endemic lot more effectively if the monkeypox virus epidemic is predict and forecast.
Susceptible infected and removed (SIR), “susceptible exposed infected and recovered”
(SEIR), susceptible infected recovered and dead (SIRD), infected, infected, and dead
(IID), infected, infected are the following terms used for prediction and forecasting the
virus [25]. Several ML-based prediction models for the monkeypox virus are depicts
as shown in Fig. 4. The performance of these models has to anticipate and improve, as
more data is available.
 Machine Learning Approaches and Particle Swarm Optimization 319

Fig. 4. ML tools for monkeypox virus

2.4 Drug Design and Molecular Research

Machine learning (ML) and molecular biology (MB) are well-known interdisciplinary
research fields [13, 25]. Modern molecular biology employs advanced tools to analyze
and interpret data and answer complex biological questions that traditional laboratory
techniques cannot solve. Deep learning has applied to molecular design, most promi-
nently in modeling, simulation, and drug discovery [25, 37]. With the emergence of the
monkeypox virus, scientists and medical specialists from all over the world have pushed
for a practical solution to build monkeypox virus treatments and vaccinations by using
machine-learning approaches [25, 36, 53]. By aiding in the creation of cutting-edge
medications and vaccines, recommending fresh inhibitors, looking into how structural
changes affect the viral genetic variation, and other activities [16, 40], these efforts can
help in the fight against the endemic spread of the monkeypox virus. The endemic may
control through treatment and vaccination programs against the monkeypox virus to
create potential drug designs for clinical trials, and the molecular and metabolic mech-
anisms of various disease pathogenesis must have understood. Drug repositioning, also
known as medicinal repositioning, is a successful method for identifying new pharma-
cological or therapeutic uses for drug molecules that are already in use. In comparison
to traditional drugs, developing new treatments frequently costs less and takes less time
[8, 13, 36]. This is particularly true for the monkeypox virus, where drugs used to treat

Fig. 5. Monkeypox virus Humans Transmission Suspected Models

320 A. K. Mandal et al.

one viral DNA group, such as smallpox, are also use to treat another. The use of ML for
molecular analysis and vaccine development is discussed in Sect. 3.2.7 and Table.3.

3 PSO Clustering for the Human Monkeypox Viruses

As a result of growing biological data volume, rapid data volume expansion, and tech-
nological innovation, bioinformatics originated [50]. Big data analytics were required
in the monkeypox virus case because there was a significant amount of patient data,
including medical records, X-rays, images of monkeypox virus patients, lists of doctors
and nurses, and their travel histories. As the number of people affected by the monkey-
pox virus grows, bio-inspired computing can be useful in investigating these data sets
and identifying trends. Bunge et al. [6] investigated open-data tools for tracking, simu-
lation, as well as the shifting epidemiology of human monkeypox as a potential issue.
Accessible organizations, open-source communities, and geographic data resources for
monkeypox viruses are represent in Tables 1, 2, 4.

Fig. 6. PSO Clustering for human monkeypox virus

3.1 Introduction to Particle Swarm Optimization (PSO)

For continuous and discrete optimization, PSO [17] is employed with the ‘dimension D
of the search space’ is the same for all particles, and each particle “1 ≤ i ≤ Nd ” has a
position“Pi ,d ” and a velocity “Vi„d ”. Before they may move about in the search space, all
particles must be assigned random starting places and velocities. Each iteration results in
the identification of the personal optimal solution “Pbest ” with the overall optimal solution
 Machine Learning Approaches and Particle Swarm Optimization 321

“Gbest ” for each particle. According to the following equations [17], the particle’s new
position and velocity are:
Updated velocity: vi t+1 = w. vi t + c1 r1 (pt best - pi t ) + c2 r2 (Gt best - pi t ).
‘w’ is the inertia weight in the above equation, while ‘c1 ’ and ‘c2 ’ are the global
constants.
Updated position: pi t+1 = pi t + vi t+1 .
The particle updates both ‘Pbest ’and ‘Gbest ’ in the following way:
Particle swarm optimization algorithm [17]:

• Calculate the fitness factor for each particle.

• Examine the modifications in position and speed.
• Track down ‘Pbest ’ and ‘Gbest ’ until the velocity value is zero or the specified number
of iterations has reached.

Fig. 7. PSO clustering for monkeypox virus using an image [58] of the infected patients

3.2 Clustering Based on PSO

Clustering is a technique that combines “similar data points” from a given data set into
a single group [2]. Few application of clustering algorithms includes ‘image compres-
sion [20], pattern analysis [22], and picture segmentation [43]’. For clustering analysis,
‘particle swarm optimization, or ‘clustering based on swarm intelligence, may be used
[43]. Data points have treated as particles in ‘particle swarm optimization approaches.
Different clustering methods are use to collect the first groups of particles, until the
cluster center, converges. The particle cluster is an update based on the cluster center,
velocity, and position. (Represent in Fig. 6 and 7).
Using clustering method based on particle swarm optimization [43].

• Determine each particle’s degree of fitness inside a cluster. ∀xi  cj (cj is a ‘cluster’ j).
• Determine the most recent positions and ‘velocities of each particle’ within ‘cluster’
cj .
322 A. K. Mandal et al.

• Replace ‘Pbest ’ with the ith cluster center & ‘Gbest ’ with the neighboring cluster
centers.

Particle swarm optimization clustering iterates until it reaches the maximum number
of iterations or converges. Previous sections described PSO clustering with the equations
for velocity and position updates using “Pbest ” and “Gbest " fitness matrices. The cluster
centers for the PSO approach to image clustering are as follows:

vit+1= w. vit + c1r1 (ptbest- pit) + c2r2 (Gtbest- pit)

pit+1= pit + vit+1
Where, pit represents Particle position.
vit represents Particle velocity.
c1, c2 represent cognitive and social parameters.
r1 r2 are random numbers between 0 and 1.

3.2.1 Clustering Algorithm Based on PSO

The micro-aneurysms are effectively separate in this study using “blended discrete PSO
clustering.“ In terms of swarm algorithms, “Blended Discrete Particle Swarm Optimiza-
tion” (BPSO) is the best [20]. It is important to determine the “swarm size of particles”
instantly as follows:
Step 1: Initialize all swarm particles.
Step 2: Determine each particle’s fitness rate.
Step 3: Continue for each particle.

• Determine the cluster block’s velocity.

• For each particle, use the k - nearest search method.
• Calculate the performance metric.
• ‘Pbest ’ and ‘Gbest ’ must be update.

Step 4: Go back to ‘Gbest ’.

3.2.2 Image Fitness Tests

The fitness test determines the primary accuracy of the segmented computed tomogra-
phy scan image. The class of the “clusters” has evaluated by applying fitness metric.
The expression xn = (xn,1 ,… xn,k ) represents each particle in a cluster. Groups of data
clustering approaches are referring to swarm. The fitness function is used to rate the par-
ticle quality in the image. Figure 7 illustrates the fitness metrics used in the previously
described enhanced “particle swarm optimization clustering.“

3.2.3 Stages of Feature Extraction

The lesions are extracts from the segmented images during the feature extraction stage of
the proposed system. Area, perimeter, circularity, and diameter are the parameters used
to identify lesions. Therefore, to improve accuracy, the segmented images should be free
 Machine Learning Approaches and Particle Swarm Optimization 323

of unnecessary and irrelevant features. This provided an automatic disease identification

of lesions in patients’ photos, which is extremely desirable to assist doctors in diagnosis.

3.2.4 Screening and Treatment

Monkeypox viruses have identified in ‘most’ or all animal fluids and excretions (includ-
ing ‘urine, faces, and oral, nasal, and conjunctival exudates’) in addition to skin lesions
[53, 54]. Consumption of infected tissues and direct injection into skin breaches are fur-
ther methods of transmission. The importance of aerosol transmission varies depending
on the species or habitat. Human infection can cause by ‘animal bites, aerosols from
proximity, or direct contact with lesions, blood, or body fluids (Fig. 5). There have a few
cases when vaginal lesions increased the probability of sexual transfer, and transplacen-
tal transmission has seen. Handling, preparing, and eating wild animals have repeatedly
linked to clinical cases in Africa. However, in many epidemics, the person-to-person
transmission was also common [31].
Tecovirimat (chemical agent ST246), also known as ‘Arestyvir’, has been licensed for
use in people who infected with orthopoxviruses, albeit its ‘specific effectiveness’ against
‘monkeypox’ in humans has yet to be demonstrated. Several possible treatments are
also undergoing clinical trials, including a cidofovir derivative (CMX001/Brincidofovir)
[45]. ‘Vaccinia immune globulin’, which was formerly used to treat ‘smallpox’, must
investigate, especially in ‘immunocompromised’ individuals [8]. The main advantage of
using this method is that it allows medical practitioners to communicate the appropriate
steps to create the essential alert system. These interactions’ accuracy and dependability
were still major challenges for this method.

Table 3. Review of clinical research for human monkeypox virus

Monkeypox virus symptoms Initial research Secondary research

Respiratory symptoms McCollm et al. [28] Parker et al. [33]
Quenelle et al. [37] Parker et al.(2007) [34]
Weinstein et al. [52] Jain et al. [14]
Kugelman et al. [19]
Rash Quenelle et al. [37] Breman et al. [4]
Petersen et al. [36] Parker et al. [33]
Weinstein et al. [52] Lamet al. [21]
Kugelman et al. [19] Sale et al. [41]
McCollm et al. [28] Parker et al.(2007) [34]
MacNeil et al. [26]
Petersen et al.[36]
(continued)
324 A. K. Mandal et al.

Table 3. (continued)

Monkeypox virus symptoms Initial research Secondary research

Lesions Weinstein et al. [52] Parker et al. [33]
Quenelle et al. [37] Weinstein et al. [52]
Petersen et al. [36] Sale et al. [41]
McCollm et al. [28] Parker et al. (2007) [34]
Kugelman et al. [19] Macneil et al. [26]
McCollm et al. [28]
Wawina-Bokalanga et al. [51]
Fever and exhaustion Petersen et al. [36] Weinstein et al. [52]
Quenelle et al.[37] Macneil et al. [26]
McCollm et al. [28] Sale et al. [41]
Kugelman et al. [19] Parker et al. [33]
Weinstein et al. [52] McCollm et al. [28]
Parker et al. (2007) [34]
Lymphadenopathy Weinstein et al. [52] Weinstein et al. [52]
Quenelle et al. [37] Sale et al. [41]
Petersen et al. [36] Parker et al. [33]
McCollm et al. [28] Macneil et al. [26]
Kugelman et al. [19] Parker et al. (2007) [34]
McCollm et al. [28]
Lam et al. [21]

3.2.5 Contact Tracing

‘Contact tracing’ is another application of PSO clustering that is essential for ‘healthcare
organizations and government officials to properly manage monkeypox infections. It
uses metadata from ‘blog posts, social media tags, metro smartcard data, automobile
records, credit card transactions, and other sources.‘ Extraction of important properties
from geolocation to ensure that a person is in a specific area at a specific time can used
to create a tracing system that can monitor people even if they are not using a tracking
device, but this method is not particularly accurate [54, 60]. Bui et al. [5] proposed an
approach for combining biological data with spatial analytical tools and GIS technology
to make it easier to gather and integrate heterogeneous data from various healthcare
information providers. The information needed to determine the data transmission on a
large scale could also be accessible by improving swarm clustering. The ability to “data
merge, share, and analyze” is still critical in a large data set.

3.2.6 Prediction and Forecasting

Researchers and academicians have proposed novel PSO Clustering-driven algorithms
to identify monkeypox virus infections in real-time. Infection control measures, such as
innovative animal isolation, aid in the prevention of epidemics in primate institutions and
those that import exotic pets. Diseases have observed in Asian monkeys combined with
African monkeys, so these animals should not house together. It is essential to prevent
 Machine Learning Approaches and Particle Swarm Optimization 325

the virus from infecting other fomites. Vaccination against the vaccinia virus protects
nonhuman primates (smallpox vaccine). This immunization, according to the study, may
also benefit other species such as prairie dogs. If you have had monkeypox, you should
avoid rats and other nonhuman primates, as well as any other animals that might be a
source of infection. The design of such a robust early warning system is made possible
by the combination of “Particle Swarm Optimization Clustering and ML” into a single
framework. In some countries, the surveillance of infectious populations is accelerating
with the use of big data analytics. For example, various governments have used Particle
Swarm Optimization Clustering and ML by employing the use of numerous surveillance
cameras, drones, and face recognition devices to monitor their citizens’ movements and
determine whether they are properly following isolation rules [17, 25]. A quick response
to the spread of monkeypox viruses by Taiwan’s healthcare system, which has already
tested for coronavirus forecasting and prediction, was crucial in reducing the infection
incidence. By carrying out an action plan for employing big data, Taiwan has led the
way in the field of human healthcare [3].

Table 4. Different parameters and Symptom alleviation of monkeypox virus

Monkeypox virus different parameters Monkeypox virus Symptom alleviation

Signs and symptoms Fever, headache, aches in the muscles, shivering,
blistering rash, enlarged lymph nodes [4, 56]
Specialty of monkeypox virus Infectious illness [6]
Complications Secondary infections include bronchopneumonia,
sepsis, encephalopathy, scarring, and eye infections,
among other conditions [6, 10]
Duration 2 to 4 weeks [7]
Usual onset 5 to 21 days after exposure [7]
Causes Monkeypox virus [6]
Variation West African,Central African(Congo Basin) [19]
Diagnostic Method DNA testing [1, 12]
Differential diagnosis Smallpox, Chickenpox [3, 12]
Protection ST-246 vaccination, hand washing, rash cover, PPE,
and staying away from sick people [15]
Diagnosis Supportive, antiviral, and vaccinia immune globulin
[27, 36]
Drug Tecovirimat [33]
Prognosis Most people recover [16, 35]
Frequency Not as uncommon as previously supposed [54]
Fatality Up to 10.6 percentage (Clade I, untreated) [53, 57]
and up to 3.6% (Clade II), respectively
326 A. K. Mandal et al.

3.2.7 Molecular Research and Drug Development

Particle Swarm Optimization Clustering can be highly useful in getting ideas for drug dis-
covery [22, 25, and 50]. Few attempts have previously made in this short period to develop
a suitable vaccination for example CMX001/Cidofovir /Brincidofovir/Tecovirimat [45].
Hence, the development of drug and vaccination programs against the monkeypox virus
is essential for controlling the endemic. The molecular and metabolic mechanisms behind
disease pathogenesis are critical in finding potential treatment alternatives for clinical
trials, which are required to achieve this understanding. The significance of viral-host
protein-protein interactions during viral infection, as well as their potential for thera-
peutic applications, has previously addressed in the context of ST-246 [15, 37]. Drug
reframing, also known as drug repositioning, is the process of discovering new phar-
macological or therapeutic uses for old, existing, or available pharmaceuticals. It has
demonstrated to be a successful strategy for discovering innovative drug compounds
[33]. The Review of Clinical Research and drug design for Human Monkeypox Virus
depicts above in Table 3.

4 Existing Challenges and Vaccination

There are still problems that need to overcome soon, although ‘ML and Particle Swarm
Optimization Clustering’ have demonstrated enormous potential in the fight against the
endemic monkeypox virus. As shown in Tables 3, 4, additional measures, such as ‘social
exclusion, mass screening, and testing, have already been implemented in response to the
rapid increase in the number of confirmed cases (both infected and dying). Individuals
have been encouraged to provide personal information such as medical records, GPS
locations, and travel data so that “ML and Particle Swarm Optimization Clustering” can
use to monitor the situation during the monkeypox virus endemic. People, on the other
hand, usually refuse to disclose such information out of fear of their privacy. This chal-
lenge may overcome by employing ML-based bio-inspired techniques. Governmental
agencies must also play a significant role in this field to harmonize the schemes used by
multiple groups. The most common issue that big data platforms and apps have is the
limited availability of standard datasets. The ML and PSO Clustering are applying for
the clinical data storage and processing in a systematic approach to enable simple and
safe access because further research on ‘ML and bio-inspired methods’ to ‘predictive,
diagnostic, and therapeutic ‘techniques against the Monkey Pox virus and other similar
pandemics will require the development of a vast cyber infrastructure to stimulate global
cooperation. Smallpox vaccine recipients may be immune to monkeypox. Younger indi-
viduals, on the other hand, are unlikely to have immunized against smallpox because the
disease was eliminate in most parts of the world in 1980. People who have had ‘smallpox
immunization’ should continue to take precautions to protect themselves and others.

5 Summary and Discussion

In the past, people believed that ‘Africa’ was the primary source of the uncommon
‘zoonotic virus known as monkeypox. Since May 2022, there have several instances
 Machine Learning Approaches and Particle Swarm Optimization 327

of monkeypox reported in several countries that do not often experience the disease,
including ‘India, the United States, the United Kingdom, Spain, Portugal, and Canada’.
Patients with monkeypox-related rash illnesses should investigate regardless of ‘gender
identity, natal sex, sex of sex partner, travel, or other special risk factors. Clinicians
should strictly adhere to infection control procedures and immediately notify their local
health department. This article summarizes an automated estimate of the monkeypox
virus that was identifying using ‘machine learning and the PSO clustering approach.
Machine learning techniques, which may employ the vast quantity of data provided by
various laboratory tests, clinical symptoms, and radiological scans to enhance diagnosis,
will be particularly valuable in this respect for improving screening and increasing
rapid diagnosis on a broad scale [55]. Machine learning techniques can predict ‘infected
patients’ using current quantitative and qualitative data [14, 25]. The results of this
strategy can help medical professionals make decisions during treatment and diagnosis,
which will help battle the endemic that represents in Fig. 2, 4, and 6. Presently no
effective treatments for people who are infect by this deadly virus. However, in extreme
cases, supportive treatments such as tecovirimat may employ.
The following are only a few of the health recommendations for the monkeypox
virus:

• Testing at Commercial and Public Health Laboratories

• ’Nucleic Acid Detection Kittest for Monkeypox virus’ (‘PCR-fluorescent probe
method-size-48T,Cat.no-PCR0101A, PCRL0101B’) [52]
• Infection Control Guidelines, Packaging, and Treatment of Monkeypox Medical
Waste
• Treatment of monkeypox with a smallpox vaccine [15, 38]

It can be challenging to distinguish monkeypox from other pox-like diseases because

of the “smallpox-like rash” and typical “prodromal symptoms.“ However, distinct mon-
keypox symptoms, including lymphadenopathy and lesions on the mucosa, palms, and
soles, are important clinical criteria. To preserve the infection in the human population, it
is believe that “recurrent animal reintroductions of the monkeypox viruses” are required
[14]. Typical hypotheses include:

• Under certain temperature and light conditions, the monkeypox virus can survive
longer without a host [19]
• Deforestation and flooding may result in more habitats for species that carry the
monkeypox virus, increasing the likelihood that those species will come into contact
more frequently and pose a greater risk of transmission.
• The monkeypox virus and its reservoirs may be able to increase their geographic range
because of the rainforest’s expansion brought on by warmer and more humid weather,
thereby hastening the spread of the virus [29].

If the monkeypox virus spreads outside of ‘Africa,’ the global public health impacts
will be disastrous. As a result, there would be no way to stop the virus while its geographic
spread is still limited [19]. Due to inadequate resources and infrastructure, insufficient
‘diagnostic resources, systemic barriers in war conditions’ [19, 54], and a lack of clinical
328 A. K. Mandal et al.

diagnosis of monkeypox [19, 28], monkeypox monitoring is challenging. Knowing the

primary ‘clinical characteristic’ features of monkeypox will assist in ‘clinical detection.
Furthermore, the development of simple ‘diagnostic algorithms’ combined with simple,
rapid testing [26] may help in the detection and hence control of a monkeypox outbreak
[54].

6 Conclusion

As illustrated in Table 2, the monkeypox virus had affected around 90 countries world-
wide as of August 2022, with ten cases and one fatality in India. People’s social and
economic security has suffered substantially because of this horrible illness. To tackle the
monkeypox virus challenge, this article discusses the most recent cutting-edge applica-
tions of ‘machine learning and particle swarm optimization cluster. Various researches
involving big data have suggested that machine learning (ML) plays a crucial role.
Advanced analytical techniques have been successful in both predicting the disease’s
emergence and identifying its symptoms and indications. Furthermore, the application of

Table 5. Supportive protection against of monkeypox

Component of Protection Symptoms/Signs Prevention/ Treatment

Protection of the skin and Rashes on the skin Mupironic Acid/Fucidin
mucous membranes [35, 54, Clean with a basic antiseptic
55] If there is an extensive lesion,
cover it with a light dressing
Avoid touching or scratching
the lesions
Appropriate systemic
antibiotics may be investigated
in the case of a subsequent
infection
Urethral lesions Sitz bathing
Oral ulcers Saline gargles with warm
water and an oral topical
anti-inflammatory gel
Conjunctivitis Usually self-defeating
If signs worsen or there is
discomfort or vision loss,
consult an ophthalmologist
Disturbances
Nutritional assistance and Dehydration may occur along Promote oral fluids or ORS
rehydration therapy [8, 28] with vomiting, nausea, Intravenous fluids if
diarrhea, or loss of appetite necessary and a balanced
Nutrient-rich diet
(continued)
 Machine Learning Approaches and Particle Swarm Optimization 329

Table 5. (continued)

Component of Protection Symptoms/Signs Prevention/ Treatment

Symptom [35, 55] Increase body temperature Tepid sponging and, if
necessary, Paracetamol
Itching/Pruritus Applying calamine lotion
orally
Antihistamines
Nausea and vomiting Consider antiemetic
Headache/ malaise Proper hydration and
paracetamol

‘machine learning-based big data analytics inspired by biological systems’ has consid-
erably improved ‘contact tracing, molecular analysis, and drug development. However,
the lack of standard data sets and privacy issues restrict the efficiency of ML and Particle
Swarm Optimization Clustering in this scenario. A single library of all the data related
to the monkeypox virus is required to determine the efficacy of ML-based bio-inspired
approaches in eliminating the endemic.

Acknowledgements. This work was supported by Dr.Satchidananda Dehuri, Professor of Com-

puter Science (Erstwhile P. G. Department of Information and Communication Technology), Fakir
Mohan University would like to thank SERB, Govt. of India for financial support under Teachers’
Associateship for Research Excellence (TARE) fellowship vide file no. TAR/2021/000065 for the
period 2021–2024.

CRediT Authorship Contribution Statement:. Akshaya Kumar Mandal (https://orcid.org/

0000-0003-0955-4096): The general idea of literature survey, Data correction, Resources,
Writing – review & editing – original draft.
Pankaj Kumar Deva Sarma (https://orcid.org/0000-0002-9748-3787): The general idea of
literature survey, Research Direction, Resources, Review, and editing.
Satchidananda Dehuri (https://orcid.org/0000-0003-1435-4531): The general idea of literature
survey, Research Direction, Resources, Review, and editing.

References
1. “About Monkeypox | Monkeypox | Poxvirus | CDC”. 2021–11–22. Archived from the original
on 2022–05–10, 2022–04–27
2. Anderberg, M.R.: The broad view of cluster analysis. Cluster Analysis for Appl. 1(1), 1–9
(1973)
3. Bauch, C.T., Lloyd-Smith, J.O., Coffee, M.P., Galvani, A.P.: Dynamically modeling SARS
and other newly emerging respiratory illnesses: past, present, and future. Epidemiology 791–
801 (2005)
4. Breman, J.G.: Monkeypox: an emerging infection for humans? Emerging infections 4, 45–67
(2000)
330 A. K. Mandal et al.

5. Bui, D.T., Bui, Q.T., Nguyen, Q.P., Pradhan, B., Nampak, H., Trinh, P.T.: A hybrid artificial
intelligence approach using GIS-based neural-fuzzy inference system and particle swarm
optimization for forest fire susceptibility modeling at a tropical area. Agric. For. Meteorol.
233, 32–44 (2017)
6. Bunge, E.M., et al.: The changing epidemiology of human monkeypox—a potential threat?
a systematic review. PLoS Negl. Trop. Dis. 16(2), e0010141 (2022)
7. Centers for Disease Control and Prevention. Interim clinical guidance for the treatment of
monkeypox (2022)
8. Centers for Disease Control and Prevention.. Monkeypox and smallpox vaccine guid-
ance. Ultimo aggiornamento 2 (2022)
9. Davenport, T., Kalakota, R.: The potential for artificial intelligence in healthcare. Future
Healthcare J. 6(2), 94 (2019)
10. Ennab, F., Nawaz, F.A., Narain, K., Nchasi, G., Essar, M.Y.: Rise of monkeypox: Lessons
from COVID-19 pandemic to mitigate global health crises. Annals of Medicine Surgery 79,
104049 (2022)
11. https://en.wikipedia.org/wiki/Monkeypox
12. Huggett, J.F., French, D., O’Sullivan, D.M., Moran-Gilad, J., Zumla, A.: Monkeypox: another
test for PCR. Eurosurveillance 27(32), 2200497 (2022)
13. Jabeen, C., Umbreen, G.: Monkeypox transmission, need and treatment of humans with an
antiviral drug. Int. J. Social Sciences Manage. 4(2), 77–79 (2017)
14. Jain, N., Lansiaux, E., Simanis, R.: The new face of monkeypox virus: an emerging global
emergency. New Microbes and New Infections (2022)
15. Jordan, R., et al.: Single-dose safety and pharmacokinetics of ST-246, a novel orthopoxvirus
egress inhibitor. Antimicrob. Agents Chemother. 52(5), 1721–1727 (2008)
16. Keasey, S., et al.: Proteomic basis of the antibody response to monkeypox virus infection
examined in cynomolgus macaques and a comparison to human smallpox vaccination. PLoS
ONE 5(12), e15547 (2010)
17. Kennedy, J.: Stereotyping: Improving particle swarm performance with cluster analysis. In:
Proceedings of the 2000 congress on evolutionary computation. CEC00 (Cat. No. 00TH8512),
Vol. 2, pp. 1507–1512. IEEE (2000)
18. Ko, Y., Mendoza, R., Mendoza, V.M., Seo, Y.B., Lee, J., Jung, E.: Estimation of monkeypox
spread in a non-endemic country considering contact tracing and self-reporting: a stochastic
modeling study. medRxiv (2022)
19. Kugelman, J.R., et al.: Genomic variability of monkeypox virus among humans, democratic
republic of the congo. Emerg. Infect. Dis. 20(2), 232 (2014)
20. Kumari, R., Gupta, N., Kumar, N.: Cumulative histogram based dynamic particle swarm
optimization algorithm for image segmentation. Indian J. Computer Science Eng. 11(5),
557–567 (2020)
21. Lam, H.Y.I., Guan, J.S., Mu, Y.: In silico repurposed drugs against monkeypox virus. bioRxiv
(2022)
22. Lee, R.C.: Clustering analysis and its applications. In: Advances in information systems
science, pp. 169–292. Springer, Boston, MA (1981). https://doi.org/10.1007/978-1-4613-988
3-7_4
23. Li, X., Fang, Z.: Parallel clustering algorithms. Parallel Comput. 11(3), 275–290 (1989)
24. Likos, A.M., et al.: A tale of two clades: monkeypox viruses. J. Gen. Virol. 86(10), 2661–2672
(2005)
25. Luo, Y., Szolovits, P., Dighe, A.S., Baron, J.M.: Using machine learning to predict laboratory
test results. Am. J. Clin. Pathol. 145(6), 778–788 (2016)
26. MacNeil, A., et al.: Transmission of atypical varicella-zoster virus infections involving palm
and sole manifestations in an area with monkeypox endemicity. Clin. Infect. Dis. 48(1), e6–e8
(2009)
 Machine Learning Approaches and Particle Swarm Optimization 331

27. Maksyutov, R.A., Gavrilova, E.V., Shchelkunov, S.N.: Species-specific differentiation of var-
iola, monkeypox, and varicella-zoster viruses by multiplex real-time PCR assay. J. Virol.
Methods 236, 215–220 (2016)
28. McCollum, A.M., Damon, I.K.:. Human monkeypox. Clinical Infectious Diseases 58(2),
260–267 (2014)
29. Nakazawa, Y., et al.: A phylogeographic investigation of African monkeypox. Viruses, 7(4),
2168-2184 (2015)
30. Nolen, L.D., et al.: Extended human-to-human transmission during a monkeypox outbreak
in the Democratic Republic of the Congo. Emerg. Infect. Dis. 22(6), 1014 (2016)
31. O’Shea, J.: Interim guidance for prevention and treatment of monkeypox in persons with HIV
infection—United States, August 2022. MMWR. Morbidity and Mortality Weekly Report 71
(2022)
32. Osorio, J.E., Yuill, T.M.:. Zoonoses. Encyclopedia of Virology 485 (2008)
33. Parker, S., Handley, L., Buller, R M.: Therapeutic and prophylactic drugs to treat orthopoxvirus
infections (2008)
34. Parker, S., Nuara, A., Buller, R.M.L., Schultz, D.A.: Human monkeypox: an emerging
zoonotic disease (2007)
35. Petersen, B.W., Damon, I.K.: Smallpox, Monkeypox, and Other Poxvirus Infec-
tions. Goldman-Cecil Medicine. 26th ed. Elsevier, Philadelphia, PA (2020)
36. Petersen, E., et al.: Human monkeypox: epidemiologic and clinical characteristics, diagnosis,
and prevention. Infect. Dis. Clin. 33(4), 1027–1043 (2019)
37. Quenelle, D.C., et al.: Efficacy of delayed treatment with ST-246 given orally against systemic
orthopoxvirus infections in mice. Antimicrob. Agents Chemother. 51(2), 689–695 (2007)
38. Ray, S., Turi, R.H.: Determination of number of clusters in k-means clustering and applica-
tion in colour image segmentation. In: Proceedings of the 4th International Conference on
Advances in Pattern Recognition and Digital Techniques, Vol. 137, p. 143 (1999)
39. Reed, K.D., et al.: The detection of monkeypox in humans in the Western Hemisphere. New
England Journal of Medicine 350(4), 342-350 (2004)
40. Reynolds, M.G., Damon, I.K.: Outbreaks of human monkeypox after cessation of smallpox
vaccination. Trends Microbiol. 20(2), 80–87 (2012)
41. Sale, T.A., Melski, J.W., Stratman, E.J.: Monkeypox: an epidemiologic and clinical compar-
ison of African and US disease. J. Am. Acad. Dermatol. 55(3), 478–481 (2006)
42. Shchelkunov, S.N., et al.: Analysis of the monkeypox virus genome. Virology 297(2), 172–194
(2002)
43. Shi, Y., Eberhart, R.: A modified particle swarm optimizer. In: 1998 IEEE International Con-
ference on Evolutionary Computation Proceedings. IEEE World Congress on Computational
Intelligence (Cat. No. 98TH8360), pp. 69–73. IEEE (1998)
44. Sklenovska, N., Van Ranst, M.: Emergence of monkeypox as the most important orthopoxvirus
infection in humans. Front. Public Health 6, 241 (2018)
45. Smee, D.F., Sidwell, R.W., Kefauver, D., Bray, M., Huggins, J.W.: Characterization of wild-
type and cidofovir-resistant strains of camelpox, cowpox, monkeypox, and vaccinia viruses.
Antimicrob. Agents Chemother. 46(5), 1329–1335 (2002)
46. Sreenivas, S.: Monkeypox: What to Know. webmd.com. WebMD (2022)
47. Stittelaar, K.J., et al.: Antiviral treatment is more effective than smallpox vaccination upon
lethal monkeypox virus infection. Nature 439(7077), 745–748 (2006)
48. Torcate, A.S., Barbosa, J.C.F., de Oliveira Rodrigues, C.M.: Utilizando o learning analytics
com o k-means para análise de dificuldades de aprendizagem na educação básica. In: Anais
do XXVI Workshop de Informática na Escola, pp. 31–40. SBC (2020)
49. UK, G. Research and analysis Qualitative assessment of the risk to the UK human population
of monkeypox infection in a canine, feline, mustelid, lagomorph or rodent UK pet
332 A. K. Mandal et al.

50. Watkins, X., Garcia, L.J., Pundir, S., Martin, M.J., UniProt Consortium: ProtVista: visualiza-
tion of protein sequence annotations. Bioinformatics 33(13), 2040-2041 (2017)
51. Wawina-Bokalanga, T., et al.: An accurate and rapid Real-time PCR approach for human
Monkeypox virus diagnosis. medRxiv (2022)
52. Weinstein, R.A., Nalca, A., Rimoin, A.W., Bavari, S., Whitehouse, C.A.: Reemergence of
monkeypox: prevalence, diagnostics, and countermeasures. Clin. Infect. Dis. 41(12), 1765–
1771 (2005)
53. World Health Organization. Multi-country monkeypox outbreak: situation update. 2022–06–
04)[2022–06–07] (2022). https://www.who.int/emergencies/disease-outbreak-news/item/
2022-DON390
54. World Health Organization. Clinical management and infection prevention and
control for monkeypox: interim rapid response guidance, 10 June 2022 (No.
WHO/MPX/Clinical_and_IPC/2022.1). World Health Organization (2022)
55. World Health Organization. Surveillance, case investigation and contact tracing for monkey-
pox: interim guidance, 22 May 2022 (No. WHO/MPX/Surveillance/2022.1). World Health
Organization (2022)
56. Yuen, C.Y., Dodgson, J.E., Tarrant, M.: Perceptions of Hong Kong Chinese women toward
influenza vaccination during pregnancy. Vaccine 34(1), 33–40 (2016)
57. https://worldhealthorg.shinyapps.io/mpx_global/
58. https://www.istockphoto.com/photos/monkeypox
59. Abdelhamid, A.A., et al.: Classification of monkeypox images based on transfer learning and
the Al-Biruni Earth radius optimization algorithm. Mathematics 10(19), 3614 (2022)
60. Patel, M., Surti, M., Adnan, M.: Artificial intelligence (AI) in Monkeypox infection
prevention. Journal of Biomolecular Structure and Dynamics, 1–5 (2022)
 Machine Learning-Based Hybrid Feature
Selection for Improvised Network Intrusion
Detection

Pritimayee Satapathy(B) and Prafulla Kumar Behera

Department of Computer Science and Applications, Utkal University, Vani Vihar,

Bhubaneswar 751004, Odisha, India
[email protected]

Abstract. With the expeditious development of network-based applications,

unusual threats erupt. Therefore, distinctive security components are required
so that they can boost the existing security mechanism. Network administra-
tors depend primarily on intrusion detection programs to spot such attacks. But
nowadays, Machine Learning-based approaches are overwhelmingly used towards
intrusion detection, where the system is made to learn from information so that
it can distinguish between anomaly and normal traffic. In this research work,
we comprehensively discuss some of the established machine learning classifiers
concerning network traffic invasions. Using the NSL-KDD dataset, we evalu-
ated five ML classifiers and highlighted their adequacy by employing a rigorous
experimental assessment.
Within the scope of this study, we propose a hybrid feature selection tech-
nique with the goal of enhancing the prediction performance of anomaly-based
intrusion detection systems (IDS). We have used information gain (IG), gain ratio
(GR) and chi-squared (CS) attribute evaluation methods for the selection of rel-
evant features. Thereafter, we have used mathematical set theory intersection to
extract the best features among the 3 evaluation methods mentioned above. Five
different machine learning algorithms (MLA): REPTree, LogitBoost, J48, NB,
and BayesNet were implemented to generate 32 machine learning models based
on each MLA, resulting in a total of 160 models. Akaike Information Criterion
(AIC) was used to select the best model out of the 160 models.

Keywords: IDS · Anomaly detection accuracy · Feature selection · Machine

learning · Classification algorithms · Attribute evaluation · NSL-KDD dataset ·
WEKA

1 Introduction
Due to the internet’s growing popularity and extensive use, security is becoming increas-
ingly important to maintain a stable and secure network, free from unauthorized activity.
The term “intrusion” refers to any malicious attempt to breach the security of a network
by compromising its secrecy, integrity, or accessibility. A variety of approaches, meth-
ods, and algorithms have been developed to capture and spot the network intruder. An

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 333–347, 2022.
https://doi.org/10.1007/978-3-031-23233-6_25
334 P. Satapathy and P. K. Behera

example of such a mechanism is the IDS, which keeps tabs on and assesses all the
activities happening on a network in real time. Thus, IDS is a crucial tool that monitors
networks and alerts the network administrator to any suspicious or anomalous activity,
protecting the network from both existing and emerging threats that put its resources at
risk [1].
IDSs are further categorized into three types based on the nature of the attacks:
signature-based, anomaly-based, and hybrid. Signature- or misuse-based IDSs look for
patterns in network traffic, such as byte sequences, that are used by known threats.
Consequently, it cannot identify zero-day threats. IDSs that are based on anomalies
are deployed to identify attacks that were previously unidentified by generating models
based on the system’s behaviour. The primary challenge with anomaly-based IDS is
the rise of false alarm rates. The hybrid IDS combines signature and anomaly-based
IDS [2]. Typically, IDS employs machine learning (ML) methods to process enormous
volumes of high-dimensional data to identify intruders, develop a trustworthy model of
activity, and evaluate new behaviours in the model. Even though there are a number of
IDS that employ machine learning, additional research is required to improve accuracy
and reduce false positive rates [3].
The pre-processing stage is accelerated and detection accuracy is increased through
feature selection, a crucial component of an ML-based solution. Because the IoT ecosys-
tem is always changing, the detection precision of anomaly-based IDSs is regarded as
the biggest issue. The goal of this study is to acquire resilient performance for ML-based
IDS in the varied IoT ecosystem [4] by proposing a novel feature selection approach.

1.1 Related Work

Numerous studies have been conducted over the past decade to investigate intrusion
detection based on NSL-KDD dataset, aiming to improve prediction accuracy and
decrease the false positive rate. Machine learning (ML) and deep learning (DL) have
been the subjects of a variety of studies aimed at enhancing IoT safety [15–17]. Uti-
lizing the Weka tool’s 10-fold cross-validation strategy [5], Alabdulwahab et al. [3]
tested the effectiveness of 6 supervised classifiers [Reduced Error Pruning Tree (REP-
Tree), Sequential Minimal Optimization (SMO), LogitBoost, BayesNet, Radial Basis
Function (RBF), and Naive Bayes Tree (NBTree)] for intrusion detection on the full
NSL-KDD training dataset. Compared to other classifiers, REPTree’s model-building
time was the shortest, and its detection accuracy was the highest (99.73%). Albulayhi K.
et al. [4] used the IoT intrusion dataset 2020 (IoTID20) and the NSL-KDD datasets to test
4 ML-based algorithms: Bagging, Multilayer Perceptron, J48, and K-Nearest Neighbour
(called IBk or instance-based learner in the Weka Tool) for detecting intrusion using a
hybrid feature selection approach that uses both IG and GR. The J48 algorithm achieved
the highest level of accuracy (99.70%) in comparison to other methods when applied
to the NLS-KDD dataset. Using the NSL-KDD dataset, Mahfouz et al. [6] evaluated
the efficacy of six different ML techniques (NB, LR, MLP, SMO, IBK, J48) in the con-
text of intrusion detection. First, they used the InfoGainAttributeEval algorithm’s ranker
method to figure out the order of attributes. Then, they used CVParameterSelection to
find the best values for the classifier’s hyper-parameters. With 97.89% precision, they
found that the J48 method efficiently classified the dataset.
 Machine Learning-Based Hybrid Feature Selection 335

Table 1. An overview of several methods for selecting features

Feature Selection Machine Learning algo- Datasets Performance Metrics Reference

Method rithm
InfoGain and 1. Bagging IoTID20 1. FP Rate [4]
GainRatio 2. Multilayer Perceptron NSL-KDD 2. Precision
3. J48 3. Recall
4. IBk 4. F-Measure
5. ROC Area
6. Accuracy
CfsSubsetEval 1. REPTree NSL-KDD 1. TP Rate [3]
InfoGain 2. SMO 2. FP Rate
3. LogitBoost 3. Precision
4. BayesNet 4. F-measure
5. RBF 5. ROC Area
6. NBTree 6. Specificity
7. Sensitivity
InfoGain 1. Naive Bayes NSL-KDD 1. Accuracy [6]
2. Logistic Regression 2. TPR
3. Multilayer Perceptron 3. FPR
4. SMO 4. Precision
5. KNN 5. Recall
6. J48 6. F-measure
7. ROC area
Subset evaluator 1. Naive Bayes CIC IDS 2017 1. Accuracy [7]
feature selection 2. Decision Tree 2. Specificity
method (wrapper 3. Recall
technique) 4. Precision
5. F1-score
6. Error rates
7. Response time.

The studies presented in Table 1 were limited to one or a combination of two feature
selection methods. Moreover, statistical model efficiency, such as AIC comparison, was
not done in any of these studies. Since an optimized method for selecting features is a
key part of machine learning, more research is needed to find the best ML classifier with
a high rate of correct detection and a low rate of false positives. Based on this literature
review, we propose a more shrewd IDS framework that eliminates redundant features
from the dataset in the pre-processing stage. Our research focuses on the application of
3 well-known individual feature reduction methods, namely information gain (IG), gain
ratio (GR), and chi-square (CS) methods to get the optimal feature set. Using the concept
of mathematical set theory intersection, a unique hybrid feature selection approach is
proposed which will rank the best minimum relevant features. The proposed IDS will
be a combination of 3 feature selection modules viz. IG, GR, and CS have been used for
dimension reduction which may help the classifier to improve efficiency over alternative
methods.

2 Materials and Methods

Data security and privacy are key concerns, and IDSs serves as the first line of protection.
Intelligent IDS are built using a variety of rule-based methods or ML frameworks. The
flow diagram describing the methods followed in our proposed work is depicted in
336 P. Satapathy and P. K. Behera

Fig. 1. In this section, we have described and explained the dataset, ML classifiers,
feature selection and reduction approaches, performance metrics, and model evaluation
methods used in our research.

Fig. 1. Machine learning-based intrusion detection with a high-performing hybrid feature

reduction approach

2.1 Dataset Description

Analysts have compiled several network traffic datasets to make it simpler to test dif-
ferent algorithms for discovering intrusions, allowing for a more thorough comparison
of available intrusion detection approaches. These datasets may be accessible either
publicly or privately or through the use of network simulation, which may be used to
monitor a variety of traffic events and monitor them. The vast majority of these datasets
were assembled with the assistance of a variety of tools that assisted in capturing traffic
and keeping track of various traffic incidents. In our research, we have used the Network
Security Laboratory Knowledge Discovery and Data (NSL-KDD) dataset, the well-
known dataset for benchmarking IDSs. The well-known KDDcup99 dataset, which is
exploited extensively in the field of IDS development, served as the basis for the creation
of the NSL-KDD dataset, which is an upgraded version of the original. Each record in
 Machine Learning-Based Hybrid Feature Selection 337

this dataset is either considered a normal operation or an attack based on the presence or
absence of 41 attributes and a single class attribute [8]. The NSL-KDD datasets used for
training and testing include 21 and 37 types of threats, respectively. Denial-of-service
(DoS), user-to-root (U2R), root-to-local (R2L), and probe attacks are the four main types
of network intrusions. Train and test data for the NSL-KDD dataset [9] are subdivided
by attack type as shown in Table 2. Table 3 lists the subcategories of the four major
attack types in both the training and testing datasets.

Table 2. The NSL-KDD core dataset

Dataset Total Normal Denial-of-service Probe Remote to local User to root

KDDTrain + 125973 67343 45927 11656 995 52
KDDTest + 22544 9711 7456 2421 2756 200

Table 3. Intrusions into the NSL-KDD dataset

Attacks in dataset Attack type

Denial-of-service Teardrop, Back, Worm, Land, Apache2, Neptune, Udpstorm, Pod,
Processtable, Smurf, Mailbomb
Probe Satan, Portsweep, IPsweep, Saint, Nmap, Mscan
Remote to local Xlock,Guess_password, Named, Ftp_write, Sendmail, Imap, Httptunnel,
Phf, Snmpgetattack, Warezmaster, Snmpguess, Multi hop, Xsnoop
User to root Buffer_overflow, Ps, Loadmodule, Xterm, Perl, Rootkit, Sqlattack

2.2 Data Pre-processing

Real-world traffic data primarily contains noise, missing values, and categorical variables
that prevent it from being directly employed in machine learning models. Therefore, data
pre-processing is an essential prerequisite step for improving the accuracy and efficiency
of a machine learning model. Out of the 41 attributes of the NLS-KDD dataset, two
single-valued attributes viz. num_outbound_cmds and is_ host_login had been removed.
Symbolic attributes like protocol type (3 different symbols; viz. TCP, UDP, ICMP), flag
(11 different symbols; viz. OTH, RSTOS0, SF, SH, RSTO, S2, S1, REJ, S3, RSTR,
S0), and service (70 different symbols) were converted to binary vector using one hot
encoding method, which resulted in a total of 120 attributes (excluding the class attribute).
Thus, the dataset obtained from the pre-processing stage is further used in the next stage
for feature reduction, as shown in Fig. 1.

2.3 Dimension Reduction

Feature selection is crucial to the development of a better classification model because it
frequently decreases the complexity of training and enhances the model’s performance
338 P. Satapathy and P. K. Behera

matrix. In our research, we implemented IG, GR, and CS attribute evaluation with the
help of the ranker method to rank the one-hot encoded attributes of the NSL-KDD data
set. The lowest 10, 20, 30, 40, 50, 60, 70, and 80 ranked attributes were sequentially
filtered from each attribute-ranked dataset, resulting in the construction of 8 datasets for
each of the three attribute evaluation methods containing 111, 101, 91, 81, 71, 61, 51,
and 41 features, respectively, ultimately having a total of 8 X 3 = 24 datasets. The hybrid
feature selection approach used by us utilizes three filter-based approaches. To design
the hybrid feature selection approach, we used mathematical set theory to consider the
attributes that are common (intersection) between the individual sequentially filtered
datasets of the three attribute evaluation methods, resulting in a total of eight datasets.
Thus, a total of 32 datasets (24 + 8) were generated.

Information Gain Feature Selection: The importance of an attribute is determined by

the Information Gain Attribute Analysis, which does this by calculating the information
gain in relation to the class [14].

Gain Ratio Feature Selection: The gain ratio can be considered of as an extension of
the information gain that was addressed earlier. It emphasizes selecting characteristics
that have a high number of values in an effort to minimize the Information Gain that
occurs as a result of its use. Therefore, we are able to say that the Gain Ratio feature
evaluation is more accurate in certain conditions, specifically those in which the data are
well-ordered and there is no redundancy [14].

Chi-Square Feature Selection: The chi-squared test is a numerical test that measures
the variation from the distribution that would be predicted if the feature event were
independent of the class value [18].

2.4 Model Evaluation

Several machine-learning approaches were employed to monitor traffic data for various
irregularities. The majority of these approaches (classifiers) detect the abnormality by
taking into account deviations from basic common network data. Typically, these models
are trained using a collection of traffic data gathered over some time. There are three
types of ML anomaly detection methods: supervised, unsupervised, and semi-supervised
learning. In our experiments, the complete NSL-KDD dataset along with the 32 datasets
generated (as described in Sect. 3.3) was evaluated based on 10 Fold Cross Validation
using five well-known ML classifiers named as REPTree, LogitBoost, J48, NB, and
BayesNet. We also train the classifiers without any feature reduction methods. This
resulted in the generation of a total of (32 X 5 + 1 X5) = 165 models.

REPTree: Reduced error pruning is a quick decision tree learner. A decision/regression

tree is created by REPTree depending on information gained or by lowering variance
[10].

LogitBoost: By reducing logistic loss, it may be applied to additive tree regression as

well as binary or multi-class classification. It is suitable for managing noisy and outlier
data when compared to other AdaBoost classifiers [3].
 Machine Learning-Based Hybrid Feature Selection 339

BayesNet: It is a frequently used technique that bases its operation on the fundamen-
tal Bayes theorem and builds a Bayesian network after determining the conditional
probability for each node [10].

J48: The C4.5 algorithm is used to create a decision tree (DT), which is a part of
supervised learning methodology and is referred to as J48 in the Weka tool [4].

NB: The Bayes Theorem serves as the foundation for the Naive Bayes (NB) classifier,
one of the most traditional methods for solving classification issues. NB does not discard
the features’ weak predictors, in contrast to other machine learning algorithms. Each
feature’s information is considered, and a probabilistic model is constructed [7].

2.5 Evaluation Metrics

To assess how well the suggested IDS performs, we employed the evaluation matrices
below:

Confusion Matrix: A confusion matrix is a visual aid that serves as the foundation for
computing all other evaluation criteria. It has the following four values: false negative
(FN), true negative (TN), false positive (FP) and true positive (TP) [3].

TP: When the IDS classifies an instance as an attack and the instance is in fact an
attack, this is considered a true positive result.

FP:When the IDS classifies an instance as an attack whereas the instance is actually
found to be normal, this is considered a false positive result. It is often referred to
as the false alarm rate

TN: When the IDS classifies an instance as normal and the instance is actually
found to be normal.

FN: Instances that should be considered attack but are misclassified as normal

Sensitivity: The percentage of true positives that are accurately determined is evaluated
by sensitivity.
Sensitivity = (TP) / (TP + FN) [6].

Specificity: The percentage of actual negatives correctly detected is regarded as a

measure of specificity.
Specificity = TN / (TN + FP) [6].

Accuracy: The result is accurate when the total amount of objects is divided by the
fraction of correctly detected components, which may include either true positives or
true negatives.
Accuracy = (TP + TN)/(TP + TN + FP + FN) [3]
340 P. Satapathy and P. K. Behera

F-measure: The F-measure is used to evaluate the efficacy of binary classification

functions, which is the harmonic mean of precision and recall. Precision is the proportion
of predicted positive outcomes that actually are in the class the classifier has predicted.
Precision = (TP) / (TP + FP)
The percentage of positive cases that the classifier correctly predicted is measured
by recall.
Recall = (TP)/(TP + FN)
F-measure = 2 × (Precision × Recall) / (Precision + Recall [6]

Matthew’s Correlation Coefficient (MCC): The MCC is an useful statistic for eval-
uating the efficiency of binary classification models developed from unbalanced data. It
accepts a number between 1 and + 1, where + 1 stands for a perfect prediction, 0 for
an average random prediction, and √ -1 for an inverse prediction.
MCC = (TP × TN-FP × FN)/ ((TP + FN) (TP + FP) (TN + FP) (TN + FN))
[11]

Root Means Squared Error (RMSE): The percentage is determined by dividing the
difference between the predicted and real values by the total number of predictions.
This indicates that the RMSE is most useful when dealing with significant mistakes, the
majority of which
√ are unfavourable [12].
RMSE = ((p1- a1 )2 +…… + (pn − an )2 ))/n)
where a1, a2, and an are the real values and p1, p2, and pn are the predicted values
for test instances.

Kappa Statistic: The agreement between expected values and values that could be
expected by chance is measured using the kappa statistic. When there is no agreement
beyond chance, its values tend to be zero, and it approaches 1.0 when there is a very
high statistical association between the projected and actual category labels. High kappa
statistics and AUC values, as well as a low RMSE value, are characteristics of a successful
generalizable prediction model [11].

Akaike’s Information Criterion (AIC): AIC is used to choose the best models.
AIC = n log (∂ˆ 2 ) + 2(m + 1)
The model that achieves the lowest AIC is the best option, where n is the quantity of
training data, ∂ˆ2 =  ∂ 2 /n is a mean squared error (MSE) between what was intended
to be achieved and what was actually achieved, and X = m + 1 is the number of total
parameters [13].

3 Results and Discussion

Here, we present our experimental results utilising several methodologies for machine
learning models and feature selection strategies that, in terms of accuracy with a minimal
amount of features, are most suited for the IoT environment. The goal of this experiment
is to see how well five classifiers (REPTree, LogitBoost, BayesNet, NB, and J48) can
find intrusions with high accuracy while only using a few features. For this, we’ve
 Machine Learning-Based Hybrid Feature Selection 341

used the NSL-KDD [8] training and test dataset together with the Weka ML software
[5]. The examination was carried out on a computer running Microsoft Windows 11
and equipped with an AMD Ryzen 7 3700U processor, a Radeon Vega Mobile Gfx
graphics card clocked at 2.30 GHz, and 8 GB of RAM. Using the Weka tool’s 10-fold
cross-validation, we compared three different feature reduction strategies on the entire
NSL-KDD dataset for training purposes that was created following the pre-processing
stage (described in Sect. 2.2) and scored each feature. This phase’s goal is to not only
identify many useful qualities but also to eliminate the unimportant ones. In the second
step, we used 10-fold cross-validation to assess each classifier’s performance on 32
datasets that are subsets of the NSL-KDD dataset and were created using various feature
selection methods.

3.1 Classifier Comparison

The results of the five classifiers using 10-fold cross-validation on the NSL-KDD dataset
are shown in Table 4 without the use of any feature reduction techniques. The research
reveals that REPTree had a detection accuracy of 99.78% and that the J48 classifier had
the highest rate of successful detection of 99.84%. Even though no feature reduction
strategies were employed, the results of the investigation make it abundantly evident that
the REPTree and J48 classifiers achieved the highest levels of accuracy when compared
to those of the BayesNet, NB, and LogitBoost classifiers.

Table 4. Evaluation of the five different classifiers with a 10-fold cross-validation procedure

Classifier Train Set Test Set TPR* FPR* Precision Recall F-Measure MCC
Accuracy Accuracy
BayesNet 96.59% 74.78% 0.966 0.038 0.967 0.966 0.966 0.933
J48 99.84% 81.60% 0.998 0.002 0.998 0.998 0.998 0.997
LogitBoost 97.10% 74.72% 0.971 0.030 0.971 0.971 0.971 0.942
REPTree 99.78% 78.72% 0.998 0.002 0.998 0.998 0.998 0.996
NB 90.09% 76.19% 0.901 0.103 0.902 0.901 0.901 0.801
TPR: True Positive Rate
FPR: False Positive Rate

3.2 Analysing the Current State of Knowledge in Comparison to the Approach

that Was Suggested

Our method was compared to four other published methods that also employed the NSL-
KDD dataset with binary classification. The goal of this comparison was to validate the
efficacy of the strategy that we have suggested. Analysis revealed that the models gen-
erated through our approach are: (i) based on the dataset with 41 features ranked by
the ChiSquaredAttributeEval method with J48 classifier, shown to have better accuracy
342 P. Satapathy and P. K. Behera

(99.85%) than the model generated by Albulayhi et. al. (99.70%) [4]. (ii) dataset with
41 features ranked by InfoGainAttributeEval method with BayesNet classifier, shown to
have better accuracy (96.82%) than the model generated by Alabdulwahab et.al (94.89%)
[3]. (iii) dataset with 41 features ranked by ChiSquaredAttributeEval method with Log-
itBoost classifier, shown to have better accuracy (97.12%) than the model generated by
Alabdulwahab et.al (96.66%) [3]. (iv) dataset with 41 features ranked by ChiSquaredAt-
tributeEval method with REPTree classifier, shown to have better accuracy (99.77%) than
the model generated by Kumar et.al (97.02%) [14]. (v) dataset with 56 features ranked
by hybrid feature selection approach with NB classifier, shown to have better accuracy
(91.30%) than the model generated by Mahfouz et.al (90.41%) [6]. In order to provide
evidence for our assertion, we have included this comparison analysis of the NSL-KDD
dataset in Table 5.

Table 5. Comparison of our analysis with published methods

Classifier Reference dataset

Current analysis dataset Reference Current
description description model Accuracy Analysis model
Reference FRM* FRM* No. of Features Accuracy

J48 [4] UMF* CS 41 99.70% 99.85%

BayesNet [3] IG IG 41 94.89% 96.82%
LogitBoost [3] IG CS 41 96.66% 97.12%
REPTree [14] IG CS 41 97.02% 99.77%
NB [6] IG IMF 56 90.41% 91.30%
FRM*: Feature Reduction Method.
UMF*: Union Mathematical set theory.
IMF*: Intersection Mathematical set theory.
CS: ChiSquaredAttributeEval.
IG: InfoGainAttributeEval.

3.3 Evaluation of Performance Based on the Elimination of Unnecessary

Features

Here, we assess the comparative effectiveness of several distinct classifiers by making

use of a wide variety of feature selection strategies in order to establish whether or not
the effectiveness of these strategies could be enhanced further. We used primarily the
three filter-based feature reductions named as InfoGainAttributeEval + Ranker, GainRa-
tioAttributeEval + Ranker, and ChiSquaredAttributeEval + Ranker, as well as a novel
approach which is the intersection of these three methods, to improve the performance
of five different classifiers. We have used a well-known dataset known as the NSL-KDD
dataset for IDS in order to validate our suggested hybrid technique. The findings that we
have acquired from the NSL-KDD dataset are displayed in Table 6, which can be found
below.
 Machine Learning-Based Hybrid Feature Selection 343

Following the discussion in Sect. 3.3, we partitioned the NSL-KDD dataset into a
total of 32 different datasets using the ranking of feature reduction. During the first stage,
we used the ranker approach using IG, CS, and GR to determine the order in which the
features should be presented. Then, from the datasets that we had gotten after one hot
encoding, we eliminated the ten features with the lowest importance, and this procedure
is repeated until the lowest 80 attributes have been eliminated. As a consequence of this,
we now have 8 datasets, each of which has 41, 51, 61, 71, 81, 91, 101, 111 features from
one of three categories—namely, IG, CS, or GR—and one from our own hybrid method.
After that, each of the 32 datasets is put through training with five different classifiers,
resulting in a total of 160 different models. In a similar manner, we also prepare the test
datasets to contain the exact same number of characteristics as are found in the training
datasets. The next thing that we did was look at the true positive (TP), false negative
(FN), false positive (FP), and true negative (TN) values of 160 different models. We
calculate their sensitivity, specificity, precision, recall, f-measure, accuracy, Matthew’s
correlation coefficient, kappa statistic, root means squared error, mean squared error,
and Akaike Information Criterion based on the values of TP, FN, FP, and TN. The lower
the value of the AIC, the more accurate the model will be, which will help reduce the
complexity of the models, which will be very helpful for the IoT system.
Out of 160 models, the best model generated for each classifier (n = 5) chosen on
the basis of AIC [13] is depicted in Table 6. It is found that J48 performed well with an
accuracy of 99.855% having a model building time of 44.54 s using only 41 attributes
out of 121 attributes and using chi-squared feature reduction. REPTree achieved an
accuracy of 99.7746% having a model building time of 7.39 s with only 41 attributes out
of 121 attributes using chi-squared feature reduction. In spite of J48’s strong performance
in terms of detection accuracy, the amount of time required to construct its models is
significantly longer than that required by REPTree. REPTree performed well in terms of
specificity %, sensitivity %, and mean squared error as shown in Table 6. Among all, J48
performed well using only 41 features with chi-squared feature reduction. Our hybrid
approach which is the intersection of three feature reduction methods (IG, CS, and GR)
also performed well in BayesNet and REPTree classifiers. The J48 classifier significantly
performed well in terms of specificity %, sensitivity %, precision, recall, accuracy, mean
square error and the kappa statistic. Also, J48 outperformed by removing the lower 80
features out of 121 features using the CS attribute evaluation method. Because of this, we
strongly suggest that the J48 classifier be utilised for the purpose of intrusion detection
because it has an incredibly high levels of accuracy in detection while requiring just a
small number of features.
Figure 2 presents a comparison of the accuracy of the models we used in our analysis
when they were trained and when they were tested. This comparison reveals that both
J48 and REPTree achieved a high level of accuracy both during the training phase and
the testing phase.
344 P. Satapathy and P. K. Behera

Table 6. Classifiers trained models accuracy metrics using feature reduction

Classifier Features FDT* Se* Sp* Precision Recall FM* AC* MCC* K* RMSE* MSE* AIC*

J48 41 CS 99.91 99.80 0.998 0.999 0.999 99.855 0.997 0.997 0.038 0.001 −825152

BayesNet 81 Our 99.40 94.39 0.953 0.994 0.973 97.066 0.942 0.941 0.162 0.026 −457952
hybrid
approach

LogitBoost 41 IG 97.86 96.25 0.967 0.979 0.973 97.107 0.942 0.942 0.149 0.022 −479911

REPTree 41 CS 99.84 99.70 0.997 0.998 0.998 99.775 0.995 0.996 0.045 0.002 −77955

NB 81 Our 94.91 86.88 0.893 0.949 0.920 91.172 0.824 0.822 0.290 0.084 −312148
hybrid
approach

FDT: Feature Reduction Techniques: Se: Sensitivity; Sp: Specificity; FM: F-measure; AC: Accu-
racy; MCC: Matthew’s correlation coefficient; K: Kappa statistic; RMSE: Root means squared
error; MSE: Mean squared error; AIC: Akaike information criterion;

Fig. 2. Train and Test accuracy of the best model among 160 models

Figure 3 displays the accuracy of each of the five classifiers after deleting the 80
attributes with the lowest rank.
 Machine Learning-Based Hybrid Feature Selection 345

Fig. 3. Training accuracy of classifiers with 41 attributes using feature reduction

Figure 4 displays the accuracy of each of the five classifiers after removing the 10
attributes with the lowest rank. Thus, it is clear from our analysis that J48 and REPTRee
achieved good accuracy with only 41 attributes out of 121 attributes using the feature
reduction method. Meanwhile, REPTreee using our intersection approach of feature
reduction achieves a model building time of 7.7 s with 41 attributes and uses CS feature
reduction to achieve a model building time of 7.39 s. This research makes it very evident
that the method that we have suggested shortens the amount of time needed to develop
models while simultaneously reducing the number of attributes involved, all without

Fig. 4. Training accuracy of classifiers with 111 attributes using feature reduction
346 P. Satapathy and P. K. Behera

sacrificing accuracy. REPTree using CS feature reduction achieved a model building

time of 18.74 s with 111 attributes, which confirms that when we increase the number
of attributes the model building time also increases.

4 Conclusion
In this article, we describe a data analysis method for the purpose of detecting intrusions
into the Internet of Things system. The NSL-KDD training dataset was pre-processed,
then one-hot encoded, and then subsetted using three alternative feature reduction algo-
rithms. With the use of WEKA’s technologies, we were able to evaluate how well five
distinct machine learning classifiers performed on these datasets. The individually cre-
ated models were ranked based on the values of the Akaike Information Criterion (AIC),
which takes into consideration both MSE and the amount of features that were utilised to
generate the model. It was observed that the model that was created by employing 41 fea-
tures, and then ranked by CS algorithm while using J48 classifier, was the best model.
This model had the lowest AIC value of -825152, and it had an accuracy of 99.855.
We compared our findings with those of four existing approaches and discovered that
the J48 classifier performed well when compared to other classifiers while utilising the
chi-squared feature reduction method with an accuracy of 99.8555%. The J48 classifier
outperformed the challenge by achieving a detection accuracy of 99.7817% by utilising
our hybrid feature reduction with 22 attributes. As a result, the accuracy of the model has
been increased, and the rate of false positives that our proposed method generates has
also been found to be minimal. In the future, we plan to extend our ML-based framework
to numerous datasets with multiclass classification, and we also plan to create a deep
learning-based framework for intrusion detection. Both of these things will take place
in the future.

Acknowledgments. I extend my acknowledgment for the financial assistance from the UGC, New
Delhi, in the form of Junior Research Fellowship (JRF) and Department of Computer Science and
Applications, Utkal University for providing me an opportunity to pursue my PhD work.

References
1. Ahanger, A.S., Khan, S.M., Masoodi, F.: An effective intrusion detection system using super-
vised machine learning techniques. In: 2021 5th International Conference on Computing
Methodologies and Communication (ICCMC). IEEE (2021)
2. Bandyopadhyay, S., et al.: A Decision Tree Based Intrusion Detection System for Identifica-
tion of Malicious Web Attacks (2020)
3. Alabdulwahab, S., Moon, B.: Feature selection methods simultaneously improve the detection
accuracy and model building time of machine learning classifiers. Symmetry 12(9), 1424
(2020)
4. Albulayhi, K., et al.: IoT intrusion detection using machine learning with a novel high
performing feature selection method. Appl. Sci. 12(10), 5015 (2022)
5. Frank, E., et al.: Weka-a machine learning workbench for data mining. In: Data Mining and
Knowledge Discovery Handbook, pp. 1269–1277. Springer (2009). https://doi.org/10.1007/
978-0-387-09823-4_66
 Machine Learning-Based Hybrid Feature Selection 347

6. Mahfouz, A.M., Venugopal, D., Shiva, S.G.: Comparative analysis of ML classifiers for net-
work intrusion detection. In: Yang, X.-S., Sherratt, S., Dey, N., Joshi, A. (eds.) Fourth interna-
tional congress on information and communication technology. AISC, vol. 1027, pp. 193–207.
Springer, Singapore (2020). https://doi.org/10.1007/978-981-32-9343-4_16
7. Shaukat, S., et al.: Intrusion detection and attack classification leveraging machine learning
technique. In: 2020 14th International Conference on Innovations in Information Technology
(IIT). IEEE (2020)
8. Tavallaee, M., et al. A detailed analysis of the KDD CUP 99 data set. in 2009 IEEE symposium
on computational intelligence for security and defense applications. 2009. Ieee
9. Revathi, S., Malathi, A.: A detailed analysis on NSL-KDD dataset using various machine
learning techniques for intrusion detection. Int. J. Eng. Res. Technol. (IJERT) 2(12), 1848–
1853 (2013)
10. Choudhury, S., Bhowal, A.: Comparative analysis of machine learning algorithms along with
classifiers for network intrusion detection. In: 2015 International Conference on Smart Tech-
nologies and Management for Computing, Communication, Controls, Energy and Materials
(ICSTM). IEEE (2015)
11. Sarangi, A.N., Lohani, M., Aggarwal, R.: Prediction of essential proteins in prokaryotes
by incorporating various physico-chemical features into the general form of Chou’s pseudo
amino acid composition. Protein Pept. Lett. 20(7), 781–795 (2013)
12. Joshi, M., Hadi, T.H.: A review of network traffic analysis and prediction techniques. arXiv
preprint arXiv:1507.05722 (2015)
13. Karegowda, A.G., Jayaram, M., Manjunath, A.: Combining Akaike’s information criterion
(AIC) and the golden-section search technique to find optimal numbers of k-nearest neighbors.
Int. J. Comput. Appl. 2(1), 80–87 (2010)
14. Kumar, K., Batth, J.S.: Network intrusion detection with feature selection techniques using
machine-learning algorithms. Int. J. Comput. Appl. 150(12) (2016)
15. Kumar, G., Thakur, K., Ayyagari, M.R.: MLEsIDSs: machine learning-based ensembles for
intrusion detection systems—a review. J. Supercomput. 76(11), 8938–8971 (2020). https://
doi.org/10.1007/s11227-020-03196-z
16. Gudla, S.P., Bhoi, S.K., Nayak, S.R., Verma, A.: DI-ADS: a deep intelligent distributed denial
of service attack detection scheme for fog-based IoT applications. Math. Probl. Eng. 8, 2022
(2022)
17. Gudla, S.P.K., Bhoi, S.K.: MLP deep learning-based DDoS attack detection framework for
fog computing. In: Rout, R.R., Ghosh, S.K., Jana, P.K., Tripathy, A.K., Sahoo, J.P., Li, K.-C.
(eds.) Advances in Distributed Computing and Machine Learning: Proceedings of ICADCML
2022, pp. 25–34. Springer Nature Singapore, Singapore (2022). https://doi.org/10.1007/978-
981-19-1018-0_3
18. Thaseen, I.S., Kumar, C.A.: Intrusion detection model using fusion of chi-square feature
selection and multi class SVM. J. King Saud Univ. Comput. Inf. Sci. 29(4), 462–472 (2017)
 Application of Efficient Feature Selection
and Machine Learning Algorithms in Mental
Health Disorder Identification

Sumitra Mallick1,2(B) and Mrutyunjaya Panda1,2

1 Department of Computer Science, Guru Nanak Institution Technical Campus, Hyderabad,
Telangana, India
[email protected]
2 Department of Computer Science and Applications, Utkal University, Bhubaneswar,
Odisha, India
[email protected]

Abstract. Mental disorder is an illness that are more common in technical

employees and are increasing among working professionals. It is an important
and challenging issue in the world as the stress among the technical employees
are high due to work culture. Mental health in the healthcare is necessary to rec-
ognize individual’s situation and continue to predict diverse situations accurately.
Therefore, this study proposes the mental health disorder prediction using various
feature selection algorithms for a specific dataset called Tech survey Datasets. We
applied multiple machine learning classification algorithms on the best features
of RFE, RFECV and LASSO to obtain the performance metrics of the model and
to decide the best accuracy, precision and recall of the respective models. There
are 61 features in the Tech survey Dataset consisting of the data in the technical
workplace worldwide that provides mental health attribute and frequency. The
results are discussed and an aggregated table is developed using the performance
metrics to understand the percentage of technical employees undergoes mental
disorder. The proposed research evaluates various classification algorithms along
with feature selection algorithm to forecast the mental disorder in a particular
workplace and found to be 79% accuracy.

Keywords: Mental disorder · Feature selection · RFE · RFECV · LASSO ·

Associative property · Distributive property

1 Introduction
More than half of the population in different technical companies are likely suffering
from mental disorder at least once during their lifetime. Depression which affects 350
million people globally suggested by World Heath Organization (WHO) (Breiman1984).
Majority of people required treatment for depression may not be provided as per the
report of WHO. In the United States, one of the most common mental health condition
(Cuellar et al. 2005) is major depressive disorder (MDD). MDD, also referred to as

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 348–364, 2022.
https://doi.org/10.1007/978-3-031-23233-6_26
 Application of Efficient Feature Selection and Machine Learning Algorithms 349

clinical depression, is a significant medical condition that can affect many areas of your
life. More than 264 people of all age globally suffering from MDD according to world
health organization (WHO Mental-disorders 2019). The combination of genetic and non-
genetic characteristics in AD is the diversity of MDD. A recent study which combines
clinical predictors and gene study variants like rs6313 of 5-hydroxytryptamine (sero-
tonin) receptor 2A (HTR2A), rs7430 of protein phosphatase 3 catalytic subunit gamma
isozyme (PPP3CC), and rs6265 of brain-derived neurotrophic factor (BDNF) from the
European Group for the Study of Resistant Depression (GSRD). A Random forest model
reaching accuracy of above 79% (Kautzky et al. 2015) and in other investigation where
a series of linear and logistic elastic net regularized regression (ENRR) models is used
for training to achieve accuracy of 25% (Aitchison et al., 2005). The different protein
level platelet has been shown among different MDD patients (Huang et al., 2014). The
peripheral marker genes are considered for alzheimer’s disease whereas Visinin-like
1 (VSNL1) has found co-expression with calcium for AD and long term depression
(Patel et al. 2015) have worked to separate the late-life depression (LLD) patients from
the normal patients for which a decision tree derived from structural imaging, age and
mini-mental status examination scores and achieve an accuracy of 89%.
In feature selection process, the subset of features is selected from the larger dataset
of original features. The feature selection steps include data quality, less computational
time for prediction of model, predictive performance improvement and efficiency in data
collection process. (Muhammad Salman et al. 2022) used filter-based feature selection
technique named as ANOVA-f (Analysis of Variance test) to determine the most impor-
tant features. In multi linear regression analysis, depression risk is set as target variable
and for predicting the depression the context variables are set as independent variables
(Zhang 2006). The two-class problem which was proposed to solve the SVM-RFE for
analysis of multi class problems (Mundra et al. 2010, Xiaohui Lin et al. 2017). This multi
class problem for feature selection increase the robustness of SVM-RFE, by removing
the number of features with the lowest weights. Rather than removing the features one
by one which is time consuming, the author has (X.G. Zhan, X.H. Lin 2015) proposed
R-SVM & F-SVM to remove the noise in the data. They have followed a method AV-MI,
perform to remove the noisy variable, and achieve high specificity and sensitivity. Fea-
ture selection process used in this paper (John scott 2018) are 10-fold cross validation
which is used to run 10 times for validation. There are 100 features subsets that computed
from the research and the feature are ranked according to their descending order. The
top ranked feature is the most necessary one selected by using PCA plotted graph on the
liver disease data set and selected 34 ion features.
In few cases, few samples of three liver disease are mixed due to noise and non-
informative variable where MI-SVM-RFE is very constructive for obtaining the selective
features. It is showing three different liver disease are clearly separated by using two
intercept method- R2 intercept and Q2 intercept (Xiaohui Lina 2012). In the process of
selecting the feature to remove the noise method used is MI-SVM-RFE, which is most
popular method to select the significant features and to validate the proposed model. SVM
and CRT model developed to predict the response and remission to duloxetine utilizing
genome-wide genetic dataset but for selecting the features, the LASSO regression was
used. LASSO model is used to predict the response or remission with the accuracy.
350 S. Mallick and M. Panda

Finally, an optimal model was evaluated on the test subset of data with the nested cross-
validation (Patel et al. 2015). The growing body of literature which addresses the role
of social networks such as breakup relationship, mental illness (‘depression’, ‘anxiety’,
‘bipolar’ etc.) (Azorin 2013, Almeida 2010).
The paper is further organized as follows: Sect. 2 elaborates the background works
given by the authors related to mental health illness, Sect. 3 generate a heat map of the
dataset considered for the research, Sect. 4 proposes the mathematical modeling of the
feature selection techniques, Sect. 5 discusses the experimental setup and implementa-
tion, Sect. 6 shows the results and discussions, and Sect. 7 deals with the conclusion for
better analysis and classification with future models.

2 Background

In order to improve the performance of machine learning models a statistical analysis is

desired as the data is huge which are collected from clinicians and pharmacists. Gaussian
mixture model (Hastie 2001) is an unsupervised clustering algorithm that fits the data
based on structures that determine the dimension in which the subgroup is found. The aim
is to identify subgroups with inpatients rather than classify them. The depression is also
detected through user’s twitter by using the Bag-of-word technique i.e. the frequency of
words that shows the depression level (Okser 2014, Dudek 2013). The author has used
morphological stemming and categorization of Japanese tweet text, where bag-of-words
is used to predict depression among participants.
The author considers a total of 571,054 SNPs genotyped samples at the Centre for
Applied Genomics (TCAG) at the Hospital for Sick Children (Toronto ON, Canada).
While doing data preprocessing, participants of non-European self-reported ancestry (n
= 29) were excluded from the analysis (Sachan et al. 2021). The remaining sample
(N = 411) follows the necessary steps to improve quality control includes the follow-
ing: heterozygosity rate, relatedness (IBD), and possible population stratification due to
ancestry. DNA samples were collected from clinical data and blood samples which are
obtained from duloxetine and placebo arms. Gene variational characteristics are used in
each step of the training set. The robust LASSO approach has been used in genomics
studies to solve the ‘curse of dimensionality’ issue (Okser et al. 2014). Machine learning
needs data to implement the algorithm (Breiman et al. 1984) of classification trees and
linear support vector machines to construct a model. A predictive model is a general
concept for building the model in a supervised machine learning method that classi-
fies the observations. (Patel et al. 2016) on the MDD and CRT dataset constructed the
SVM model using tenfold cross-validation, during the training stage where the best
parameters are determined for the SVM model. To evaluate the model different metrics
Accuracy, specificity, and sensitivity were computed. The classification performance for
the training and testing of each pair of datasets implemented in R package caret. By the
LASSO model, MADRS score computed on the two types of models, the first model
is constructed based on the original genotype variant, and the second one is based on
genotype and imputed genetic variants. Xiaohui et al. took metabolomics which is the
study of small molecules composed of cells, tissues, or organisms of stimuli or genetic
modifications. To determine several human diseases in the human body and to improve
 Application of Efficient Feature Selection and Machine Learning Algorithms 351

the diagnosis and prevention of disease. This study was also carried out in the area
of toxicology and pharmacology, crop breeding plant biotechnology and drug research
(Aitchison 2005). The emerging field of metabolomics can improve and explore the
mechanism of cancer.
Data was collected from patients who are suffering from chronic hepatitis B (CHB),
cirrhosis (CIR), and hepatocellular carcinoma (HCC) of Sixth People’s Hospital (Aitchi-
son 2005). Each patient data was collected from MRI scan, ultrasonography, CT scan and
tumor markers test and the classification accuracy rate is computed firstly by comparing
SVM-RFE-OA with SVM-RFE. To determine the performance of the SVM-RFE-OA,
it shows that it performs outstanding for all the eight biological datasets (Xiaohui Lin
2017). The performance of the M-SVM-RFE-OA compared with SVM-RFE-OA was
computed by performing the classification and determining the high-value features.
Ultimately, it is observed that the M-SVM-RFE-OA performs better (John Scott).

3 Data Exploration
Data exploration is the study and correlation of huge amount of data in an easiest way for
visualization. Understanding the structure and exploring data to identify how features are
distributed and whether there is a link within the dataset. In the below Fig. 1, correlation
within the variable is observed between the dependent variable and independent variable
are depicted in this graph. it is also explained the columns in the dataset are correlated
and linked between the columns of the dataset.

Fig. 1. The correlation values for employee in Tech survey features datasets.

The main motivation of this research is to evaluate the state-of-the-art methodolo-

gies to determine the accuracy of mental disorder prediction for technical employees.
A state-of-art technique is used to assess the type of mental disorder problem a tech
employee is experiencing based on the awareness of mental disorder among employ-
ees. The aim of the research is to apply machine learning model using feature selection
method namely SVM, Naive Bayes, Stochastic Gradient Descent, KNN, Decision tree,
Random Forest, Logistic Regression, Neural Nets, Cross Gradient Booster, Cross Gradi-
ent Booster (Random Forest) for mental disorder and the model is tested using different
352 S. Mallick and M. Panda

performance measurements. It is necessary to determine which approaches are more

effective for identifying the mental disorder and which machine learning algorithm pro-
vide superior results. (Zhang et al. 2019) describe how to extract efficient features based
on the weighting-and-ranking-based hybrid feature selection (WRHFS), to identify the
heart stroke risk. The weighting and ranking techniques are filter based feature selection
(Ruba Skaik et al. 2020).

4 Mathematical Modeling of Feature Selection

We have implemented three feature selection algorithms i.e. LASSO, RFE and RFECV
to calculate the score of feature with the targeted variable. The target class feature for
this research is mh_disorder_current which finds the percentage of technical people who
gets mental illness compared to nontechnical people in the tech survey dataset. The best
features are extracted by using all the above three feature selection algorithms. We have
collected the mental health in technical survey dataset consisting of 1235 records and
61 features from various people worldwide that measure attitudes towards mental health
and frequency of mental health illness in the technical workplace. In this dataset, there
are a total of 61 features and mostly the NaN values are present as the data values are
removed. The goal is to find a subset of size f features among N universal features to
maximize the performance of the model using a machine learning classifier.
There is a requirement of computing the correlation coefficient between two variables
to find the relationship. It needs to generate a correlation matrix to summarize the dataset
to find the best features for further processing of data. This paper considers 51 features
out of 61 features of the dataset after implementing a heatmap that removed highly
correlated features. Correlation is a measure of the dependent variables in a way to
reflect two variables move together and strongly relate each other. It is said that highly
correlated variables become redundant for classification algorithms and degrade the
performance if considers.
Let us assume N the universal set consisting of all 51 features i.e. N = { f1,
f2,……….f51}.
The features set is { f1: “self_empl_flag”, f2: “comp_no_empl”, f3: “tech_comp_flag”,
f4: “mh_coverage_flag”, f5: “
mh_coverage_awareness_flag”, f6: “ mh_employer_discussion”, f7: “mh_resources_
provided”, f8:
“mh_anonimity_flag”, f9: “mh_medical_leave”, f10: “mh_discussion_neg_impact”,
f11: “ph_discussion_neg_impact”,
f12: “mh_discussion_cowork”, f13: “mh_discussion_supervis”, f14: “ mh_eq_ph_
employer”, f15:
“mh_conseq_coworkers”, f16: “prev_employers_flag”, f17: “prev_mh_benefits“, f18:
“prev_mh_benefits_awareness”,
f19: “prev_mh_discussion”, f20: “prev_mh_resources”, f21: “prev_mh_anonimity”,
f22:
“prev_mh_discuss_neg_conseq”, f23: “prev_ph_discuss_neg_conseq”, f24: “prev_mh_
discussion_cowork”, f25:
 Application of Efficient Feature Selection and Machine Learning Algorithms 353

“prev_mh_discussion_supervisor”, f26: “prev_mh_importance_employer”, f27: “prev_

mh_conseq_coworkers”, f28:”
future_ph_specification”, f29: “why/why_not”, f30: “future_mh_specification”, f31:
“why/why_not2 , f32:”
mh_hurt_on_career”, f33: “mh_neg_view_cowork”, f34:” mh_sharing_friends/fam_
flag”, f35:
“mh_bad_response_workplace”, f36: “mh_family_hist”, f37: “ mh_disorder_past”, f38:
“ mh_disorder_current”, f39: “mh_diagnos_proffesional”, f40:“mh_sought_proffes_
treatm”, f41:” mh_eff_treat_impact_on_work”, f42:” mh_not_eff_treat_impact_on_
work”, f43: “age”, f44: “sex”, f45: “country_live”, f46: “live_us_teritory”, f47: “coun-
try_work”, f48: “work_us_teritory”, f49: “work_position”, f50: “remote_flag”, f51: “
tech_flag”}.
Let F1 , F2 and F3 be the feature subsets of universal set N.
F1 is the feature subset that are extracted from the universal set N using LASSO
technique denoted as F1 = { f:F is a feature subset using LASSO}, f is a member of
feature set N, i.e. f ∈ N.
F2 is the feature subset that are extracted from the universal set N using RFECV
technique denoted as F2 = { f:F is a feature subset using RFECV}, f is a member of
feature set N, i.e. f ∈ N.
F3 is the feature subset that are extracted from the universal set N using RFE technique
denoted as F3 = {f:F is a feature subset using RFE}, f is a member of feature set N, i.e.
f ∈ N.

4.1 Feature Selection Using LASSO

LASSO (Least Absolute Shrinkage and Selection Operator) is a regression analysis that
can be used for both feature selection and regularization. LASSO Minimizes the error
between data points and the actual fit points by avoiding overfitting problem i.e., it
works well on the dataset where a model fit well on the training dataset but will not
fit well on the testing dataset. The main objective of LASSO regularization is to find
the best set of optimal features that helps to minimize the error between the actual and
the target feature. The unwanted features are removed by making the coefficients to
zero in the shrinking process. By using wrapper-based feature selection techniques like
LASSO Regularization, F1 features are extracted from the universal set N. The most
relevant features extracted based on the rank values in the employee disorder dataset are
19 features shown below. This also helps to deal with the overfitting and underfitting
problems of machine learning. Hence the feature set is derived as follows:

F1 = {f48, f29, f46, f21, f47, f18, f17, f9, f4, f8, f22, f36, f8, f34, f39, f37, f42, f49, f43, f31}

Each feature f ∈ F1 ⊆ N refer to a single feature, where {f48: work_us_teritory, f28:

why/why_not, f46: live_us_teritory,f21: prev_mh_anonimity, f47: country_work, f17:
prev_mh_benefits_awareness, f17: prev_mh_benefits, f9: mh_medical_leave, f4: mh_
coverage_flag, f22: prev_mh_discuss_neg_conseq, f36: mh_family_hist, f8: prev_mh_
discuss_neg_conseq, f34: mh_sharing_friends/fam_flag, f39: mh_diagnos_proffesional,
f37: mh_disorder_past, f42: mh_not_eff_treat_impact_on_work, f49: work_position,
f29: age, f31: why/why_not2}.
354 S. Mallick and M. Panda

4.2 Feature Selection Using RFECV

Recursive Feature Elimination with Cross-validation (RFECV) is a technique to evaluate
the performance of all the machine learning models. This is a powerful method that helps
to select the best model with its efficiency score. The paper considers the Tech survey
dataset consists of 51 preprocessed features by implementing k-fold cross validation
technique with the value of k equal to 5 where K data is the training dataset, and the
remaining K-1 data is the test dataset. The performance of the model is computed. This
process is continued multiple times with different values of k then the model performance
is evaluated. All the process is done multiple times to evaluate the robustness of the
model. By using RFECV feature selection technique, F2 features are extracted from the
universal set N. The most relevant features extracted based on the rank values in the
employee disorder dataset are 14 features shown below. Hence the feature set is derived
as follows:

F2 = {f43, f2, f46, f39, f37, f41, f9, f42, f34, f40, f29, f31, f49, f48}

Each feature f ∈ F2 ⊆ N refer to a single feature, where {f43: “age”, f2: “comp_no_
empl”, f46: “live_us_teritory”, f39:” ‘mh_diagnos_proffesional’,f37:” ‘mh_disorder_
past’,f41:” ‘mh_eff_treat_impact_on_work’,f9:” ‘ ‘mh_medical_leave’,f42:” ‘mh_not_
eff_treat_impact_on_work’,f34:” ‘mh_sharing_friends/fam_flag’,f40: ‘mh_sought_
proffes_treatm’,f29: ‘why/why_not’,f31: ‘why/why_not2’,f49: ‘work_position’, f48:
“work_us_teritory”}.

4.3 Feature Selection Using RFE

RFE is a wrapper method that uses random forest machine learning algorithm for its
evaluation criteria. It is a backward compactable way of doing feature selection or elim-
ination which works well in both the classification and the regression techniques. RFE
builds a model using all the features and then calculates the model’s accuracy. it removes
one feature per iteration and then builds the model again and sees how much variation is
caused in the accuracy score. It is based on the feature importance to keep the feature or
remove feature. At each iteration, feature importance is measured and the less relevant
one is removed to speed up the process. It is to construct a model by selecting best or
worst features in each iteration and repeating the process until all features are exhausted.
Features are eliminated based on their rank like Greedy optimization to find best subset
of features. It is the idea of repeatedly constructing a model and choosing either the best
or worst performing feature, setting the feature aside, and then repeating the process
with the rest of the features. This process is applied until all features in the dataset are
exhausted. Features are then ranked according to when they were eliminated. As such,
it is a greedy optimization for finding the best-performing subset of features. By using
wrapper-based feature selection techniques like Recursive Feature Elimination (RFE),
F3 features are extracted from the universal set N. The most relevant features extracted
from the employee disorder dataset are 10 features shown below. Hence the feature set
is derived as follows:

F3 = {f43, f39, f37, f41, f42, f34, f29, f31, f49, f46}
 Application of Efficient Feature Selection and Machine Learning Algorithms 355

Each feature f ∈ F3 ⊆ N refer to a single feature, where {f43:age, f39: mh_

diagnos_proffesional, f37: mh_disorder_past, f41: mh_eff_treat_impact_on_work, f42:
mh_not_eff_treat_impact_on_work, f34: mh_sharing_friends/fam_flag, f29: why/why_
not2, f31: why/why_not, f49: work_position, f46:work_us_teritory}.

4.4 Feature Selection Using Associative Property of Union

Let F1 , F2 and F3 be subsets of universal set N.
The union of F1, F2 , and F3 written as F1 U F2 U F3 = { f ∈ N | f ∈ F1 , f ∈ F2 and f
∈ F3 }.
Union operations are associative where (F1 U F2 ) U F3 = F1 U (F2 U F3 ).
By using union operation of associative property on F1 , F2 and F3 , the F4 features
are extracted. The most relevant features extracted from the employee disorder dataset
is 15 features shown below. Hence the feature set is derived as follows:

F4 = {f9, f43, f22, f21, f17, f41, f39, f36, f2, f49, f31, f42, f8, f38, f4, f18, f46, f47, f48, f40, f29, f34}

It is represented as F4= {f9:’mh_medical_leave’,f43:’age’, f22:’prev_mh_discuss_

neg_conseq’, f21:’prev_mh_anonimity’, f17:’prev_mh_benefits’, f41:’mh_eff_treat_
impact_on_work’, ‘,f39:’mh_diagnos_proffesional’, f36:’mh_family_hist’,f2: ‘comp_
no_empl, f49:’work_position’, f31:’why/why_not2’, f42:’mh_not_eff_treat_impact_
on_work’, f8:’mh_anonimity_flag’, f38:’mh_disorder_past’, f4:’mh_coverage_flag’,
f18:’prev_mh_benefits_
awareness’,f46: ‘live_us_teritory’,f47:’country_work’,f48: ‘live_us_teritory’, f40:’mh_
sought_proffes_treatm’,f29: ‘why/why_not’,f34: ‘mh_sharing_friends/fam_flag’}.

4.5 Feature Selection Using Associative Property of Intersection

Let F1 , F2 and F3 be subsets of universal set N.

The intersection of F1 , F2 and F3 written as F1 ∩F2 ∩F3 , is defined by the associative
property as set of all element that are in F1 , F2 and F3 .

F1 ∩ F2 ∩ F3 = {f ∈ N | f ∈ F1 , f ∈ F2 and f ∈ F3 }

(F1 ∩ F2 ) ∩ F3 = F1 ∩ (F2 ∩ F3 ) Hence, intersection operations are associative.

By using Intersection operation of associative property on F1 , F2 and F3 , the F5
features are extracted. The most relevant features extracted from the employee disorder
dataset is 8 features shown below. Hence the feature set is derived as follows:

F5 = {f49, f43, f37, f29, f31, f41, f34, f39}

F5 = {f49:’work_position’, f43:’age’, f37:’mh_disorder_past’,f29: ‘why/why_

not’, f31:’why/why_not2’, f41:’mh_not_eff_treat_impact_on_work’, f34:’mh_sharing_
friends/fam_flag’, f39: ‘mh_diagnos_proffesional’}.
356 S. Mallick and M. Panda

4.6 Feature Selection Using Distributive Property

Let F1 , F2 and F3 be subsets of universal set N.

The Union of F1 , F2 and intersection with F3 written as F1 ∪ (F2 ∩ F3 ), is defined
by the distributive property as set of all elements that are in F1, F2 and F3 are The
Distributive law shows as:

F1 ∪ (F2 ∩ F3 ) = (F1 ∪ F2 ) ∩ (F1 ∪ F3 )

By using union distributed over intersection of distributive property on F1 , F2 and F3 ,

the F6 features are extract ed. The most relevant features extracted from the employee
disorder dataset is 10 features shown below. Hence the feature set is derived as follows:

F6 = {f49, f48, f43, f37, f29, f31, f42, f34, f41, f39}

F6 = {f49:’work_position’, f48:’work_us_teritory’, f43:’age’, f37:’mh_disorder_

past’, f29:’why/why_not’, f31:’why/why_not2’, f42:’mh_not_eff_treat_
impact_on_work’, f34:’mh_sharing_friends/fam_flag’, f41:’mh_eff_treat_impact_on_
work’, f39:’mh_diagnos_proffesional’}.

5 Implementation of Classification Algorithms

The implementation is carried out with preprocessed 51 features and 10 classification

algorithms such as Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees,
Random Forest, Support Vector Machine, Logistic Regression, Neural Nets, Cross Gra-
dient Booster and Cross Gradient Booster (Random Forest) to find the accuracy, recall
and precision. The accuracy of Random Forest and Cross Gradient Booster (Random
Forest) is around 78%, the recall of Random Forest and Cross Gradient Booster (Random
Forest) is around 73% and the precision of Random Forest and Cross Gradient Booster
(Random Forest) is more than 75%. Considering all the 51 features, the accuracy, preci-
sion and recall of Random Forest and Cross Gradient Booster (Random Forest) is better
than all other machine learning algorithms shown in Table 1.

5.1 Classification with LASSO

We have implemented with 19 best features of LASSO and 10 classification algorithms

such as Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random Forest,
Support Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and
Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision. The
accuracy of Random Forest and Cross Gradient Booster (Random Forest) is more than
78%, the recall of Random Forest and Naïve Bayes is more than 73% but equivalent to
Cross Gradient Booster (Random Forest) and the precision of Random Forest and Cross
Gradient Booster (Random Forest) is 77%. Considering best 19 features of Tech survey
dataset, the accuracy, precision and recall of Random Forest and Cross Gradient Booster
(Random Forest) is better than all other machine learning algorithms shown in Table 1.
 Application of Efficient Feature Selection and Machine Learning Algorithms 357

Table 1. Considering 51 best features

Algorithms Accuracy Recall_score Precision_score

Naive Bayes 0.74394 0.7074 0.70435
Stochastic Gradient Descent 0.45553 0.45735 0.45424
KNN 0.39353 0.34693 0.34917
Decision trees 0.65229 0.637 0.62987
Random Forest 0.78167 0.72876 0.75662
Support Vector Machine 0.42588 0.33333 0.14196
Logistic Regression 0.69811 0.61629 0.66023
Neural Nets 0.41779 0.327 0.14078
Cross Gradient Booster 0.74394 0.68906 0.69643
Cross Gradient Booster (Random Forest) 0.78706 0.73563 0.77349

5.2 Classification with RFECV

We have implemented with 14 best features of RFECV and 10 classification algorithms
such as Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random For-
est, Support Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster
and Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision.
The accuracy of Random Forest and Cross Gradient Booster (Random Forest) is more
than 78%, the recall of Random Forest and Cross Gradient Booster (Random Forest) is
more than 77% and the precision of Random Forest and Cross Gradient Booster (Ran-
dom Forest) is 77%. Considering best 14 features of Tech survey dataset, the accuracy,
precision and recall of Random Forest and Cross Gradient Booster (Random Forest) is
better than all other machine learning algorithms shown in Table 1.

5.3 Classification with RFE

We have implemented with 10 best features of RFE and 10 classification algorithms such
as Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random Forest,
Support Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and
Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision. The
accuracy of Random Forest and Cross Gradient Booster (Random Forest) is more than
78%, the recall of Naive Bayes and Random Forest is more than 74% and the precision
of Random Forest and Cross Gradient Booster (Random Forest) is more than 77%.
Considering best 10 features of tech survey dataset, the accuracy, recall and precision
of Random Forest and Cross Gradient Booster (Random Forest) is better than all other
machine learning algorithms shown in Table 1.
358 S. Mallick and M. Panda

5.4 Classification with Feature Selection Using Associative Property of Union

We have implemented with 15 best features and 10 classification algorithms such as

Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random Forest, Sup-
port Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and
Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision. The
accuracy of Cross Gradient Booster (Random Forest) is more than 78%, the recall of
Cross Gradient Booster (Random Forest) is more than 73% and the precision of Cross
Gradient Booster (Random Forest) is 77%. Considering best 15 features of Tech survey
dataset, the accuracy, precision and recall of Cross Gradient Booster (Random Forest)
is better than all other machine learning algorithms shown in Table 1.

5.5 Classification with Feature Selection Using Associative Property

of Intersection
We have implemented with 15 best features and 10 classification algorithms such as
Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random Forest, Sup-
port Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and
Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision. The
accuracy of Cross Gradient Booster (Random Forest) is 79%, the recall of Cross Gradient
Booster (Random Forest) is 74% and the precision of Cross Gradient Booster (Random
Forest) is 78%. Considering best 8 features of Tech survey dataset, the accuracy, pre-
cision and recall of Cross Gradient Booster (Random Forest) is better than all other
machine learning algorithms shown in Table 1.

5.6 Classification with Feature Selection Using Distributive Property

We have implemented with 10 best features and 10 classification algorithms such as

Naive Bayes, Stochastic Gradient Descent, KNN, Decision trees, Random Forest, Sup-
port Vector Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and
Cross Gradient Booster (Random Forest) to find the accuracy, recall and precision. The
accuracy of Random Forest is 79%, the recall of Naïve Bayes is 74% and the precision of
Cross Gradient Booster (Random Forest) is 77%. Considering best 10 features of Tech
survey dataset, the accuracy, precision and recall of Cross Gradient Booster (Random
Forest) and Random Forest is better than all other machine learning algorithms shown
in Table 1.

6 Results and Discussions

The best subset of features are selected and the unimportant features are eliminated that
are not contributing to the accuracy of the model from the dataset. We have implemented
three algorithms for feature selection i.e. RFE, RFECV and LASSO to calculate the
feature score with respect to targeted variable. The target class feature for this research is
mh_disorder_current which leads to find the percentage of technical people gets mental
illness compared to non technical people in the tech survey dataset. By using RFE,
 Application of Efficient Feature Selection and Machine Learning Algorithms 359

RFECV and LASSO algorithms, the best features are extracted and shown in the Table 1
given above. The results are computed from the best features of RFE, RFECV and
LASSO algorithms by using ten machine learning algorithms called as Naive Bayes,
Stochastic Gradient Descent, KNN, Decision trees, Random Forest, Support Vector
Machine, Logistic Regression, Neural Nets, Cross Gradient Booster and Cross Gradient
Booster (Random Forest) to find the accuracy, recall and precision. The Table 2 is
prepared as an aggregated accuracy, recall and precision to visualize the importance of
the classification algorithms with respect to the feature selection methods.

Table 2. Aggregated Accuracy, Recall and Precision considering the best features

Feature selection Algorithms Classification Accuracy Recall_ Precision_

Algorithms score score
LASSO Naive Bayes 0.77089 0.73668 0.7392
Stochastic Gradient 0.32345 0.36213 0.55758
Descent
K-Nearest Neighbor 0.39623 0.34926 0.35247
Decission trees 0.66846 0.64281 0.63796
Random Forest 0.78437 0.73883 0.76686
Support Vector 0.42588 0.33333 0.14196
Machine
Logistic Regression 0.66577 0.59539 0.64776
Neural Nets 0.42588 0.33333 0.14196
Cross Gradient 0.75202 0.71616 0.7178
Booster
Cross Gradient 0.78706 0.73563 0.77349
Booster (Random
Forest)
RFECV Naive Bayes 0.77089 0.7392 0.7392
Stochastic Gradient 0.32345 0.55758 0.55758
Descent
K-Nearest Neighbor 0.39623 0.35247 0.35247
Decission trees 0.66846 0.63796 0.63534
Random Forest 0.78706 0.76686 0.77124
Support Vector 0.42588 0.14196 0.14196
Machine
Logistic Regression 0.66038 0.64776 0.64016
Neural Nets 0.42588 0.14196 0.2712
(continued)
360 S. Mallick and M. Panda

Table 2. (continued)

Feature selection Algorithms Classification Accuracy Recall_ Precision_

Algorithms score score
Cross Gradient 0.75202 0.7178 0.7178
Booster
Cross Gradient 0.78437 0.77349 0.7702
Booster (Random
Forest)
RFE Naive Bayes 0.77089 0.73668 0.7392
Stochastic Gradient 0.32345 0.36213 0.55758
Descent
K-Nearest Neighbor 0.39623 0.34926 0.35247
Decission trees 0.68194 0.658 0.65251
Random Forest 0.78437 0.73883 0.76686
Support Vector 0.42588 0.33333 0.14196
Machine
Logistic Regression 0.66577 0.59539 0.64776
Neural Nets 0.42588 0.33333 0.14196
Cross Gradient 0.75202 0.71616 0.7178
Booster
Cross Gradient 0.78706 0.73563 0.77349
Booster (Random
Forest)
LASSO ∪ RFECV ∪ RFE Naive Bayes 0.7655 0.72738 0.7296
Stochastic Gradient 0.19677 0.3348 0.5956
Descent
K-Nearest Neighbor 0.39084 0.34482 0.34712
Decission trees 0.67116 0.64867 0.64162
Random Forest 0.77628 0.72432 0.74748
Support Vector 0.42588 0.33333 0.14196
Machine
Logistic Regression 0.69272 0.59903 0.6449
Neural Nets 0.36927 0.30673 0.24677
Cross Gradient 0.74933 0.70566 0.71144
Booster
Cross Gradient 0.78706 0.73563 0.77349
Booster (Random
Forest)
(continued)
 Application of Efficient Feature Selection and Machine Learning Algorithms 361

Table 2. (continued)

Feature selection Algorithms Classification Accuracy Recall_ Precision_

Algorithms score score
LASSO ∩ RFECV ∩ RFE Naive Bayes 0.77089 0.73933 0.74166
Stochastic Gradient 0.38005 0.38543 0.56776
Descent
K-Nearest Neighbor 0.40431 0.35316 0.35371
Decision trees 0.66307 0.63969 0.63403
Random Forest 0.78706 0.7416 0.75857
Support Vector 0.42588 0.33333 0.14196
Machine
Logistic Regression 0.68464 0.6053 0.65834
Neural Nets 0.62534 0.51401 0.53059
Cross Gradient 0.75472 0.72314 0.7232
Booster
Cross Gradient 0.78976 0.7404 0.77673
Booster (Random
Forest)
LASSO ∪ (RFECV ∩ RFE) Naive Bayes 0.77089 0.73955 0.74172
Stochastic Gradient 0.45822 0.3969 0.40541
Descent
K-Nearest Neighbor 0.39084 0.34482 0.34626
Decision tree 0.69811 0.67773 0.66852
Random Forest 0.78706 0.73873 0.75867
Support Vector 0.42588 0.33333 0.14196
Machine
Logistic Regression 0.68733 0.59967 0.64024
Neural Nets 0.42588 0.33333 0.14196
Cross Gradient 0.75741 0.72259 0.7251
Booster
Cross Gradient 0.89706 0.73563 0.77349
Booster (Random
Forest)

The accuracy of LASSO feature selection with Cross Gradient Booster (Random
Forest) is around 79% which is best among all other classification algorithms but Naive
Bayes and Random Forest also performs well in this scenario. The accuracy of RFE
feature selection with Cross Gradient Booster (Random Forest) is around 79% which is
best among all other classification algorithms but Naive Bayes and Random Forest also
performs well in this scenario. The accuracy of RFECV feature selection with Random
362 S. Mallick and M. Panda

Forest is around 79% which is best among all other classification algorithms but Naive
Bayes and Cross Gradient Booster (Random Forest) also performs well in this scenario.
The classification accuracy of Cross Gradient Booster (Random Forest) is more than 78%
with feature selection using associative property of union, The classification accuracy of
Cross Gradient Booster (Random Forest) is 79% with feature selection using associative
property of intersection, and the classification accuracy of Random Forest is 79% with
feature selection using distributive property. We have concluded that Cross Gradient
Booster (Random Forest) performs very well among all the classification algorithms
among all the feature selection algorithms. The prediction for the mental health illness
in tech survey dataset for technical people compared to non technical people is 79%.
The recall of LASSO feature selection with is around 74% which is best among all
other classification algorithms but Naive Bayes and Cross Gradient Booster (Random
Forest) also performs well with nearer value. Similarly the recall of RFE feature selection
with Random Forest is around 74% which is best among all other classification algo-
rithms but Naive Bayes and Cross Gradient Booster (Random Forest) also performs well
with nearer value. The recall of RFECV feature selection with Cross Gradient Booster
(Random Forest) is around 77% which is best among all other classification algorithms,
but Random Forest also performs well. The recall of Cross Gradient Booster (Random
Forest) is more than 73% with feature selection using associative property of union,
the recall of Cross Gradient Booster (Random Forest) is 74% with feature selection
using associative property of intersection, and the recall of Naïve Bayes is 74% with
feature selection using distributive property. We have concluded that Cross Gradient
Booster (Random Forest) performs very well among all the classification algorithms
with RFECV feature selection algorithm is 77%. The recall for the mental health illness
in tech survey dataset for technical people compared to non technical people is 77%. It
means, the model correctly identifies 77% technical people have mental illness disease
in the organization.
The precision of LASSO feature selection with Cross Gradient Booster (Random
Forest) is around 77% which is best among all other classification algorithms but Naive
Bayes also performs well with nearer value. Similarly the precision of RFE feature selec-
tion with Cross Gradient Booster (Random Forest) is around 77% which is best among
all other classification algorithms but Random Forest also performs well. The precision
of RFECV feature selection with Random Forest is around 77% which is best among
all other classification algorithms but Cross Gradient Booster (Random Forest) also per-
forms well in this scenario. The precision of Cross Gradient Booster (Random Forest)
is 77% with feature selection using associative property of union, the precision of Cross
Gradient Booster (Random Forest) is 78% with feature selection using associative prop-
erty of intersection, and the precision of Cross Gradient Booster (Random Forest) is 77%
with feature selection using distributive property. We have concluded that Cross Gradi-
ent Booster (Random Forest) performs very well among all the classification algorithms
with feature selection using associative property of intersection is 78% but RFECV with
Random Forest, Cross Gradient Booster (Random Forest) with feature selection using
associative property of union and Cross Gradient Booster (Random Forest) with feature
selection using distributive property also performed well with nearer value. The preci-
sion for the mental health illness in tech survey dataset for technical people compared
 Application of Efficient Feature Selection and Machine Learning Algorithms 363

to non technical people is 78%. It means, the classification model predicts the technical
people have mental illness disease is correct 78% of the time.

7 Conclusion

We proposed a healthcare system for mental health disorders that can perform better with
better feature selection algorithms. Though mental illness is a chronic illness, the results
are compared by implementing various machine learning algorithms with respective
feature selection algorithms for accurate predictions and justifying the suitability of the
algorithms. The research was carried out with 61 features, 3 feature selection algorithms,
and by implementing 10 classification algorithms. The accuracy of the classification
algorithms is evaluated and the better accuracy with Cross Gradient Booster (Random
Forest) with all feature selection algorithm is found to be 79%. Therefore, considering
a workplace, the inference drawn is 79% of technical people are predicted that they
suffer from mental health illness disease according to the tech survey dataset. We have
calculated the recall with all ten classification algorithms with the best features using
feature selection algorithms and the model correctly identifies 77% of technical people
who have mental illness disease in the organization. We have calculated the precision with
all ten classification algorithms with the best features using feature selection algorithms
and the model predicts that the technical people who have mental illness disease is
identified as correct 77% of the time. In the future, we will implement a model that
considers various operations on feature selection algorithms by classifying the mental
health disease prediction.

References
Skaik, R., Inkpen, D.: Using social media for mental health surveillance: a review Canada. ACM
Comput. Surv. 53(6), 1–31 (2021). https://doi.org/10.1145/3422824. University of Ottawa
(2020)
Cuellar, A.K., Johnson, S.L., Winters, R.: Distinctions between bipolar and unipolar depression.
Clinical Psychol. Rev. 25(3), 307–339 (2005)
Kautzky, A., et al.: The combined effect of genetic polymorphisms and clinical parameters on
treatment outcome in treatment-resistant depression. Eur. Neuropsychopharmacol. 25(4), 441–
453 (2015)
Aitchison, K.J., Basu, A., McGuffin, P., Craig, I.: Psychiatry and the ‘new genetics’: hunting for
genes for behaviour and drug response. Br. J. Psychiatr. 186, 91–92 (2005)
Huang, T.L., Sung, M.L., Chen, T.Y.: 2D-DIGE proteome analysis on the platelet proteins of
patients with major depression. Proteome Sci. 12(1), 1 (2014)
Patel, M.J., Andreescu, C., Price, J.C., Edelman, K.L., Reynolds, C.F., Aizenstein, H.J.: Machine
learning approaches for integrating clinical and imaging features in late-life depression
classification and response prediction. Int. J. Geriatr. Psychiatr. 30(10), 1056–1067 (2015)
Pathan, M.S., Nag, A., Pathan, M.M., Deva, S.: Analyzing the impact of feature selection on the
accuracy of heart disease prediction (2022). arXiv:2206.03239v1
Sachan, S., Almaghrabi, F., Yang, J.-B., Xu, D.-L.: Evidential reasoning for preprocessing uncer-
tain categorical data for trustworthy decisions: an application on healthcare and finance. Expert
Syst. Appl. 185(2021), 115597 (2021)
364 S. Mallick and M. Panda

Guyon, I., Weston, J., Barnhill, S., et al.: Gene selection for cancer classification using support
vector machines. Mach. Learn. 46(1–3), 389–422 (2022)
Almeida, J.R., Versace, A., Hassel, S., et al.: Elevated amygdala activity to sad facial expressions:
a state marker of bipolar but not unipolar depression. Biol. Psychiatr. 67(5), 414–421 (2010)
Mundra, P.A., Rajapakse, J.C.: SVM-RFE with MRMR filter for gene selection. IEEE Trans.
Nanobiosci. 9, 31–37 (2010). CrossRef PubMed
Lin, X., Li, C., Zhang, Y., Su, B., Fan, M., Wei, H.: School of Computer Science and Technology,
Dalian University of Technology, Dalian. Selecting Feature Subsets Based on SVM-RFE and
the Overlapping Ratio with Applications in Bioinformatics (2017). https://doi.org/10.3390/mol
ecules23010052
https://methods.sagepub.com/book/social-network-analysis-4e/i829.xml
Polat, Ö.: A robust regression based classifier with determination of optimal feature set. J. Appl.
Res. Technol. 13(4), 443–446 (2015)
Scott, J.: https://methods.sagepub.com/book/social-network-analysis-4e/i829.xml. https://doi.
org/10.4135/9781529716597.n9
Islam, M.R., Kabir, M.A., Ahmed, A., Kamal, A.R.M., Wang, H., Ulhaq, A.: Depression detection
from social network data using machine learning techniques. Health Inf. Sci. Syst. 6(1), 1–12
(2018). https://doi.org/10.1007/s13755-018-0046-0
Azorin, J.M., et al.: Characteristics and profiles of bipolar I patients according to age-at-onset:
findings from an admixture analysis. J. Affect. Disord. 150, 993–1000 (2013)
Yu, Y., et al.: How to conduct dose-response meta-analysis by using linear relation and piecewise
linear regression model. J. Evid. Based Med. 16(1), 111–114 (2016)
Hastie, T., Tibshirani, R., Friedman, J.: The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. SpringerVerlag, New York (2001)
Okser, S., Pahikkala, T., Airola, A., Salakoski, T., Ripatti, S., Aittokallio, T.: Regularized machine
learning in the genetic prediction of complex traits. PLoS Genet. 10(11), e1004754 (2014)
Breiman, L., Friedman, J.H., Olshen, A., Stone, C.J.: Classification and Regression Trees.
Wadsworth Publishing Company, Belmont, California, USA (1984)
Lin, X., et al.: A support vector machine-recursive feature elimination feature selection method
based on artificial contrast variables and mutual information. J. Chromatography B 910, 149–
155 (2012). https://doi.org/10.1016/j.jchromb.2012.05.020
Dudek, D., Siwek, M., Zielinska, D., et al.: Diagnostic conversions from major depressive disorder
into bipolar disorder in an outpatient setting: results of a retrospective chart review. J. Affective
Disorders 144(1–2), 112–115 (2013)
World Health Organization (WHO). Mental Disorders (2019). WHO. https://www.who.int/news-
room/fact-sheets/detail/mental-disorders
Pathan, M.S., Nag, A., Pathan, M.M., Deva, S.: Analyzing the impact of feature selection on the
accuracy of heart disease prediction (2022). arXiv:2206.03239v1
Zhang, X.G., et al.: Recursive SVM feature selection and sample classification for mass-
spectrometry and microarray data. BMC Bioinformatics 7, 197 (2006). https://doi.org/10.1186/
1471-2105-7-197
Zhang, Y., Zhou, Y., Zhang, D., Song, W.: A stroke risk detection: improving hybrid feature
selection method. J. Med. Internet Res. 21(4), e12437 (2019). https://doi.org/10.2196/12437
 A Survey on Sentimental Analysis of Student
Reviews Using Natural Language Processing
(NLP) and Text Mining

J. Jayasudha1(B) and M. Thilagu2

1 Research Scholar, Department of Computer Science, Avinashilingam Institute for Home
Science and Higher Education for Women, Coimbatore, India
[email protected]
2 Department of Computer Science, Avinashilingam Institute for Home Science and Higher
Education for Women, Coimbatore, India
[email protected]

Abstract. Sentiment Analysis (SA) is a Natural Language Processing (NLP)

application. It’s sometimes referred to as “opinion mining” or “emotion extrac-
tion”. Sentiment Analysis is a growing area of study in text mining and a popular
research subject for opinion mining in education, which analyses and compre-
hends students’ attitudes toward their learning in order to improve the quality of
decision-making. More innovative approaches have been developed in online edu-
cation, which provide an efficient way to learn irrespective of Gender, geographical
discrimination, age, etc. As more and more reviews or feedback are posted during
online classes, these messages should be correctly classified to identify the theme
behind them. Machine Learning (ML) and Deep Learning (DL) based classifica-
tion models are used to solve sentiment classification problems. Machine Learning
(ML) algorithms Support Vector Machine (SVM), Naïve Bayes (NB), Random
Forest (RF) and Decision Tree (DT) and Deep Learning (DL) algorithms especially
Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN)
are most frequently used algorithms for Natural Language processing (NLP). It
allows us to have a deeper understanding of the student’s emotions. It may also
be used to recognize and analyse feelings in user comments or reviews on var-
ious blogs, websites, social media, online communities, and other sites in order
to better understand their perspectives. Thus, the most addressed domain was the
use of sentiment analysis in the context of higher education and the evaluation
of teaching quality. It would be very helpful for instructors to improve learning
strategies by assessing both the instructors and students’ progress.

Keywords: Sentimental analysis · Opinion mining · Emotion extraction ·

Polarity · Natural Language Processing (NLP) · Machine learning · Deep
learning · Text mining

1 Introduction
The current educational system offers a large E-Learning platform that enables students
to learn new information and abilities. It also offers a means of enhancing the teaching

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 365–378, 2022.
https://doi.org/10.1007/978-3-031-23233-6_27
366 J. Jayasudha and M. Thilagu

and learning process through the analysis of feedback or Comments in order to take
better decisions. Sentiment analysis has been widely used for a variety of objectives in
numerous application fields. When taking into account customer ratings, social media
monitoring, and other factors, the education sector should be given priority. It is very
important to make a proper analysis of textual data as the data collected is unstructured.
Sentimental Analysis is used for analyzing the positive, negative and neutral sentiments
of user’s opinion. This uses the concept of Natural Language Processing and text analysis
to analyze user’s sentiment using tweets, comments, reviews posted. By analyzing the
context it identifies the sentiment of users and it helps them to makes some changes for
better approach and make users to meet their needs. Sentiment analysis can be carried
out at the word, sentence, or document level. Because handling these reviews manually
is difficult due to the large amount of data, automatic processing is required.
The popularity and importance of students feedback have also increased recently
especially during COVID-19 when compared to movie reviews or tweets (Santos,
Cicero, & Gatti, 2014.), With a few notable exceptions, who examined the connection
between a ratio based on the positive and negative phrases in the postings and dropout
(Wen, 2014.). There are very few contributions in the domain of MOOCs and educa-
tion evaluations. Additionally, neither a defined strategy for handling sentiment analysis
in MOOCs nor a comparison of various methodologies exist nor role of teachers has
surprisingly shifted over the years. Teachers must now not only get familiar with new
tools, but also keep up with cutting-edge technology. The skeleton of any educational
system is its teachers. A teacher’s effectiveness is judged not just by his or her academic
qualifications, but also by skill, and commitment. Feedback at regular intervals is con-
sidered as a most effective approach for a teacher to improve teaching methodologies.
Feedback might be an open-ended or closed-ended as they are loaded with observations
and insights, open ended textual feedbacks are difficult to examine manually and draw
conclusions.More number of research are being carried out in identifying sentimental
Analysis and Opinion mining. To be specific opinion mining helps teachers to evaluate
the attitudes and behavior of students towards specific subject, platforms and teachers.
Machine Learning (Baidal, Karen, Carlota, & Vera, 2018) and Deep Learning algo-
rithms are used to identify the meaning behind the reviews and it focuses on polarity
i.e., positive, negative, or neutral. Machine Learning is classified as supervised or unsu-
pervised, whereas the Lexicon-based approach analyses and identifies the scores for the
phrases already assigned, and it is used to identify the polarity. As shown in Fig. 1. It is
divided into two approaches: dictionary-based and corpus-based, (Bhalla, 2022) which
use semantic approaches or statistical methods to identify sentiment polarity. If the
dataset is small, algorithms such as Naive Bayes, Support Vector Machines, and Ran-
dom Forest are used to find polarization. Deep Learning concepts like Convolutional
Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term
Memory Networks (LSTM) have also contributed a lot to Natural Language Processing
(NLP).
This study aims to provide a comprehensive comparison between various approaches
in ML and DL in detecting the polarity of sentiments behind students’ feedback based
on various datasets. Students’ feedback is viewed as an effective tool in this research
paper that provides valuable insights when compared to other entities such as courses,
 A Survey on Sentimental Analysis of Student Reviews Using 367

Naïve Bayes

Machine learning

Deep Learning
Decision Tree
Convolutional
Neural Network
K-Nearest
Neighborhood Dictionary

Approach
Lexicon-based
Recurrent Neural Based
Network Approach
Random Forest
Deep belief Corpus
Logistic Network
Based
Regression Approach
Feed Forward
Support Vector Neural etwork
Machines

Fig. 1. Sentiment analysis classifiers

institutions, and so on. The analysis here is based on students’ perspectives, taking into
account all aspects of teaching and learning, including the course, teaching methodology,
course structure, assessment, teachers’ knowledge, behavior, and presentation, among
others. During the last five years, the majority of articles published focused on gathering
opinions and attitudes toward teachers, with only a few papers focusing on institution-
related questions and some on general feedback. According to the literature review,
a greater number of papers focus on gathering opinions and thoughts about student
attitudes toward teachers.

2 Related Work: A Literature Review

Sentimental Analysis focuses mainly on polarity detection and recognition of emotions
towards an entity, which could be based on any event. Sentimental Analysis uses NLP,
ML and DL techniques to discover users’ opinions, identify the sentiments from a vast
amount of information. It can be done at the document, sentence, or aspect level (Cam-
bria, Schuller, Liu, Wang, & Hava, 2012). SA is an interdisciplinary field that deals
with data preprocessing, information retrieval, statistics, and computational linguistics.
Research work related to student feedback is analyzed, which is collected from various
sources, and a few research works related to other domains are also considered to analyse
the various algorithms, methodologies, and models’ accuracy in Machine Learning and
Deep Learning.
A research mainly focuses on sentimental analysis of user preferences with Google
Classroom data for Indonesian language (Situmorang, Chairunnas, & Bon, 2020).
Reviews about Google Classroom are obtained through the Google Play store and data
scrapping techniques for mobile applications, which are stored as .csv format for fur-
ther processing. Preprocessing is performed on the collected data to remove stop words,
punctuations, and stemming. A framework is being developed using impKNN and uses
term weighting using the term frequency-inverse document frequency (TF-IDF) method
to characterize the polarity of opinion in E-Learning. When the TF-IDF calculation is
applied to training and testing data, number of matches is found using confusion matrix
368 J. Jayasudha and M. Thilagu

(Kandhro, 2019) that divides the results of predictions into True Positive (TP), True
Negative (TN), False Positive (FP) and False Negative (FN).
Sentimental Analysis is used to measure, analyze, report, and predict data about
learners to optimize teaching and learning. Here, a method is developed to determine
whether the learner is satisfied with the learning experience. Initially, data is collected
from Amazon, Yelp, and IMDB to find similar characteristics in data. (Omar, 2020) In the
second phase, data transformation takes place. Preprocessing includes tokenization, case
conversion, normalization, stemming, transliteration, and removal of irrelevant content.
During the third phase, data is processed using lexicon and ML classifiers such as Naive
Bayes, Decision Trees. After evaluation, the best approach is selected to know the actual
feedback collected towards the E-Learning environment. Finally, VADER yields the best
results when compared to other approaches, and it is applied in an e-learning environment
to get the best results.
A comparative analysis was carried out with 140 responses collected through Google
forms (Fahmi, 2020). For the best results, Multinomial Nave Bayes, Support Vector
Machine, and Random Forest with unigram and bi-gram were compared. In both cases,
the Multinomial Naïve Bayes classifier produced the best results of 80% when compared
with the Support Vector Machine and Random Forest. A hybrid architecture, combination
of keyword-based components and learning system components, is used to detect hidden
phrase patterns, used to identify syntactic and semantic details, and emotion prediction
is identified based on the knowledge-rich linguistic resources and trained classifiers.
Initially, text is tokenized and split into sentences.
(Dsouza, 2019) POS Tagger is used to annotate information by identifying verbs
and nouns. This content is provided as input for Emotion Intelligence PR, followed by
Machine Learning PR. Finally, it predicts the emotions by building a training model and
applying it. So the output obtained from the keyword-based component acts as an input
for the learning-based component. Here, a hybrid-based architecture is used to detect
emotion that has been validated with experimental results.
An international questionnaire Turkish Tale dataset (Binali, 2010) is used to identify
the sentiments where the data in tales were split into paragraphs and sentences, and
typo errors were also corrected using the Zemberek Library. In supervised machine
learning approaches, multinomial, logistic regression, and Gaussian NB were used. Due
to insufficient data, results were not accurate, and the dataset was improved with more
data to get better accuracy. A final result with 75% accuracy was also obtained using the
Logistic Regression algorithm.
Students’ feedback from Vietnam universities was collected through surveys and
collected feedback was classified into sentiment-based tasks and topic-based tasks, and
labelling was done (Osmanoğlu, 2020). The vntokenizer tool is used for tokenization,
and Maximum Entropy and Naïve Bayes algorithms are compared here. Finally, MaxEnt
performs well when compared to Naïve Bayes with both tasks in corpus.
The research work (Almalki, 2022) used a twitter dataset in the distance learning
domain for identifying sentiments. Using the Twitter API or third-party libraries, data
scraping is done. The API extracted nearly 14,000 tweets and preprocessing is done.
Apache Spark is used for parallel computing, and the Flask API is used for identifying
the sentiments behind Arabic tweets. Finally, Logistic Regression produced an accuracy
 A Survey on Sentimental Analysis of Student Reviews Using 369

of 91%. An approach (Nimala, 2021) collected feedback from around 4895 students and
compared it with four different models like STM,ETM, SWAT, and ET were constructed,
and STM has an accuracy of 86.5 and produced better results compared to other models.
Another model uses the Movie Review dataset to build a model that analyses sen-
timents.The initial model was trained with a 50% manually annotated dataset in the
Hindi language (Rani, 2019) and the experimental results are compared with other ML
approaches, but the proposed CNN model overcomes the performance of other models
and achieves 95% accuracy. The next model uses Deep Learning NN and Applied CNN
on the Lithuanian Internet Comment dataset (Kapočiūtė-Dzikienė, 2019) to analyse the
emotions behind the dataset. It is built on Word2Vec and FastText word embedding. So
the best results were reported with 70.6% accuracy using CNN. So it works better and
produces good results if the dataset is very small.
It was suggested to use an attention-based LSTM model that includes aspect infor-
mation. A novel model was developed to predict the sentiment score of financial opinion
mining and categorise it into a particular pre-defined class. The LSTM layers maintain
the intermediate output to link it to the final distribution representation and take the word
embedding as input. (Shijia, 2018) Additionally, 504 news headlines and 675 microblog
messages are included in the dataset utilised for the experiment. In addition, there were
27 classification labels and an emotion score scale from -1 to 1. The neural network
used the word2vec model as an input for word embedding. A number of deep learning
models were compared, and the results showed that ALA performed the best.
To present a sentiment lexicon (Chen, 2020) established a sentiment analysis frame-
work based on a deep neural LSTM network. The information was created using the
Military Life PTT board from Taiwan’s largest internet forum. The testing findings
demonstrated that the suggested framework’s accuracy and F1-measure were higher
than the results obtained by employing solely existing sentiment dictionaries. LSTM is
used to create a sentiment analysis model for the Roman Urdu language (Ghulam, 2019),
When compared to other machine learning algorithms, the testing findings demonstrated
significant accuracy.
A work that compared five Machine Learning techniques (Altrabsheh, 2014) using
Naïve Bayes, Complement Naïve Bayes, Maximum Entropy, and Support Vector
Machines that automatically analyse the student’s feedback in real-time with a mini-
mum time. Preprocessing is done to remove stop words, punctuation, and numbers, case
conversions and spell checking have been done for a collected 1036 instances. Here,
data is collected in two different ways: real-time collection of feedback in lectures and
end-of-unit feedback. Naive Bayes will not work with uneven datasets and Complement
NB addressed this problem and produced better results where the Naive Bayes algorithm
was implemented in R and CNB was implemented in Weka. Maximum Entropy finds
the weights of features that maximise the likelihood using search-based optimization.
Its performance is very poor when dealing with real-time problems. Support Vector
Machines is effective in text categorization and SVM Radial Basis Kernel works well
with Natural Language Processing (NLP) and it is flexible to use in experiments, SVM
performed better compared to NB. Maximum Entropy performed better for neutral class
results without change in F1-Score, precision, and recall and also prevents over fitting.
370 J. Jayasudha and M. Thilagu

In (Toçoğlu, 2020), three conventional text representation schemes and three N-

grams have been used with classifiers like SVM, Naïve Bayes, Logistic Regression,
and the Random Forest algorithm. Ensemble learners like AdaBoost, Bagging, Random
Space, and Voting algorithms were also evaluated for producing better results. Here the
dataset is divided into 10 mutual folds. When three different classifiers with text repre-
sentation and stemming schemes were used, F5, raw, and snow ball stemming scheme,
raw outperformed the SVM classifier and TF-IDF text representation. Again, N-gram
models are compared against classifiers and text representations, which produced the
results as SVM and RF with unigram performed well. Considering ensemble methods,
it doesn’t produce more changes in the results, so finally TF-IDF outperformed well.
The research work (Kastrati Z. A., 2020) compared Deep Learning and Machine
Learning algorithms using student reviews (21,940 reviews) from an E-Learning plat-
form. SVM, Nave Bayes, Boosting, and DT ML algorithms are combined with TF-IDF,
a 1D-CNN model is built to extract sentiment, and FastText, Word2Vec, and Glove word
embedding’s are used for a better F1 Score of 88.2%. A work consists of 2 datasets:
the rating score dataset and the textual comment dataset, and various ML classifiers
(MLP, Logistic Regression, SVM,RF are compared to identify the best classifier (Lwin,
2020).For labelling the rating score dataset, K-Means clustering is used and trained with
various models where SVM produced a better result, and Naïve Bayes is used for textual
comment classification as it consists of both English and Myanmar language and tested
using 10-fold cross validation with F-Measure-95% best results. 30,0000 textual feed-
back has been collected from college students. After preprocessing (Katragadda, 2020),
it is evaluated using NB, SVM, and ANN classifiers. Finally, ANN produced the best
accuracy of 88.2%. It clearly states that classifiers’ performance is purely based on the
parameters considered and the size of the dataset.
A study was conducted to improve the sentiment classifier by using TF-IDF and n-
grams and ensemble techniques (Pacol, 2021). It converts text into vector values instead
of count vectors in base models where it produces the minimum amount of information
when compared to vector values. Teaching performance data is collected from various
campuses of Pangasinan State University and consists of 9140 sentences, which are
manually cleaned and labelled in an excel file. Considering the performance of n-gram
and base models, n-gram with SVM performed with 0.98 accuracy, whereas n-gram
shows improvement in the performance in ensemble models with Naive Bayes, Logistic
Regression, and Support Vector Machine.
Large volumes of data, such as discussion statements, likes, and follows, as well as
individual system interactions, including timestamps of activities, videos watched, test
outcomes, and logins, are generated by MOOCs and serve as evidence of participant
involvement and experience. The study’s low response rate, (Lundqvist, 2020) which
can skew the sample, is one of its limitations. Only 264 people commented in the free
text comment section during the first week of the course, which is a reflection of how
few people comment in MOOCs overall. According to the study, there is a correlation
between the general tone of postings and comments made about the MOOC, and this
course had a generally positive tone compared with others.
Four Machine Learning techniques such as Support Vector Machines, Naïve Bayes
(NB), Maximum Entropy (ME), and Complement Naïve Bayes (CNB) were compared.
 A Survey on Sentimental Analysis of Student Reviews Using 371

1036 responses were collected, and 641 cases were positive and 292 cases were con-
sidered negative. (Ullah, 2016) After several preprocessing steps, Maximum Entropy
and Support Vector Machines perform well, and the possibility of including more
preprocessing will result in a better outcome.
A model is built with different layers that use the data collected from students of
Vietnam University (Sangeetha, 2021). The inputs are trained using embedding models
such as Glove and Cove, and the output features are fed into the encoding layer. Next,
the output of the encoding layer is fed into a dropout layer that prevents over fitting. The
results of these dropout layers are combined using LSTM, which produces only minimum
features. Finally, everything will be fed into the dense layer for producing output. This
new architecture is less error-prone when compared to the other three models: LSTM
with 86%, LSTM with Attention with 87%, Multi-head LSTM with Attention with 90%,
and the proposed model with 94% accuracy.
MOOC reviews have been collected from Coursera and a framework has been pro-
posed for aspect-based sentimental analysis.This model identified the critical factors that
determine the effectiveness of online courses and also the attitude of students. With key-
words and manually annotated reviews, the proposed framework (Kastrati Z. I., 2020)
is being tested and validated against datasets using LSTM and CNN. LSTM with glove
performed well, with an F1-Score of 86.13%.
Two algorithms have been compared to analyse the Twitter data based on hashtag
keywords and compared the performance with Recurrent Neural Network and Support
Vector Machine through R-Tool. SVM and RNN have been carried out with a sample
of 20, 50, and 600 tweets, Recurrent Neural Network produced better results compared
to Support Vector Machine (Kaur, 2021). Even though Machine Learning and lexicon-
based approaches are used in feedback analysis, Deep Learning methods propose a
better and more efficient system that uses Word2Vec for text processing, CNN (Asmita
S, Anuja T, & Ash, 2019) for automatic feature extraction, and final classification is done
using Support Vector Machine. So, Deep Learning can be used in the education domain
to classify the strengths, weaknesses, and suggestions for a trainer. Various datasets
from Twitter sentiment, the sentiment Tree bank, and online movie reviews (Rojas-
Barahona, 2016) were considered and applied to various Deep Learning architectures
such as convolutional, recurrent, and combined methods. Both fine-grained and binary
classification have been carried out, and comparatively, CNN performs well in polarity
detection compared to supervised models.

3 Methodology
3.1 Datasets
Most of the related research findings identified a few important aspects regarding the
teaching-learning process, including students’ opinions about teaching, institutions, and
courses. So a study was carried out nearly using more than 25 papers based on educational
feedback reviews from students. Mostly, for this kind of analysis, open datasets are lim-
ited and are not available with the required attributes that suit our needs. The data used in
the research papers was taken from various sources like real-time feedback collected from
students of various universities through Google forms, Survey/Questionnaire, social
372 J. Jayasudha and M. Thilagu

media, blogs or forums, MOOC educational platforms (Course era and EdX), online
reviews, and some used blended datasets which may be manually annotated to obtain
the result. Duplicate records should be removed after data collection and, in many cases,
data collected from online forums is unstructured and should be labelled manually and
trained to be used with supervised Machine Learning approaches.

3.2 Approach
The most crucial phase of every machine learning model is data preprocessing (Gottipati,
2018). The performance of the model is significantly influenced by how effectively the
raw data is cleaned and preprocessed. Text preprocessing aims to clean up the data so that
it can be processed in the next step. Text preprocessing makes the input documents more
consistent to facilitate text representation, which is necessary for most text analytics
tasks. Tokenization breaks the sentences into words, stemming obtains the root words
from tokens; and lemmatization creates a single base word for various inflected forms,
Part of Speech is used for identifying parts of speech in tokens, Named Entity Recognition
(NER) classifies the text into various categories.
Figure 2 shows preprocessing techniques in NLP like tokenization, lemmatization,
parsing, and bag of words that help us analyse the sentiments in text. Initially, after
getting feedback, input is split into sentences or words, and, using tokenization, POS
tagging, and lemmatization, it converts sentences into words. Then a sentiment score is
calculated from these tokenized words and, based on the score, the sentiment is cate-
gorized. Sentiment can be classified using either Machine Learning or Deep Learning
Models, so appropriate models should be selected to get the relevant output.

Pre-Processing Feature Extraction Modelling

TF-IDF Naïve Bayes

Tokenization Decision Tree
Stemming 1-Hot Encoder
Count Vectorizer Random Forest
Input Lemmatization SVM, KNN Output
Part of Speech N-gram
Word Embed- Logistic
NER Regression,
ding,LSTM
CNN,RNN

Fig. 2. Sentimental analysis using machine learning/deep learning

An automated process that analyses and discriminates sentiments as positive, nega-

tive, or neutral based on the emotions expressed by users in the form of text.NLP identifies
the core data that helps us understand students’ sentiments. SA classifies huge amounts
of real-time data in a few seconds, which is very complicated to manually classify and it
also results in erroneous results. Sentiment can be analysed through Machine Learning
and lexicon-based approaches. Machine learning algorithms need data for training as
it is to be collected and labelled manually and it is a time-consuming process (Singh,
2017). Supervised ML algorithms like Naïve Bayes, Decision Tree, K-Nearest Neigh-
borhood, Random Forest, Logistic Regression (Ghulam, 2019) Support Vector Machine
 A Survey on Sentimental Analysis of Student Reviews Using 373

algorithms are mostly used to identify the sentiment of any review or sentence. Natural
language processing, text analysis, computational linguistics, and other approaches are
used to identify and quantify the sentiment of text in sentiment analysis. For better results,
these algorithms require labelled data as a dataset and can be combined with lexicon
approaches (dictionary-based or corpus-based) and classifiers such as Ada boost, N-
Model, Bagging, and Random Space and Aspect based classification methods (Toçoğlu,
2020) are also used with TF-IDF statistical method (Kastrati Z. A., 2020), activation
functions like Tanh, sigmoid, and ReLu (Chen, 2020) are combined with ML algorithms
like SVM, Multinomial NB, and Logistic regression for better accuracy. Table 1 demon-
strates various methods used in literature works, datasets used, algorithms compared,
new models developed, and results.
Deep learning has emerged as a powerful machine learning technique that learns
multiple layers of representations or features of the data and produces results (Zhang,
2018). Along with the success of deep learning in many other application domains, it
has also been popularly used in sentiment analysis in recent years. Due to their high
performance (Yadav, 2020) Deep Learning algorithms like Convolutional Neural Net-
work (CNN), Recurrent Neural Network (RNN), and Deep Belief Networks (DBN) are
the most frequently used algorithms for Sentimental Analysis. Ensemble methods and
architectures can also be used to identify the polarity of words that learn features through
layers that are fed directly into another layer as input (Ghorbani, 2020).
Lexicon approaches like corpus-based and dictionary-based approaches can perform
sentiment analysis without training, and it needs explicit vocabulary for separate lan-
guage as it is unsupervised. The tokenization process separates each word in a review or
post and, based on the emotion found, it is assigned as positive, negative, or neutral flags
based on words already available in the lexicon. Some tools like VADER (Kastrati Z. D.,
2021; Dsouza, 2019) & (Kastrati Z. D., 2021),SentiWordNet, WordNet and SentiBank
are publicly available to analyse the sentiment behind the words.

Table 1. Literature review- machine learning, lexicon based and deep learning methods for
sentimental analysis

Year Reference Model Dataset Results

2022 (Almalki, 2022) Logistic Regression with Arabic Tweets from Logistic Regression- 91%
Apache spark, Flask API Twitter API SVM -69%
2021 (Pacol, 2021) Naïve Bayes, SVM, Pangasinan State N-Gram with
Logistic University-1827 SVM -98%
regression.TF-IDF,n-grams, statements
ensemble methods
2020 (Situmorang, Chairunnas, & Imp KNN with TF-IDF LMS reviews Accuracy -84%
Bon, 2020) through Google
Classroom and Play
Store
(continued)
374 J. Jayasudha and M. Thilagu

Table 1. (continued)

Year Reference Model Dataset Results

2020 (Omar, 2020) VADER Classifier, IMDB, Yelp and VADER
Naïve Bayes and Decision Amazon Accuracy– 90%
Tree NB (amazon)-81%
DT (Amazon)-76%
2020 (Fahmi, 2020) SVM & SVM-PSO, Naïve LMS Dataset SVM-PSO with
Bayes,KNN through API (2154 Sastrawi – 82.27%
Reviews) SVM-PSO without
(Feature Reduction) Sastrawi – 82.09%
2020 (Osmanoğlu, 2020) Decision tree, KNN, MLP, International & Logistic Regression
XGB, NB, SVM Turkish Tale (CT + SC) – 77.5%
Logistic Regression, Dataset, Data from DT- 76.5%
LMS (Study
materials) 2421
Comments
2020 (Toçoğlu, 2020) Classifiers, Text Student Reviews SVM Accuracy
Representations and 698 TF *idf = 73%
Stemming Schemes Reviews TF – 70.6%
TP – 69%
2020 (Sangeetha, 2021) LSTM,LSTM + ATT, Vietnam University Fusion Model Achieves
MULTIHEAD, (16,175 statements) better Accuracy
FUSION
2020 (Kastrati Z. D., 2021) LSTM,CNN Coursera reviews F1-score
(5989 reviews) LSTM + Glove -93.3%
2020 (Katragadda, 2020) Naïve Bayes,SVM and ANN 30000 Students NB-50%
Review SVM -60.8%
ANN -88.2%
2020 (Kastrati Z. A., 2020) SVM, Naive Bayes, 21940 students’ CNN -88.2%
Boosting, Decision Tree, reviews DT (TF)-77.9%
1D-CNN (Aspect based DT(TF*idf)– 88.67%
Classification)
2019 (Dsouza, 2019) MNBC, SVM, RF for Google Form data MNBC -80% produced
unigram and bi-gram (140 Responses) best results
2019 (Nimala, 2021) STM,ETM,SWAT, Feedback form STM Accuracy = 86.5%
ET Models (4895 Responses) ETM -74.5%
2019 (Kapočiūtė-Dzikienė, 2019) ML (SVM,MNB) Lithuanian LSTM,CNN– 61%,70%
DL (LSTM,CNN) Internet comments SVM,MNB–72%,73.5%
(10,570 comments)
2019 (Chen, 2020) LSTM,Bi-LSTM Military life PTT LSTM – 90.11%
Bi-LSTM – 92.68%
2019 (Ghulam, 2019) NB,RF,SVM,Deep Neural Roman Urdu dataset NB – 77.2%
Network RF – 88.6%
SVM – 92.4%
DNN- 95.2%
(continued)
 A Survey on Sentimental Analysis of Student Reviews Using 375

Table 1. (continued)

Year Reference Model Dataset Results

2018 (Rani, 2019) CNN Movie Review CNN (3 Layers) - 95%
Dataset CNN (2 Layers) – 93.4%
2018 (Shijia, 2018) RNN-LSTM 504 news headlines ALA outperformed well
and 675 microblog (Attention based LSTM
messages model with aspect
information)
2018 (Moreno-Marcos, 2018) Logistic Regression, MOOC AUC Performed well
NB, SVM,DT, RF, (13,300 Reviews) compared to Kappa
UnSup (Dictionary,
SentiWordNet)
2018 (Van Nguyen, 2018) Naïve Bayes and Maximum Vietnam Students ME(Sentiment Based)-
Entropy Dataset 87.9%
(16000 Responses) ME (Topic Based)-84%
2017 (Singh, 2017) Naïve Bayes,J48, BFTree Woodlands Wallet J-48 & OneR
and OneR (100 Reviews), performed well
Camera Review
(8000 Reviews),
Movie Review
(2900 Reviews)
2016 (Ullah, 2016) SVM,ME,NB,CNB University of SVM- 96.2%
Portsmouth ME – 87%
(1036 reviews)
2014 (Altrabsheh, 2014) NB,CNB,SVM Real-time feedback CNB(With Kernal)
Linear, Poly and collection and end - 84%
SVM RB of unit feedback CNB (Without Kernel)
(1036 Instances) – 80%
SVM(Without Kernel)
– 94%
SVM (with Kernel) -93%

4 Research Challenges and Future Directions

Instead of a fine-grained analysis that is related to various characteristics and the sen-
timents associated with them, the majority of the study concentrated on a thorough
examination of sentimental analysis. Possibility of using figurative language, such as
sarcasm, irony, and other complex types of discourse in which authors use language
to convey the opposite meaning. Even for humans, recognizing irony can be difficult
and complex, resulting in a lot of misunderstandings in everyday life. Sarcasm and
irony detection in natural language processing is challenging, particularly for sentiment
analysis. For a better classification, it’s also important to consider identifying irony, sar-
casm, and figurative speech, examining emoticons, emoji’s, and emoticons, as well as
assessing negative and double-negative words. Only a small number of open datasets
with insufficient properties are available. Only a few sources are accessible to identify
the lexical components. Therefore, there is a demand for more datasets and more emo-
tional analysis publications. Researchers have utilized a variety of evaluation metrics to
assess how well different methodologies perform, and by selecting the best measures,
any mode’s effectiveness is improved. The extraction of named entities, which can refer
376 J. Jayasudha and M. Thilagu

to concrete people or abstract notions, is one of the fundamental difficulties with natural
texts. Expand the most sophisticated model to include sentiment analysis across several
languages. Consequently, a multi-language processing paradigm must be put in place.
There are over 80 dictionaries available for other languages. Very little research is done
in code-mix languages, and the majority of work is done in English. So this case needs to
be considered in future work. Many studies simply translate all comments into English
using machine translation, although this may cause emotions to be lost in the translation.
Finding irony in Natural Language Processing (NLP), especially in sentiment anal-
ysis, is a difficult issue. Satirical statements may lead to inaccurate categorization in the
findings of sentiment analysis and data mining. Huge datasets are typically needed for
deep learning techniques, but real-world datasets are frequently unbalanced.

5 Conclusions
The educational sector is undergoing a tremendous transformation because of online
education. Massive amounts of structured and unstructured data are collected through
online platforms, which include MOOCs, EdX, Coursera, etc. Students’ feelings can be
gauged through posted feedback, forums, and reviews. This helps the teachers to identify
the students’ attitudes, feelings, and behavior to make necessary improvements in teach-
ing methodology to maintain the students’ retention rate. Sentimental Analysis research
has increased in recent years, and while initially only Machine Learning algorithms
were considered, Deep Learning algorithms are now being used to identify the polarity.
Examining the Literature for Support Vector Machines and Naive Bayes are most fre-
quently used in Machine Learning models, while Convolutional Neural Networks and
Recurrent Neural Networks (RNN) are used in Deep Learning models (CNN). Research
is required, taking into account the difficulties in analysing feedback and determining
the best model in Machine Learning or Deep Learning that results in high accuracy. As
a result, Sentimental Analysis based recommendation systems analyse comments and
extract precise information about students’ expectations, teaching methodology, their
interests, and expected curriculum updates to be embedded in the education sector to
vastly improve every aspect of education.

References
Almalki, J.C.: A machine learning-based approach for sentiment analysis on distance learning
from Arabic Tweets (2022)
Altrabsheh, N., Cocea, M., Fallahkhair, S.: Learning sentiment from students’ feedback for real-
time interventions in classrooms. In: Bouchachia, A. (ed.) ICAIS 2014. LNCS (LNAI), vol.
8779, pp. 40–49. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-11298-5_5
Asmita, S.S., Anuja, T.D., Ash, D.: Analysis of student feedback using deep learning. Int. J.
Comput. Appl. Technol. Res. 8, 161–164 (2019)
Mite-Baidal, K., Delgado-Vera, C., Solís-Avilés, E., Espinoza, A.H., Ortiz-Zambrano, J., Varela-
Tapia, E.: Sentiment analysis in education domain: a systematic literature review. In: Valencia-
García, R., Alcaraz-Mármol, G., Del Cioppo-Morstadt, J., Vera-Lucio, N., Bucaram-Leverone,
M. (eds.) CITI 2018. CCIS, vol. 883, pp. 285–297. Springer, Cham (2018). https://doi.org/10.
1007/978-3-030-00940-3_21
 A Survey on Sentimental Analysis of Student Reviews Using 377

Bhalla, R.: A review paper on the role of sentiment analysis in quality education. SN Comput.
Sci. 3(6), 1–9 (2022)
Binali, H.W.: Computational approaches for emotion detection in text. In: 4th IEEE International
Conference on Digital Ecosystems and Technologies, pp. 172–177. IEEE (2010)
Cambria, E., Schuller, B., Liu, B., Wang, H., Havasi, C.: Guest Editorial Special Issue on Concept-
Level Opinion and Sentiment Analysis. IEEE, (IF:2.570, 5-year IF:2.632(2010)), pp. 15-21
(2012)
Chen, L.C.: Exploration of social media for sentiment analysis using deep learning. Soft. Comput.
24(11), 8187–8197 (2020)
Dsouza, D.D.: Sentimental analysis of student feedback using machine learning techniques. Int.
J. Recent Technol. Eng. 8(14), 986–991 (2019)
Fahmi, S.P.: Sentiment analysis of student review in learning management system based on sastrawi
stemmer and SVM-PSO. In: 2020 International Seminar on Application for Technology of
Information and Communication (iSemantic), pp. 643–648 (2020)
Ghorbani, M., Bahaghighat, M., Xin, Q., Özen, F.: ConvLSTMConv network: a deep learning
approach for sentiment analysis in cloud computing. J. Cloud Comput. 9(1), 1–12 (2020).
https://doi.org/10.1186/s13677-020-00162-1
Ghulam, H.Z.: Deep learning-based sentiment analysis for roman urdu text. Procedia Comput.
Sci. 147, 131–135 (2019)
Gottipati, S., Shankararaman, V., Lin, J.R.: Text analytics approach to extract course improvement
suggestions from students’ feedback. Res. Pract. Technol. Enhanc. Learn. 13(1), 1–19 (2018).
https://doi.org/10.1186/s41039-018-0073-0
Kandhro, I.A., et al.: Student feedback sentiment analysis model using various machine learning
schemes: a review. Indian J. Sci. Technol. 12(14), 1–9 (2019)
Kapočiūtė-Dzikienė, J.D.: Sentiment analysis of lithuanian texts using traditional and deep
learning approaches. Computers 8(1), 4 (2019)
Kastrati, Z.A.: Aspect-based opinion mining of students’ reviews on online courses. In: Pro-
ceedings of the 2020 6th International Conference on Computing and Artificial Intelligence,
pp. 510–514 (2020)
Kastrati, Z.D.: Sentiment analysis of students’ feedback with NLP and deep learning: a systematic
mapping study. Appl. Sci. 11(9), 3986 (2021)
Kastrati, Z.I.: Weakly supervised framework for aspect-based sentiment analysis on students’
reviews of MOOC. IEEE Access 8, 106799–106810 (2020)
Katragadda, S.R.: Performance analysis on student feedback using machine learning algorithms.
In: 2020 6th International Conference on Advanced Computing and Communication Systems
(ICACCS), pp. 1161–1163. IEEE (2020)
Kaur, H.A.: A proposed sentiment analysis deep learning algorithm for analyzing COVID-19
tweets. Inf. Syst. Front. 23(6), 1417–1429 (2021)
Lundqvist, K.L.: Evaluation of student feedback within a MOOC using sentiment analysis and
target groups. Int. Rev. Res. Open Distrib. Lear. 21(3), 140–156 (2020)
Lwin, H.H.: Feedback analysis in outcome base education using machine learning. In: 17th Inter-
national Conference on Electrical Engineering/Electronics, Computer, Telecommunications
and Information Technology (ECTI-CON), pp. 767–770. IEEE (2020)
Moreno-Marcos, P.M., Alario-Hoyos, C., Muñoz-Merino, P.J., Estévez-Ayres, I., Kloos, C.D.:
Sentiment analysis in MOOCs: a case study. In: 2018 IEEE Global Engineering Education
Conference (EDUCON), pp. 1489–1496. IEEE (2018)
Nimala, K.: Sentiment topic emotion model on students feedback for educational benefits and
practices. Behav. Inf. Technol. 40(3), 311–319 (2021)
Omar, M.A.: Sentiment analysis of user feedback in e-learning environment. Int. J. Eng. Trends
Technol. (IJETT), 153–157 (2020)
378 J. Jayasudha and M. Thilagu

Osmanoğlu, U.Ö.: Sentiment analysis for distance education course materials: a machine learning
approach. J. Educ. Technol. Online Learn. 3(1), 31–48 (2020)
Pacol, C.A.: Enhancing sentiment analysis of textual feed-back in the student-faculty evaluation
using machine learning techniques. Eur. J. Eng. Sci. Technol. 4(1), 27–34 (2021). https://doi.
org/10.33422/ejest.v4i
Rani, S.: Deep learning based sentiment analysis using convolution neural network. Arab. J. Sci.
Eng. 44(4), 3305–3314 (2019)
Rojas-Barahona, L.M.: Deep learning for sentiment analysis. Lang. Linguist. Compass 10(12),
701–719 (2016)
Sangeetha, K., Prabha, D.: Sentiment analysis of student feedback using multi-head attention
fusion model of word and context embedding for LSTM. J. Ambient Intell. Humanized Comput.
12(3), 4117–4126 (2020). https://doi.org/10.1007/s12652-020-01791-9
Dos Santos, C., Gatti, M.: Deep convolutional neural networks for sentiment analysis of short
texts. In: Proceedings of COLING 2014, The 25th International Conference on Computational
Linguistics: Technical Papers, pp. 69–78 (2014)
Shijia, E.Y.: Aspect-based financial sentiment analysis with deep neural networks. In: WWW
(Companion Volume) (2018)
Singh, J., Singh, G., Singh, R.: Optimization of sentiment analysis using machine learning
classifiers. HCIS 7(1), 1–12 (2017). https://doi.org/10.1186/s13673-017-0116-3
Situmorang, B.H., Chairunnas, A., Bon, A.T.: Sentiment analysis of user preferences on learning
management system (Lms) platform data. In: 2nd African International Conference on Industrial
Engineering and Operations Management, IEOM 2020, pp. 1784–1789 (2020)
Toçoğlu, M.A., Onan, A.: Sentiment analysis on students’ evaluation of higher educational insti-
tutions. In: Kahraman, C., Cevik Onar, S., Oztaysi, B., Sari, I.U., Cebi, S., Tolga, A.C. (eds.)
INFUS 2020. AISC, vol. 1197, pp. 1693–1700. Springer, Cham (2021). https://doi.org/10.1007/
978-3-030-51156-2_197
Ullah, M.A.: Sentiment analysis of students feedback: a study towards optimal tools. In: 2016
International Workshop on Computational Intelligence (IWCI), pp. 175–180. IEEE (2016)
Van Nguyen, K.N.: UIT-VSFC: vietnamese students’ feedback corpus for sentiment analysis. In:
2018 10th International Conference on Knowledge and Systems Engineering (KSE), pp. 19–24
(2018)
Wen, M.Y.: Sentiment analysis in MOOC discussion forums: what does it tell us? In: Proceedings
of the 7th International Conference on Educational Data Mining (EDM 2014), pp. 130–137
(2014)
Yadav, A.: Sentiment analysis using deep learning architectures: a review. Artif. Intell. Rev. 53(6),
4335–4385 (2020)
Zhang, L.W.: Deep learning for sentiment analysis: a survey. Wiley Interdisc. Rev. Data Min.
Knowl. Disc. 8(4), e1253 (2018)
 Rice Yield Estimation Using Deep Learning

Niyati Mishra, Sushruta Mishra(B) , and Hrudaya Kumar Tripathy

School of Computer Engineering, Kalinga Institute of Industrial Technology (KIIT) Deemed

to be University, Bhubaneswar, Odisha, India
[email protected], {sushruta.mishrafcs,
hktripathyfcs}@kiit.ac.in

Abstract. Global and regional food security heavily relies on effective yield esti-
mation results. Thus precise and on-time rice yield estimate or prediction is a
pivotal factor not only to ensure food security but also for sustainable develop-
ment of agricultural resources. Machine learning and deep learning are proving to
be exemplary support tools for decision making for rice yield estimation or predic-
tion, such as selection of the rice varieties that need to be grown and also decisions
involving the management of crops during growing season. Several researchers
have put forth a variety of deep learning as well as machine learning algorithms
that have helped estimate rice yield time and again. This paper proposes a LSTM
based model to predict the Rice yield of the data collected for all 314 blocks of
Odisha by ICAR - National Rice Research Institute (NRRI), Odisha. In this study,
we get 0.07 RMSE score for training data and 0.21 RMSE score for test data. The
model is also evaluated based on the various performance metrics for three rice
datasets. The overall performance for the rice datasets is evaluated to be 0.989
recall, 0.979 precision, 0.989 accuracy and 0.984 F1 score.

Keywords: Machine learning · Deep learning · Rice yield estimation · Food

security · Sustainable development

1 Introduction

The proper and effective estimates of rice yield helps strengthen national food security
by ensuring timely export and import decisions, thereby helping the management make
proper and informed decisions [1]. For instance, in order to breed superior varieties of
crops, seed companies need a prediction of the relative performance measures of the
newly developed hybrids in diverse environments and yield prediction would also help
farmers or cultivators to make informed financial and management related decisions.
Yet, estimation of rice yield is one of the most challenging field out there due to various
complicated factors associated with it. One such factor being genotype data which com-
prises of high-dimensional marker data which in turn contains thousands to millions of
markers for every individual plant which is an arduous task as there needs to an accurate
estimation of the effects of these genetic markers which under control of a diversity of
field management practices or environmental conditions.
Estimation of rice yield is recently heavily focused on the use of various deep learning
techniques, for instance deep neural networks. Deep neural networks (DNN) are a type of

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 379–388, 2022.
https://doi.org/10.1007/978-3-031-23233-6_28
380 N. Mishra et al.

representation learning models which help in extraction of underlying data representation

without the need of custom-made input attributes. Some major instances of the use of
deep learning techniques are described here. In [2], authors proposed a technique for
the prompt and expeditious estimation of yield of rice at pixel level by incorporating
a deep learning based model and a crop model for varied agriculture based systems
available in South Korea and North Korea. Similarly in [3], authors incorporated CNN
based framework using RGB images for estimating rice yield. The said model estimated
a variation of 70% in rice yield along with a RMSE score of 0.22.
Although Deep Learning is one of the promising technologies used for estimation of
rice yield, there are still a few challenges that are faced here in terms of accuracy of the
model, inclusion of a greater diversity of features, etc. Many existing works based on
different classifiers on rice dataset are implemented but majority of these models lack
preciseness and are not so reliable. Also the overall error rate is found to be on higher
end with all the existing models. In this study, we have developed a LSTM based model
to predict the rice yield of data collected from 314 blocks of Odisha. The main objective
of the study is to discuss the use of deep learning techniques in estimation of rice yield.
The main contributions of the work are as follows:

• Aim is to address technical reliability issues associated with the rice crop yield
prediction domain.
• Develop a novel hybrid LSTM model for predicting rice yield in farming zones with
better performance.
• The proposed model is validated against performance metrics and the outcome noted
were promising.

The paper is segregated into various sections which are briefly described here. In
Sect. 1, a brief introduction of the previous research papers is discussed. Section 3 talks
about the problem generally faced in yield prediction. Section 4 discusses the proposed
model. In Sect. 5, we discuss about the result and the overall capabilities of the proposed
model, and lastly, we summarize the various observations formed with the help of this
model under conclusion and future scope title in Sect. 6.

2 Literature Review
In this section, we have listed quite a number of reputed works that have been done
in the domain of Estimation of yield of various crops with the aid of various Machine
Learning and Deep Learning algorithms. Tian et al. have proposed a model that incorpo-
rates deep learning based methods for the estimation of yield of winter wheat. The said
model uses meteorological data and two remotely sensed indices [4]. The said incor-
porates Long Short-Term Memory (LSTM) neural network with an attention feature
(ALSTM) attached to the model, where the pivotal idea is to focus the attention to the
quintessential parts of the sequence of the input which has maximum impact on the target
vectors, thereby ensuring the accurate extraction of the specific features or attributes.
Sonal Agarwal and Sandhya Tarar have proposed a model which is enhanced by the
incorporation of deep learning methodologies that not only provide accurate yield esti-
mates but also precise information in regard to the quantities of soil ingredients needed
 Rice Yield Estimation Using Deep Learning 381

and also the necessary monetary expenses that come with it [5]. The said model incor-
porates a dynamic combination of RNN, LSTM and SVM algorithms. Wang et al. have
devised a model that helps estimate the yield of winter wheat in the primary harvesting
areas in China with the use of various deep learning methodologies [6]. E. Banu and
Dr. A. Geetha incorporated both DNN and random forest in their research work for rice
yield estimation [7]. Sun et al. have proposed a model that incorporates both LSTM and
CNN algorithms for the estimation of yield of soybean [8]. This paper was trained with
the help of both environment parameters and crop growth parameters such as weather
data, Surface Reflectance data as well as the Land Surface Temperature data. Nevavuori
et al. have incorporated CNN based framework in their study that uses RGB data along
with a particular index data that are collected by UAVs [9]. Chen et al. proposed a model
that incorporates R-CNN (region-based convolutional neural network) for detecting the
quantity of flowers, unripe strawberries, and ripe strawberries with regard to the estima-
tion of strawberry yield [10]. A 3-dimensional CNN based framework was incorporated
into the model devised by Russello for estimation of crop yield purpose which showed
quite fascinating results in comparison to various dynamic machine learning method-
ologies [11]. Jiang et al. devised a model incorporating LSTM framework for estimating
yield of corn crops by primarily focusing on soil and meteorological data which paved
way to quite a good performance score [12]. Kulkarni et al. devised a model incorpo-
rating RNN framework that made use of soil data along with various rainfall sequences
in a particular designated region for enhancement of yield estimates of the crops [13].
You et al. proposed a model that introduced a component involving Gaussian process
into a LSTM or CNN framework into his model and this said model outperformed prim-
itive various remote sensing methodologies by 30% with regard percentage error [14].
Alhnaity et al. incorporated LSTM framework into his model for predicting both plant
growth variation and crop yield for two different layouts, i.e., Ficus benjamina stem
growth and tomato yield prediction, in supervised greenhouse environments [15]. Some
studies have shown that while CNN has the ability to explore more number of spatial
features, LSTM can help divulge phenological characteristics [16]. Khaki et al. proposed
a model that incorporates YieldNet which is one of the latest deep learning technique
that makes use of a novel deep neural network (DNN) framework [17].

3 Problem Statement
Precise and timely prediction of crop yield information is crucial for making accurate
agricultural resolutions and non-expendable decisions which has a direct impact on a
country’s economy [18]. Deep learning has emerged as a front runner among the vari-
ous promising technologies in this regard. However, we can see that there are multiple
challenges associated with using deep learning. These include integrating various inter-
dependent factors into the model for better results, incorporating prior knowledge with
existent and current data, increasing the accuracy and thus credibility of the model etc.
So, in this paper we have tried to tackle some existing issues by proposing a deep learn-
ing based model which uses LSTM as the primary deep learning algorithm to provide
more accurate and faster crop estimation results. We intend to explore further into the
issues involved with the aforementioned domain of crop yield estimation so that we can
build a robust deep learning model.
382 N. Mishra et al.

4 Proposed Model
In this study, we are proposing a LSTM (Long Short-Term Memory) network based
model as can be seen in Fig. 1. LSTM is a better improved version of RNN that helps
overcome the vanishing gradient problem [19]. LSTM has the capability of processing
complete data sequences such as videos for example and not just singe data points like
images. An LSTM unit normally comprises of a cell, forget gate, output gate and input
gate. A cell is responsible for keeping track of values over random time periods, while
the three aforementioned gates helps modulate the flow of data into and out of the cell.
Let’s have a detailed explanation on how a LSTM network works. These are the
steps normally followed by a LSTM network:

Step – 1: Initially we start with forget gate. Here we can make a decision as to which bits
of the cell state can prove to be of use with the help of the new input data and previous
hidden state. So, basically the forget gate is responsible for making a decision regarding
which fragments of the long term memory can be forgotten based on the given new input
data and the previous hidden state.
Step – 2: The following step includes the input gate along with the new memory network.
The objective here is to decide what latest information is to be added to the network’s
long term memory (or) the state of the cell, with the help of the new input data and
previous hidden state.
Step – 3: Once we have made changes to the long term memory of the network, we can
then jump to the final step that focuses on determining the new hidden state by making
use of the output gate. The output gate will make use of these three things, i.e., the new
input data, the newly updated cell state and the prior hidden state.

Fig. 1. Architecture of a LSTM network

In this study, we make use of a data set comprising of the various soil parameters
such as soil pH, bulk density, soil organic carbon, clay fraction, soil nutrients (nitrogen,
phosphorus and potassium), farm yard manure and rice yield of all 314 blocks of state
Odisha collected by ICAR - National. Rice Research Institute (NRRI), Odisha. Once the
 Rice Yield Estimation Using Deep Learning 383

data is fed to the model, it is then used to predict the yield. 70% of the data is taken as the
training data while the remaining 30% is taken up as the testing data. Once the estimation
of the rice yield is obtained, RMSE score is computed for both the aforementioned sets of
data to give an estimate of the capabilities of the proposed model [20–22]. The overview
of the proposed framework is highlighted in Fig. 2.

Fig. 2. Overview of proposed model

The overall pseudocode of LSTM model is highlighted below.

Algorithm: Long Short-Term Memory (LSTM)

Input: Standardized Rice dataset
Output: Rice yield prediction

Step 1: Input considered as the present value, past state and the past internal cell
state
Step 2: Set values for forget gate, input gate, input modulation gate and output gate.
Step 3: Updated values for different gates computed as:
Step 3.1: Find parameterized values of the present input and past state for
each gate.
Step 3.2: Activation function applied to each gate on parameterized vectors.
Step 4: Present internal state calculated and then find the addition of two vectors.
Step 5: Compute present hidden state and perform multiplication with output state.
384 N. Mishra et al.

5 Results and Discussion

Upon running the code in jupyter notebook using python, using the Rice crop dataset
comprising of the various soil parameters such as soil pH, bulk density, soil organic
carbon, clay fraction, soil nutrients (nitrogen, phosphorus and potassium), farm yard
manure and rice yield of all 314 blocks of state Odisha collected by ICAR - National
Rice Research Institute (NRRI), Odisha, we get 0.07 RMSE score for training data and
0.21 RMSE score for test data. As we know that lower RMSE score indicates that the
said model is comparatively fits the given dataset better, we can say the model is a better
fit as it has comparatively lower RMSE score, though there is still a bit of an overfitting
issue as the training data performs slightly better than the testing data. We gave also used
other performance metrics for calculating the efficiency of the model that is proposed in
this study.

Table 1. Performance Metrics values of proposed LSTM model

Performance metrics Rice Yield dataset by Rice: All-India Area, Rice: State wise Yield
ICAR [Dataset – 1] production and Yield (Source: Directorate of
along with coverage Economics &
under Irrigation Statistics, DAC&FW)
[Dataset – 2] [Dataset – 3]
recall 0.989 0.977 0.982
precision 0.979 0.970 0.969
accuracy 0.989 0.969 0.977
F1 score 0.984 0.971 0.965

Table 1 is the overview of the different performance metrics like recall, precision, F1
score and accuracy of the proposed LSTM model for different rice datasets. Precision can
be measured as the ratio of number of accurately classified Positive samples to the total
number of positively classified samples, thus precision indicates the reliability of the
model in the classification of the Positive samples. Higher value of precision indicates
that the model is capable of making more accurate Positive classifications while making
fewer wrong Positive classifications. Recall can be measured as the ratio of number of
accurately classified Positive samples to the total number of Positive samples, thus recall
indicates the capability of the model to discover Positive samples. Higher value of recall
indicates more number of positive samples identified [21]. Accuracy helps determine
the overall performance of the model. It is the ratio of number of the accurately made
predictions to the total number of predictions. F1 score is measured as the mean of recall
and precision, so higher F1 score is achieved only when precision and recall are on the
higher side. Our study shows significantly higher values and thus better performance
with regard all the aforementioned datasets, for all the three different datasets that are
used in this study.
The data set is fed into a LSTM network with input shape of dimension (1, 12) and
run for 100 epochs. Batch size is chosen to be 10. RMSE (Root Mean Square Error) is the
 Rice Yield Estimation Using Deep Learning 385

loss function that has been applied on the said model. The graph for the loss function is
depicted in Fig. 3. This graph is based on dataset – 1. The activation function is Rectified
Linear Unit (ReLU).

Fig. 3. A graphical depiction of the loss function, i.e., mean squared error

We also perform validation of the proposed model. Initially we split the whole dataset
into a 70% training and 30% testing dataset. Further, we take 20% of the training data
as the validation dataset. So, we have 30% testing data, 56% training data and 14%
validation data. Then we plot a training loss versus validation loss plot to get an estimate
about the performance of the model as can be seen in Fig. 4. The green curve in the plot
represents the training curve which gives us an estimation of how well the model can
learn based on the training dataset while, the blue curve, i.e., the validation curve gives
us an estimate of how well the model is generalizing based on the validation dataset.
As we can see from the plot, our model is facing some overfitting issues as the training
loss happens to keep decreasing with further learning while the validation loss starts
increasing again with further learning after decreasing to a certain point (Fig. 5).
Figure 4 depicts the analysis of the proposed LSTM model with other classification
models in context to accuracy rate. It is observed that LSTM model generates an optimum
accuracy of 0.989 as compared to other models. SVM records the least accuracy value.
386 N. Mishra et al.

Fig. 4. Training loss vs Validation loss (Color figure online)

0.989
1
0.945
0.95 0.934
0.92 0.915

0.9 0.876

0.85

0.8
LSTM KNN RNN MLP SVM CNN

Accuracy rate

Fig. 5. Accuracy analysis with different classifiers

6 Conclusion
Estimation of rice yield plays a pivotal role in global food production. In order to
strengthen national food security by ensuring timely export and import decisions, the
management folks as well as farmers heavily rely on accuracy of yield predictions [23–
25]. Estimation of rice yield is one of the most challenging field out there due to various
complicated factors associated with. This paper focuses on integrating one of the most
sought-after deep learning algorithms with the aim of estimating the yield of rice. It
proposes a LSTM network based model with the intention of estimating the rice yield
which makes use of a data set comprising for various soil parameters for rice research
provided by ICAR - National Rice Research Institute (NRRI), Odisha. The said model
still has issues like overfitting, etc. that needs to be addressed in the future works so that
we can expand the research further in this domain. Another important future scope would
be to avail datasets of different varieties of crops other than rice and also to integrate
more features besides soil parameters that also affect crop yield.
 Rice Yield Estimation Using Deep Learning 387

References:
1. Chaudhury, P., Mishra, S., Tripathy, H.K., Kishore, B.: Enhancing the capabilities of stu-
dent result prediction system. In: Proceedings of the Second International Conference on
Information and Communication Technology for Competitive Strategies, pp. 1–6 (2016)
2. Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.C.: Convolutional LSTM
network: a machine learning approach for precipitation nowcasting. arXiv 2015 arXiv:1506.
04214 (2015)
3. Tripathy, H.K., Mishra, S., Thakkar, H.K., Rai, D.: CARE: a collision-aware mobile robot
navigation in grid environment using improved breadth first search. Comput. Electr. Eng. 94,
107327 (2021)
4. Tian, H., et al.: A deep learning framework under attention mechanism for wheat yield esti-
mation using remotely sensed indices in the Guanzhong Plain, PR China (2021). https://doi.
org/10.1016/j.jag.2021.102375
5. Agarwal, S., Tarar, S.: A hybrid approach for estimation of rice yield using machine learning
and deep learning algorithms. J. Phys.: Conf. Ser. 1714, 012012 (2021)
6. Wang, X., Huang, J., Feng, Q., Yin, D.: Winter Wheat yield prediction at county level and
uncertainty analysis in main wheat-producing regions of china with deep learning approaches
(2020). https://doi.org/10.3390/rs12111744
7. Mishra, S., Tripathy, H.K., Acharya, B.: A precise analysis of deep learning for medical image
processing. In: Bhoi, A., Mallick, P., Liu, C.M., Balas, V. (eds.) Bio-inspired Neurocomput-
ing. Studies in Computational Intelligence, vol. 903, pp. 25–41. Springer, Singapore (2021).
https://doi.org/10.1007/978-981-15-5495-7_2
8. Sun, J., Di, L., Sun, Z., Shen, Y., Lai, Z.: County-level soybean yield prediction using deep
CNN-LSTM model. Sensors 19, 4363 (2019). https://doi.org/10.3390/s19204363
9. Nevavuori, P., Narra, N., Lipping, T.: Estimation of rice yield with deep convolutional neural
networks. Comput. Electron. Agric. 163 (2019)
10. Chen, Y., et al.: Strawberry yield prediction based on a deep neural network using high-
resolution aerial orthoimages. Remote Sens. 11, 1584 (2019)
11. Russello, H.: Convolutional neural networks for estimation of rice yield using satellite images.
Master’s thesis, University of Amsterdam, Amsterdam, The Netherlands (2018)
12. Jiang, Z., Liu, C., Hendricks, N.P., Ganapathysubramanian, B., Hayes, D.J., Sarkar, S.: Pre-
dicting county level corn yields using deep long short term memory models. arXiv 2018,
arXiv:1805.12044 (2018)
13. Kulkarni, S., Mandal, S.N., Sharma, G.S., Mundada, M.R., Meeradevi: Predictive analysis
to improve crop yield using a neural network model. In: Proceedings of the 2018 Interna-
tional Conference on Advances in Computing, Communications and Informatics (ICACCI),
Bangalore, India, 19–22 September 2018, pp. 74–79 (2018)
14. You, J., Li, X., Low, M., Lobell, D., Ermon, S.: Deep Gaussian process for estimation of rice
yield based on remote sensing data. In: Proceedings of the thirty-First AAAI Conference on
Artificial Intelligence, San Francisco, CA, USA, 4–9 February 2017, pp. 4559–4566 (2017)
15. Alhnaity, B., Pearson, S., Leontidis, G., Kollias, S.: Using deep learning to predict plant
growth and yield in greenhouse environments. arXiv 2019 arXiv:1907.00624 (2019)
16. Sahoo, S., Das, M., Mishra, S., Suman, S.: A hybrid DTNB model for heart disorders pre-
diction. In: Mallick, P.K., Bhoi, A.K., Chae, G.S., Kalita, K. (eds.) Advances in Electronics,
Communication and Computing. ETAEERE 2020. Lecture Notes in Electrical Engineering,
vol. 709, pp. 155–163. Springer, Singapore (2021). https://doi.org/10.1007/978-981-15-8752-
8_16
17. Khaki, S., Pham, H., Wang, L.: YieldNet: a convolutional neural network for simultaneous
corn and soybean yield prediction based on remote sensing data. https://doi.org/10.1101/
2020.12.05.413203
388 N. Mishra et al.

18. Jeong, S., Ko, J., Yeom, J.-M.: Predicting rice yield at pixel scale through synthetic use of crop
and deep learning models with satellite data in South and North Korea, Science of The Total
Environment, vol. 802, pp. 149726 (2022). https://doi.org/10.1016/j.scitotenv.2021.149726,
ISSN 0048-9697
19. Jena, L., Kamila, N.K., Mishra, S.: Privacy preserving distributed data mining with evolution-
ary computing. In: Satapathy, S., Udgata, S., Biswal, B. (eds.) Proceedings of the International
Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA) 2013.
Advances in Intelligent Systems and Computing, vol. 247, pp. 259–267. Springer, Cham
(2013). https://doi.org/10.1007/978-3-319-02931-3_29
20. Mishra, S., Mallick, P.K., Tripathy, H.K., Jena, L., Chae, G.S.: Stacked KNN with hard
voting predictive approach to assist hiring process in IT organizations. Int. J. Electr. Eng.
Educ. 0020720921989015 (2021)
21. Dutta, A., Misra, C., Barik, R.K., Mishra, S.: Enhancing mist assisted cloud computing toward
secure and scalable architecture for smart healthcare. In: Hura, G., Singh, A., Siong Hoe,
L. (eds.) Advances in Communication and Computational Technology. Lecture Notes in
Electrical Engineering, vol. 668, pp. 1515–1526. Springer, Singapore (2021). https://doi.org/
10.1007/978-981-15-5341-7_116
22. Rath, M., Mishra, S.: Security approaches in machine learning for satellite communication.
In: Hassanien, A., Darwish, A., El-Askary, H. (eds.) Machine Learning and Data Mining
in Aerospace Technology. Studies in Computational Intelligence, vol. 836, pp. 189–204.
Springer, Cham (2020). https://doi.org/10.1007/978-3-030-20212-5_10
23. Chakraborty, S., Sahoo, K.S., Mishra, S., Islam, S.M.: AI driven cough voice-based COVID
detection framework using spectrographic imaging: an improved technology. In: 2022 IEEE
7th International conference for Convergence in Technology (I2CT), pp. 1–7. IEEE (2022)
24. Mishra, S., Thakkar, H.K., Singh, P., Sharma, G.:. A decisive metaheuristic attribute selec-
tor enabled combined unsupervised-supervised model for chronic disease risk assessment.
Comput. Intell. Neurosci. 2022, 1–17(2022)
25. Mohanty, A., Mishra, S.: A comprehensive study of explainable artificial intelligence in health-
care. In: Mishra, S., Tripathy, H.K., Mallick, P., Shaalan, K. (eds.) Augmented Intelligence in
Healthcare: A Pragmatic and Integrated Analysis. Studies in Computational Intelligence, vol.
1024, pp. 475–502. Springer, Singapore (2022). https://doi.org/10.1007/978-981-19-1076-
0_25
 Sentiment Analysis on Movie Review Data Using
Ensemble Machine Learning Approaches

Om Prakash Jena1(B) , Alok Ranjan Tripathy1 , Manas Prasad Rout2 ,

and Partha Sarathi Pattnayak3
1 Ravenshaw University, Cuttack, Odisha, India
[email protected],
[email protected]
2 FM University, Balasore, India
3 KIIT Deemed to be University, Bhubaneswar, Odisha, India

Abstract. Today’s machine learning application has a huge impact on the current
scenario that is big changes is afoot in the marketing world and shifts are largely
down to the power of machine learning. It is about finding pieces of predictive
knowledge, it has capable in the field of NLP to understand, analyze, manip-
ulate and potentially generate human language, Truly listening to a customer’s
voice requires deep understanding of what they have expressed in natural lan-
guage. NLP is the best way to understand this and uncover the sentiment behind
it. Here we have proposed a sentiment analysis for movie review, from the movie
review dataset. We have implemented the data set by using of Naive Bayes, Logis-
tic Regression, Random Forest algorithm, Support Vector Machine, Multinomial
Naïve Bayes, Stochastic Gradient descent classifier along with combining all those
as an ensemble with voting majority techniques with features parameters such as
positive, negative, neutral, partially positive, partially negative. We get the result
of the proposed ensemble technique is better than individual algorithms with 94%
of accuracy.

Keywords: Sentiment analysis · Naive Bayes · Logistic Regression · Random

Forest · Multinomial Naïve Bayes · Stochastic gradient descent classifier ·
Majority voting · Movie review · IMDB dataset

1 Introduction to Sentiment Analysis

NLP increase computational power to achieve accurate results in different area i.e. health-
care, media, finance and human. To understanding human language, speech or text for a
machine is a great challenge because machine can understand any things through binary
representation, the human generated data are large volume, this generated data may
contain unstructured format and the difficulties is the machine could generate accurate
feedback to considering the human generated data, so NLP can build a bridge in between
the human and computer and solve this problem. NLP used some technique i.e. syntactic
analysis and semantic analysis. Syntactic analysis refers to the arrangement of word in a

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 389–400, 2022.
https://doi.org/10.1007/978-3-031-23233-6_29
390 O. P. Jena et al.

sentence that make grammatical sense. It used for, how the natural language deals with
the grammatical value. There are some syntax techniques are used such that lemmatiza-
tion, word segmentation, Part-of-speech tagging, parsing, sentence breaking, stemming.
But semantic analysis used to know the meaning and interpretation of word and how
sentence are structured. There are also some technique i.e. Named entity recognition
(NER), word sense disambiguation, natural language generation.
Sentiment Analysis is also termed as emotion AI or pulling of opinion. Basically
it focuses on identify subjective information. It checks the polarity of text i.e. positive
or negative. By the help of these polarities, change the concept, improve productivity
and advertising so it helps reduce some negativity. Sentimental analysis is also referred
as opinion mining. It is component part of NLP (natural language process), that helps
to distinguish and pull opinions within a given block of speech/text. Basically the aim
of sentiment analysis is to determine the frame of mind, sentiments and response of
writer based on the topic. Sentimental analysis have so many advantage in various field
like business related area, researchers etc. It is also use for monitoring i.e. in a social
media; there are huge amount of data in form of short-message, memes and emoticons.
These are some challenges for the micro-blogging content that is coming from Twitter
and Facebook, it is because of the kind of language used for convey the sentiments i.e.
short forms, memes and emoticons. Sentimental analysis is helpful for the researchers
especially in the fields like Sociology, marketing, advertising, economics and political
science. Sentimental analysis is also used in companies. Companies have customer
feedback, so those companies manipulate his plans/structure according to the feedback
system, but it is still impossible to analyze it manually.
There are various steps to analyze sentiment data. Here we have performed sev-
eral phases as data Collection, Text Preparation, Sentiment Detection, and Sentiment
Classification.
The entire work is organized as follow: Sect. 1 describe the basic introduction to
sentiment analysis, Sect. 2 describe the related work, Sect. 3 describe the proposed
methodology with different ML methods along with ensemble method, Sect. 4 describe
performance measurement for the proposed models, Sect. 5 elaborate the result analysis
and Sect. 6 describe the conclusion and future work.

2 Recent Work

The author taking social media sentiment analysis using SVM and NB algorithms and
using Ant Colony and particle Swarm optimization method getting 73.62%, 77.30%
accuracy for NB and 76.71%, 80.54% for SVM respectively [1]. Considering the social
network Uri attack tweets the author use two methods i.e. sentiment score and polarity
count getting 94.3% accuracy of negative and 5.7% of positive result [2]. Taking Viet-
namese student feedback corpus with LSTM using support vector machine algorithm,
the author achieved the F1-score of 90.2% accuracy [3]. Using different machine learn-
ing algorithm such as NB, SVM, DT, RNN, taking movie reviews the author learn how
to face ML problem and how to do data analysis to make the work easier. He noticed
that applying transformation on the data can improve the performance of classification
method and he found the RNN gives the better result [4]. With the help of the Urdu
 Sentiment Analysis on Movie Review Data 391

sentiment corpus, taking Urdu tweets to analysis and polarity detection, the dataset
comprising over 17,185 token with obtain 52% of positive and 48% of as negative [5].
The author considered Particle Swarm optimization (POS) and Ant Colony optimiza-
tion (ACO) with SVM and NB classifier getting 86.29% of accuracy than the SVM-PSO
[6]. Taking Deep learning neural network (DNN) for customer sentimental analysis and
review classification achieving high level accuracy is 0.9248% with average F1-score of
0.925% [7].
The author proposed sentimental analysis for hotel rating using NB algorithm using
60% training data and 40% of testing data and getting 45 key values for positive and
45 key for negative. He noticed that for scaling the dataset Naïve Bayes is good and
implementing the linear equation on features and predicators [8]. Taking sentiment
analysis using SVM with Osgood values and Turnery values obtaining accuracy 68.3%
in Osgood but Turnery gives 65.8% of accuracy. But the hybrid SVM (Turnery and
Lemma) gives accuracy 86.0% and hybrid SVM (Osgood and Lemmas) gives accuracy
84.6% and 86.0% with 3-fold and 10-fold experiment [9]. Taking Amazon and IMDB
movie review, the author considering Naïve Bayes, J48, BF Tree and oneR classifier but
Naïve Bayes is quite faster in learning whereas oneR more promising, it gives 91.3% of
accuracy, 97% in F-measure [10]. The author considered 2 dataset HI-EN and BE-EN
and using SVM and voting classifier with neural network obtaining 0.569% f-score of
HI-EN and 0.526% F-score BN-EN [11]. The author tried to exceed domain-transfer
problem with some effective measure feature, using Adapted Naïve Bayes, taking 3
Chinese domain-specific dataset and use Chinese text POS tool ICTCLAS and getting
average of 600 word in education review, 460 terms in stock review and 120 word in
computer review [12]. Taking movie and product reviews dataset of sentiment analysis
in Turkish and English language using SVM classifier obtain 91.33% accuracy [13].
In this paper, the author takes car reviews as dataset contain pre-labeled sentence of
10,000, positive and negative of 5000, overly the author going through sentence level
analysis so employing conjunct analysis with sentence level it gives better accuracy. He
found that ML algorithm cannot be efficiently so using WordNet substantially enhance
the accuracy is about 80% [14]. Taking dataset as movie and product reviews hat are in
English and Turkish language using SVM algorithm and considering some parameter i.e.
kernel type, weighting schemes (i.e. TF-IDF, tokenization,feature selection), the author
getting accuracy is about 91.33% [15]. Processing micro blogs, it is very challenging
task, it have noise. In this paper the author taking method as combining social and topic
context to analyze micro blog sentiment. To analyze this, using Laplacian matrix of the
graph model [16]. A noble approach has been discussed to improve sentiment analysis
with patterns lexicons and negations with hybrid and Sentic net4 with a commendable
accuracy 86.32% [17].

3 Proposed Methodology

Here we consider a dataset as IMDB movie dataset and applied different machine learn-
ing classifier as Naïve Bayes, Random Forest, Logistic Regression and applying some
pre-processing steps as shown in Fig. 1.
392 O. P. Jena et al.

Fig. 1. Overall Diagram of Proposed Method

3.1 Naïve Bayes Algorithm

Naïve Bayes algorithm Bayes theorem for probabilistic prediction. It is used for find
out assumption of independence between predictors. This is used for very large dataset
because it has no complicated iterative parameter estimation.
Considering IMDB movie review as a dataset. Before applying algorithm, it should be
pre-processed. Converting for word vector. We separate reviews as positive and negative.
It stores like positive folder and negative folder (such as positive review in positive folder
and negative reviews in negative folder. In this folder, we noted that positive reviews as
“pos” tag and negative reviews as “neg” tag. After that, it finds out best word are taken
 Sentiment Analysis on Movie Review Data 393

by applying conditional probability. Apply Naïve Bayes classifier with testing data and
best word to obtain matrix.
Steps of the Algorithm

1. Initialize the prior probability of positive or P(positive).

2. Initialize the prior probability of negative or P(negative).
3. Conversion of sentence into word.
For each class of (Positive, negative):
For each word in (phrase)

P(W |C) < num(W |C)|num(Class) + numtotal (1)

P(C) = P(C) ∗ P(W |C)

Return max {P (+ve), P (−ve)}

P(S|G) ∗ P(G)
P(G|S) = (2)
P(S)

P(G|S) = P(S1 |G) ∗ P(S2 |G) ∗ P(S3 |G) . . . . . . · · · ∗ P(SN |G) ∗ P(G) (3)

P (G|S)-- > posterior probability of class (d, target) given predictor(y, attribute).
P(G) -- > class of prior probability.
P(S|G) -- > likelihood.
P(S) -- > predictor as prior probability.
P(S|G) ∗ P(G)
P(G|S) = (4)
P(S)
where S is feature and G is sentiment value.
P(S1 , S2 , S3 ) ∗ P(G)
P(G|S1 , S2 , S3 ) = , S1 , S2 , S3 are feature (5)
P(S1, S2, S3)
P(S1 |G) ∗ P(S2 |G) ∗ p(S3 |G) ∗ P(G)
P(d|S1 , S2 , S3 ) = (6)
P(S1 ) ∗ P(S2 ) ∗ P(S3 )
n P(S1 |G)
P(G|S) =
i=0 P(S)

Calculation of P(S) is done only once, so we calculate only

n
P(S1 |G)
i=0

1 n
P(Gk |S1 , . . . . . . Sn ) = P(Gk ) P(Si |Gk ) (7)
Z i=1
394 O. P. Jena et al.

Ŵ = argmaxgG P(G|S) (8)

P(S|G) ∗ P(G)
P(G|S) = (9)
P(S)

P(S|G)∗P(G)
Ŵ = argmaxdD P(G|S) = argmaxdD (10)
P(S)

Ŵ = argmaxdD P(S|G)∗P(G)

Ŵ = P(S1 , S2 , S3 |G) * P(S)

Ŵ = Gk for some K as follows:


n
ŵ = argmaxP (Gk ) P(Si |Gk ) (11)
i=1

To avoid underflow,

ŵ = argmax(ln P(Gk ) + n
i=1 ln P(Si |Gk )) (12)

If W1 and W2 are two independent variable,

P(W 1, W 2|S) = P(W 1|S) ∗ P(W 2|S).

3.2 Logistic Regression

It is a statistical learning technique in supervised classification algorithm. Here we used
sigmoid function.
1
(A) =
1 + e−A
mi,j --> feature vector of length ‘N’.
j = 1, . . . . . . . . . ., N , i = 1, . . . . . . . . . .., h.
⎡ ⎤
m11 , m12 , . . . .m1N
⎢ m21 , m22 , . . . .m2N ⎥
M =⎢
⎣
⎥
⎦ (13)
.
mh1 , mh2 , . . . ..mhN h∗N
⎡ ⎤
r1
⎢r ⎥
⎢ 2⎥
⎢ ⎥
R = ⎢ . ⎥ P(R|M ) = 1
1+e −−f(m)
(posterior)f(m) is the feature. mj
⎢ ⎥
⎣ . ⎦
rh
P(R|M)
log[ ] = m0 + m1 β1 + . . . . . . . . . . . . + mN βN +  = f (m)
1 − P(R|M)
 Sentiment Analysis on Movie Review Data 395

Maximum likelihood estimation,

h
argmax : log{ P(Ri |Mi )Ri (1 − P(Ri |Mi )(i−Ri ) } (14)
i=1

3.3 Random Forest Classifier

It is based upon ensemble tree based learning algorithm. It decides the final class of the
test object, considering the vote from different decision tree. Here in this classifier we
have adopts boats tarp bagging aggregation.
Training set W = w1 , w2 , . . . . . . .., wn .
With responses S = s1 , s2 , . . . . . . . . . , sn .

Bagging repeatedly “E” times.

For e = 1, 2, . . . . . . .E.

Prediction for unseen samples y’, by averaging the prediction from all individual
regression tree on y’:

1 E
f̂ = fe (y ) (15)
E e=1

To find out uncertainty of the prediction,

3.4 Multinomial Naïve Bayes Classifier

It is mostly used for discrete counts. Here we have taken the feature vectors
(g1 , g2 , ..., gk ) to solve the problem of text classification task with the integer value
of word frequency by using TF-IDF method.
The conditional probability distribution Pr ob(doc/classm ) can be given by

Multinomial Pr ob(doc/classm ) = Pr ob((g1 , g2 , ..., gk )/classm )


= Pr ob(gn /classm ) (17)
1≤n≤k

After applying the Bayes rule we have reduced the equation to.

CMAP = arg max Pr ob(cm ) Pr ob(gn /classm ) (18)
cm ∈C 1≤n≤k
396 O. P. Jena et al.

With the feature available in the documents bag of word, we have calculated
Pr ob(gn /classm ). The probability of wn in classm is calculated in training dataset as

count number(wn ∗ classm )

Pr ob(wn /classm ) = (19)
w∈Voc count number(wn , classm )

3.5 Support Vector Machine

It’s a supervised machine learning algorithm that aims to locate a hyperplane in a multi-
dimensional space. The SVM’s goal is to find the best hyperplane so that the two groups
can be segregated. This ideal hyperplane divides the two groups while also increasing
the margin between them. The distance between the hyperplane and the SVs is known as
the margin. It is widely used because of its primary benefit, which is that it can be very
efficient even in high-dimensional spaces. However, the key flaw in this method is that
it does not have probabilistic estimations. High precision can be achieved by fine-tuning
hyper parameters such as gamma, coat, and kernel level, but in practice, defining the
exact hyper parameters can be difficult which directly enhances the computational cost
and overhead.
The hyperplane optimization can be obtained by Equation.
Minimize 21 w(v) 2 , where w(v) 2 = w(v) T x

Subject to yj (w(v) x j + b) ≥ 1 where j = 0,1

T

1 −yj (w(v)
T x + b) ≤ 0 where j = 1,2, …
j
The Lagrangian optimization problem is defined in Equation.

1 2

n
τ= w(v) + αj (1 − yj (w(v)
T
xj + b)) (20)
2
j=1

The first derivative of the above equation is taken with respect to w(v) and b, then we
have the following Equation.
w(v) + nj=1 α j ( − yj) xj = 0.
w(v) = nj=1 α j yj xj.
n
j=1 α j yj = 0 where α j ≥ 0.
For the non-linearly classification problem, we introduce another dimension to cre-
ate a bigger dimensional space. Here ξj is introduced to represent the approximated
misclassified data samples. So, the classification model is represented as in Equation
1 T n
f (x) = w(v) w(v) + c ξj (21)
2 j=1

subjecttoyj (w(v)
T x + b) ≥ 1 Where j = 1,2,… And ξ ≥ 1
j j
Where c represents the trade-off between the margin and training error and c is a
constant. We have to minimize the f (x) for a better-optimized hyperplane.
 Sentiment Analysis on Movie Review Data 397

3.6 Stochastic Gradient Descent Classifier

It is mostly used for optimizing problem in machine learning which find best parameter
that corresponds to best fitted predicted result.
Let input is xi and corresponding output is yi then

H =− (yi log(p(xi )) + (1 − yi ) log(1 − p(xi )) (22)
i

Here we have used both Adam and RMSProp optimizer for our simulation based on
Batch stochastic gradient descent classifier.

4 Performance Measurement Parameter

Here we have considered some measurement parameter for our model to find out the
solution of our model. The parameters are Accuracy, Precision, Recall and F-measure.

Correctpredictionsnumber
Accuracy =
Totalnumberofprediction
T (+ve) + T (−ve)
Accuracy =
T(+ve) + T(−ve) + F(+ve) + F(−ve)
T (+ve)
Precision =
T(+ve) + F(−ve)
(β2 + 1)RT
Fβ = (RandTareprecisionandrecall)
β2 R + T
where β > 1, then it is favorable for recall.
β < 1, then it is for precisions.
β = 1, here presicion and recall are equal.
When Fβ=1 orF 1 F1 = R+T 2RT
.

5 Result Analysis
Considering on a movie review dataset we have taken 151871 amount of data out of
which we segregate 101870 amount of training and 50001 amount of testing data. Sim-
ulating on python with Intel®Core™ i5-5005u cpu @ 2.00 GHz,64-bit, 8 GB RAM and
we taking performance measurement parameter(precision, recall, accuracy) and imple-
menting three classifier(such as: Naïve Bayes, Logistic Regression, Random Forest)
and getting the following table with individual accuracy 84% of Naïve Bayes, 86% of
Logistic Regression, 88% of Random Forest, 85%.of Multinomial Naïve Bayes, 90%
of Support Vector Machine, 0.91 of Stochastic Gradient Classifier and 0.94 with our
proposed ensemble technique. The output of some sample positive review words and
negative review words are given in Figs. 2 and 3 respectively (Table 1).
398 O. P. Jena et al.

Table 1. Result analysis of different ML algorithms.

Fig. 2. Some positive review words

 Sentiment Analysis on Movie Review Data 399

Fig. 3. Some negative review words

6 Conclusion and Future Work

We have performed movie review sentiment analysis by using Naïve Bayes, Logistic
Regression, Random Forest, method and from which it will conclude that, due to the
ensemble mechanism used in Random Forest, Multinomial Naïve Bayes, Support Vector
Machine, Stochastic Gradient Classifier and ensemble technique algorithm it result 94%
of accuracy which more optimistic than the other methods. Based on this result our pro-
posed model worked efficiently for sentiment prediction of the movie reviewer to a great
intent. This can be applied to collect the relevant information to model a recommender
system which analyzes the marketing, advertising and promoting the movie.
In the future work, we have to focus on some ensemble method and optimization
technique for other parameter, parameter reduction, so that our model will work on both
robust recognition system and recommender system.

References
1. Badr, E.-S., Salam, A., Mustafa & Ali, Mahmoud & Ahmed, Hagar.: Social media sentiment
analysis using machine learning and optimization techniques. Int. J. Comput. Appl. 178,
975–8887 (2019). https://doi.org/10.5120/ijca2019919306
2. Kawade, D.: Sentiment analysis: machine learning approach. Int. J. Eng. Technol. 09, 2183–
2186 (2017). https://doi.org/10.21817/ijet/2017/v9i3/170903151
3. Vu, N., Kiet, N., Ngan, N.: Variants of long short-term memory for sentiment analysis on
Vietnamese students’ feedback corpus, pp. 306–311 (2018). https://doi.org/10.1109/KSE.
2018.8573351
4. Liombart, O.R.: Using machine learning techniques for sentiment analysis. University
Autonoma De Barcelona (UAB) (2017)
5. Khan, M.Y., Nizami, M.S.: Urdu Sentiment Corpus (v1.0): Linguistic Exploration and Visu-
alization of Labeled Dataset for Urdu Sentiment Analysis (2020). https://doi.org/10.1109/ICI
SCT49550.2020.9080043
6. Bansal, P., Kaur, R.: Twitter sentiment analysis using machine learning and optimization
techniques. Int. J. Comput. Appl. 179(19), 5–8 (2018). https://doi.org/10.5120/ijca20189
16321
7. Shreyas, R.L.: Sentiment analysis of customer satisfaction using Deep learning. IRJCS 6(12)
(2019). https://doi.org/10.26562/IRJCS.2019.DCCS10083
400 O. P. Jena et al.

8. Bhargav, P.S., Reddy, G.N., Chand, R.R., Pujitha, K., Mathur, A.: Sentiment analysis for hotel
rating using machine learning algorithm. IJITEE 8(6) (2019)
9. Mullen, T., Collier, N.: Sentiment analysis using support vector machines with diverse infor-
mation sources. In: 2004 Conference on Empirical Methods in Natural Language Processing
, EMNLP (2004)
10. Singh, J., Singh, G., Singh, R.: Optimization of sentiment analysis using machine learning
classifiers. HCIS 7(1), 1–12 (2017). https://doi.org/10.1186/s13673-017-0116-3
11. Mishra, P., et al.: Code-mixed sentiment analysis using machine learning and neural network
approaches. arXiv abs/1808.03299 (2018)
12. Tan, S., Cheng, X., Wang, Y., Xu, H.: Adapting Naive Bayes to domain adaptation for senti-
ment analysis. In: Boughanem, M., Berrut, C., Mothe, J., Soule-Dupuy, C. (eds.) ECIR 2009.
LNCS, vol. 5478, pp. 337–349. Springer, Heidelberg (2009). https://doi.org/10.1007/978-3-
642-00958-7_31
13. Gözükara, F., Özel, S.: An experimental investigation of document vector computation meth-
ods for sentiment analysis of Turkish and English reviews 31, 467–481 (2016). https://doi.
org/10.21605/cukurovaummfd.310341
14. Meena, A., Prabhakar, T.V.: Sentence level sentiment analysis in the presence of conjuncts
using linguistic analysis. In: Amati, G., Carpineto, C., Romano, G. (eds.) ECIR 2007. LNCS,
vol. 4425, pp. 573–580. Springer, Heidelberg (2007). https://doi.org/10.1007/978-3-540-
71496-5_53
15. Zou, X., Yang, J., Zhang, J.: Microblog sentiment analysis using social and topic context.
PLoS ONE 13(2), e0191163 (2018). https://doi.org/10.1371/journal.pone.0191163
16. El Alaoui, I., Gahi, Y., Messoussi, R., Chaabi, Y., Todoskoff, A., Kobi, A.: A novel adaptable
approach for sentiment analysis on big social data. J. Big Data 5(1), 1–18 (2018). https://doi.
org/10.1186/s40537-018-0120-0
17. Pradhan, A., Senapati, M.R., Sahu, P.K.: Improving sentiment analysis with learning concepts
from concept, patterns lexicons and negations. Ain Shams Eng. J. 13(2) (2022). https://doi.
org/10.1016/j.asej.2021.08.004, ISSN 2090–4479
 Comparative Analysis of COVID 19 Detection
from Cough Speech Using Machine Learning
Classifiers

Soumya Mishra(B) , Tusar Kanti Dash, Ganapati Panda, Amit Kumar,

and Sushant Kumar Singh

Department of Electronics and Tele Communication, C. V. Raman Global University,

Bhubaneswar 752054, Odisha, India
[email protected]

Abstract. Corona Virus Disease-2019, or COVID-19, has been on the rise since
its emergence, so its early detection is necessary to stop it from spreading rapidly.
Speech detection is one of the best ways to detect it at an early stage as it exhibits
variations in the nasopharyngeal cavity and can be performed ubiquitously. In
this research, three standard databases are used for detection of COVID-19 from
speech signal. The feature set includes the baseline perceptual features such as
spectral centroid, spectral crest, spectral decrease, spectral entropy, spectral flat-
ness, spectral flux, spectral kurtosis, spectral roll off point, spectral skewness,
spectral slope, spectral spread, harmonic to noise ratio, and pitch. 05 ML based
classification techniques have been employed using these features. It has been
observed that Generalized Additive Model (GAM) classifier offers an average of
95% and a maximum of 97.55% accuracy for COVID-19 detection from cough
signals.

Keywords: COVID-19 · Speech feature · Machine learning classifier

1 Introduction
According to the statistics given by the World Health Organization (WHO) on October
14th , 2022, there have been a total of 620,301,709 confirmed COVID-19 cases [1].
COVID-19 symptoms include body discomfort, high body temperature, strong coughing,
and significant breathing difficulties [3]. The ludicrous and rising spread of the COVID-
19 virus has resulted in unprecedented coordination among several fields in order to
limit the infection’s spread and prevent collateral damage on a regular basis [4]. Here
in this work, COVID-19 has been attempted to be diagnosed using human cough by
adopting datasets from multiple databases of healthy and positive individuals. Moreover,
an advanced and combative way to early speech detection is by implementing machine
learning classifiers to classify COVID-19 [2, 5]. K-Nearest Neighbour (KNN) is a proven
classifier [4] to detect COVID-19 because of its excellent efficiency and capacity [5] to
address a variety of challenging pattern categorization tasks. Moreover, Decision Tree
[6, 7], Naïve Bayes [8] and Support Vector Machine (SVM) [9] have previously shown

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 401–408, 2022.
https://doi.org/10.1007/978-3-031-23233-6_30
402 S. Mishra et al.

enhanced classification results in speech problems. GAM classifier is compared against

KNN, SVM, Binary tree and Naïve Bayes to achieve the best outcomes. In our work, we
are mainly evaluating three major corpuses for extracting baseline perceptual spectral
features for COVID-19 detection from the coughing sounds of different individuals. The
paper proceeds with the methodology, results and discussions followed by concluding
the work.

1.1 Related Works

[13] showcases exploiting temporal and spectral features with a VGGNet classifier,
delivering 0.82 AUC for Cough based COVID-19 classification. A highest accuracy of
89.79% using Random Forest (RF) ML-classifier based on spectral features is delivered
in [30]. [11] portrays exploiting Cepstral and spectral features through an RF Classifier,
delivering 69% test data classification accuracy. Pahar.et.al. [31] have in detail experi-
mented the cough detection using deep learning structures and achieved a maximum of
95% using ResNet50 architecture.

2 Methodology
Speech segments were obtained for baseline feature extraction, using a 25-millisecond
window length [10] post normalizing the signal amplitudes as a part of pre-processing
technique. Here, the audio files of the standard databases are subjected to perceptual audio
feature extraction consisting of 13 feature vectors. GAM, KNN, SVM, Binary Tree and
Naïve Bayes are the potential classifiers used to classify the positive and healthy cough
sounds based on the extracted features in this work. The classifier hyperparameters are
tuned using Bayesian Optimizer Fig. 1.

COVID-19
Speech Positive
input Classification
Pre-processing Speech Feature
Extraction using ML
Classifiers

Healthy

Fig. 1. System model

2.1 Databases

Three standard available labelled corpuses were adopted in the work for conducting the
experiment. The databases are as follows:
 Comparative Analysis of COVID 19 Detection 403

2.1.1 Coswara (DB-1)

The Coswara-Cough Audio samples were originally obtained using an online tool that
allows for crowdsourcing, from all over the world. The audio samples were recorded at
48000 samples per second. The dataset comprised of 6507 clean audio files, 1117 noisy
files, and the remaining files are of poorer quality, all belonging to 941 test subjects [11].
Here the individual cough categories are combined together and considered for further
feature extraction.

2.1.2 Coughvid (DB-2)

The Coughvid collection contains over 25,000 cough recordings from people of all
ages, genders, and geographical locations. It includes over 2,800 samples used to detect
medical issues in coughs, making it one of the most comprehensive cough datasets
with expert annotations accessible for a variety of cough audio sample classification
applications. Before downsampling, these audio samples were subjected to a lowpass
filter with a cutoff frequency of 6 KHz, which was downsampled to 12 KHz [12].

2.1.3 KDD Cambridge (DB-3)

This Cambridge database, comprised of cough and breathing samples, as well as com-
plete information about the volunteers’ age, gender, demography, medical histories and
symptoms collected via an app. The audio samples were captured at 44.1 kHz sampling
rate. The audio samples were taken from 582 healthy individuals and 141 individuals
who had COVID-19 positive tests. Cough symptoms were reported by 264 healthy peo-
ple and 54 COVID-19 patients, whereas no symptoms were reported by 318 healthy
people and 87 COVID-19 patients [13].

2.2 Feature Extraction

The baseline perceptual audio feature vectors in this work comprise of:

a. Spectral Centroid: The spectral centroid represents the centre of the spectral power
distribution of a signal and also highlights the intensity of an audio signal [14].
b. Spectral Crest: This feature denotes the maximum power spectrum of a sound signal.
Spectral crest offers easy differentiation of harmonics from noise-like sounds [15].
c. Spectral Entropy: Signal information and spectral distribution spikiness are mea-
sured by spectral entropy [16].
d. Spectral Roll-off: Speech transmissions often have less energy at high frequencies.
This characteristic can be seen in spectral roll-off, which characterises an energy
and frequency connection. Previous research has focused on the spectral roll-off that
contains the majority of the energy or on a narrow range of spectral roll-off values
[14].
e. Spectral Kurtosis: Spectral kurtosis highlights the transients with their indices in
frequency domain. It is a statistical instrument that may detect the presence of a
series of transients and their positions in the frequency domain [18].
404 S. Mishra et al.

f. Spectral Flux: The power spectrum’s cycle-to-cycle variability is measured by

spectral flux. The spectral flux was calculated as the Euclidian distance between
consecutive power spectrums [14].
g. Spectral Skewness: A spectrum symmetricity around its arithmetic mean is com-
puted using spectral skewness. Skewness equal to zero denotes symmetric distribu-
tion, a value less than zero denotes more energy compaction on the right hand side
of the spectrum, and a value larger than zero denotes more energy compaction on
the left hand side of the spectrum [15].
h. Spectral spread: Spectral Dispersion is another name for it. The bandwidth of the
transmission has a direct impact on this aspect. The rate-average map’s deviation
from the centroid can be used to define it. Pure tonal sounds have a tiny spectral
spread compared to the vast spectral spread of noise-like signals [15].
i. Spectral slope: It is determined via linear regression and represents the slope of the
signal’s amplitude which is vitally used in speech detection purpose [15].
j. Spectral decrease: The spectral decrease characterises the rate-map representation’s
average spectral slope, with a greater emphasis on low frequencies [15].
k. Harmonic to noise ratio: Harmonic to noise ratio indicates the noise amount in an
audio speech signal [22].
l. Spectral flatness: Spectral flatness measures the similarity of the speech spectrum
with a noisy spectrum. Acoustic signals include a property called spectral flatness
that has proven advantageous in a variety of audio signal processing scenarios. The
geometric mean of the magnitude spectrum of the signal, as determined by the
discrete Fourier transform, divided by the arithmetic mean, is the standard definition
of spectral flatness [17].
j. Pitch: Pitch represents the bandwidth or the correlation between any two sounds and
measure of voice quality respectively. The delayed duration between two subsequent
pitch intervals is governed by vocal fold tension and the increase of air pressure in
the lungs. It may vary, resulting in pitch variations of 2–10% across two consecutive
intervals [23].

2.3 Machine Learning Classifiers

2.3.1 SVM Classifier
[25] flaunts a high classification accuracy obtained by SVM in speech classification
problem. It is a well-known binary classifier employing a decision boundary between
two classes to create a hyperplane [21].

2.3.2 KNN Classifier

KNN classifier uses the Euclidian distance to determine the distance between the stored
records and unidentified records to classify whether the person has COVID-19 or not
[19, 20].

2.3.3 Naïve Bayes

The Naive Bayes
  method uses sample data information to estimate the posterior
probability p yx of each class y given an item x [24].
 Comparative Analysis of COVID 19 Detection 405

2.3.4 Binary Tree

To construct a binary decision tree, class labels are divided into two groups at each branch
of the tree until the binary tree reaches a single class label. Distance measurements
between the retrieved characteristics of the class labels are utilised to create a binary
decision tree. Principal component analysis is used to minimise the high dimension of
the feature set in order to measure the distance between disease categories [26]. To create
a speaker model with the goal of achieving a high average prediction power, as measured
by the average prediction entropy of the binary tree leaf distributions Low entropy, such
as anticipating unique symbols, correlates to the desirable prediction quality in this case,
and vice versa.

2.3.5 GAM Classifier

The Generalised Additive Model (GAM) was created in 1990 by Hastie and Tibshirani
as a statistical model. It can be used in a variety of prediction situations [29]. GAM
exploits regression properties for a thorough analysis of the data using a ‘Smoothing
function’. This ‘smoother’ is competent to estimate the trend from the data plot which
can subsequently establish the dependency of the mean of the response variable to the
predictors.

2.4 Experimental Setup

All simulations were performed on MATLAB 2021a software, with Windows 10 operat-
ing system, 8 GB RAM. GAM, SVM, KNN, Binary tree and Naïve Bayes were utilised
as potential classifiers to detect COVID-19 infection. Cross validation was performed
using a k-fold with k = 10 [20]. For performance evaluation of the classifier, factors
such as recall, sensitivity, F-score, precision, and accuracy were evaluated. The audio
was digitised and pre-processed using suitable functions, following which the speech
baseline features were extracted and reposited in MS Excel sheets. The classifiers were
then applied to these extracted feature vectors to obtain the necessary outcome of the
detection. The classification accuracy is evaluated as follows: [2]. The hyperparameters
of all the classifiers are selected using Bayesian Optimization technique [28].

3 Result and Discussion

Cough speech classification was performed on three databases, Coswara, Coughvid, and
the KDD Cambridge consisting of different categories of cough samples collected from
all over the globe.
Extracted baseline perceptual audio features were fed into GAM, SVM, KNN, Naïve
Bayes and Binary Tree. For performing the simulations, we have used MATLAB R2021a
in windows 10 operating system with intel i5 processor, 8GB RAM ,4GB SSD and 4 GB
graphics. The classification outcomes were interpreted in terms of confusion matrix and
classifier parameters expressed in Accuracy, Precision, Specificity, F-Score and Area
Under the Curve (AUC) Table 1.
406 S. Mishra et al.

Table 1. Classification accuracies and outcomes for DB-1, DB-2, DB-3.

COVID_19 & Healthy classification

Databases Classifiers Accuracy Precision Specificity F-Score AUC
Coswara GAM 0.9755 0.9677 0.9751 0.9754 0.97
[DB-1] KNN 0.8244 0.8462 0.7857 0.8148 0.82
SVM 0.8062 0.8125 0.8125 0.8125 0.8
Naïve Bayes 0.8589 0.9255 0.9060 0.8172 0.85
Decision Tree 0.9395 0.8423 0.7875 0.8894 0.93
Coughvid GAM 0.9722 0.9714 0.9833 0.9645 0.96
[DB-2] KNN 0.898 0.9375 0.9167 0.9091 0.9118
SVM 0.8865 0.8235 0.8865 0.8235 0.8833
Naïve Bayes 0.8852 0.9231 0.8889 0.9057 0.9139
Decision Tree 0.8357 0.95 0.7037 0.8085 0.8757
KDD GAM 0.9035 0.7891 0.8055 0.7470 0.89
Cambridge KNN 0.8357 0.95 0.7037 0.8085 0.8757
[DB-3]
SVM 0.7777 0.7945 0.5690 0.7893 0.72
Naïve Bayes 0.8707 0.8296 0.7215 0.8918 0.85
Decision Tree 0.8741 0.9882 0.9886 0.9223 0.857

From the above outcomes, it is observed that GAM classifier, delivers best results
i.e. 97.55% classification accuracy, 96.77% precision, 97.51% specificity, 97%AUC, and
97.54% F-Score for DB-1; 97.22% accuracy, 97.14% precision, 98.33% specificity, 96%
AUC, and 96.45% F-Score for DB-2; and 90.35% accuracy, 78.9% precision, 80.55%
specificity, 89% AUC, and 74.7% F-Score for DB-3.

4 Conclusion
COVID-19 detection from speech signals can be a valuable and cost-effective method
because it does not require any complex medical invasive tests or exposure to clinics. This
work highlights a ubiquitous non-invasive detection strategy which can quickly diagnose
a suspected patient’s preliminary state of being COVID-19 positive or healthy without
the need to visit a hospital or seek medical assistance. In this paper, simple baseline
perceptual audio features have been used for detection of COVID-19 cough using 03
databases and 05 Machine learning classifiers. GAM outperformed the rest in obtaining
best results from all the databases followed by Binary Tree based classifier delivering the
second best detection performance in the first database. However, this disease detection
approach is yet to be validated by hospitals and physicians. In future, attempt can be
made to develop a similar ML classifier model based on perceptual baseline features for
real time detection of speech related diseases such as Parkinson’s disease, Alzheimer’s
disease, Dysarthria, dementia, oral cancer etc.
 Comparative Analysis of COVID 19 Detection 407

References
1. https://covid19.who.int/
2. Shaban, W.M., Rabie, A.H, Saleh, A.I., Abo-Elsoud, M.A.: A new COVID-19 patients detec-
tion strategy (CPDS) based on hybrid feature selection and enhanced KNN classifier. Knowl.
Based Syst. 205, 106270 (2020)
3. Chowdhury, M.E., et al.: QUCoughScope: an artificially intelligent mobile application to
detect asymptomatic COVID-19 patients using cough and breathing sounds. arXiv preprint
arXiv:2103.12063. (2021)
4. Rasjid, Z.E., Setiawan, R.: Performance comparison and optimization of text document clas-
sification using k-NN and naïve bayes classification techniques. Procedia Comp. Sci. 1(116),
107–112 (2017)
5. Hassan, A., Shahin, I., Alsabek, M.B.: Covid-19 detection system using recurrent neural
networks. In: 2020 International Conference on Communications, Computing, Cybersecurity,
and Informatics (CCCI), pp. 1–5. IEEE (2020)
6. Aich, S., Younga, K., Hui, K.L., Al-Absi, A.A., Sain, M.: A nonlinear decision tree-based clas-
sification approach to predict the Parkinson’s disease using different feature sets of voice data.
In: 2018 20th International Conference on Advanced Communication Technology (ICACT),
pp. 638–642. IEEE (2018)
7. Petti, U., Baker, S., Korhonen, A.: A systematic literature review of automatic Alzheimer’s
disease detection from speech and language. J. Am. Med. Inform. Assoc. 27(11), 1784–1797
(2020)
8. Meghraoui, D., Boudraa, B., Merazi-Meksen, T., Boudraa, M.: Parkinson’s disease recogni-
tion by speech acoustic parameters classification. In: Chikhi, S., Amine, A., Chaoui, A., Khol-
ladi, M.K., Saidouni, D.E. (eds.) Modelling and Implementation of Complex Systems. LNNS,
vol. 1, pp. 165–173. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-33410-3_12
9. Jain, M., Narayan, S., Balaji, P., Bhowmick, A., Muthu, R.K.: Speech emotion recognition
using support vector machine. arXiv preprint arXiv:2002.07590. (2020)
10. Asgari, M., Shafran, I.: Predicting severity of Parkinson’s disease from speech. In: 2010
Annual International Conference of the IEEE Engineering in Medicine and Biology, pp. 5201–
5204. IEEE (2010)
11. Sharma, N., Krishnan, P., Kumar, R., Ramoji, S., Chetupalli, S.R., Ghosh, P.K., Ganapathy, S.:
Coswara--a database of breathing, cough, and voice sounds for COVID-19 diagnosis. arXiv
preprint arXiv:2005.10548. (2020)
12. Orlandic, L., Teijeiro, T., Atienza, D.: The COUGHVID crowdsourcing dataset, a corpus for
the study of large-scale cough analysis algorithms. Scientific Data. 8(1), 1 (2021)
13. Brown, C., et al.: Exploring automatic diagnosis of COVID-19 from crowdsourced respiratory
sound data. arXiv preprint arXiv:2006.05919. (2020)
14. Stolar, M.N., Lech, M., Stolar, S.J., Allen, N.B.: Detection of adolescent depression from
speech using optimised spectral roll-off parameters. Biomed. J. 2, 10 (2018)
15. Sharma, G., Umapathy, K., Krishnan, S.: Trends in audio signal feature extraction methods.
Appl. Acoust. 15(158), 107020 (2020)
16. Toh, A.M., Togneri, R., Nordholm, S.: Spectral entropy as speech features for speech
recognition. Proc. PEECS. 1, 92 (2005)
17. Madhu, N.: Note on measures for spectral flatness. Electron. Lett. 45(23), 1195–1196 (2009)
18. Antoni, J.: The spectral kurtosis: a useful tool for characterising non-stationary signals. Mech.
Syst. Signal Process. 20(2), 282–307 (2006)
19. Hossain, E., Hossain, M.F., Rahaman, M.A.: A color and texture-based approach for the
detection and classification of plant leaf disease using KNN classifier. In: 2019 International
Conference on Electrical, Computer and Communication Engineering (ECCE), pp. 1–6. IEEE
(2019)
408 S. Mishra et al.

20. Fushiki, T.: Estimation of prediction error by using K-fold cross-validation. Stat. Comput.
21(2), 137–146 (2011)
21. Ismael, A.M., Şengür, A.: Deep learning approaches for COVID-19 detection based on chest
X-ray images. Expert Syst. Appl. 1(164), 114054 (2021)
22. Krom, G.D.: A cepstrum-based technique for determining a harmonics-to-noise ratio in speech
signals. J. Speech Lang. Hear. Res. 36(2), 254–266 (1993)
23. Medan, Y., Yair, E., Chazan, D.: Super resolution pitch determination of speech signals. IEEE
Trans. Signal Process. 39(1), 40–48 (1991)
24. Webb, G.I., Keogh, E., Miikkulainen, R.: Naïve Bayes. Encycl. Mach. Learn. 15, 713–714
(2010)
25. Jain, M., Narayan, S., Balaji, P., Bhowmick, A., Muthu, R.K.: Speech emotion recognition
using support vector 23
26. Yüncü, E., Hacihabiboglu, H., Bozsahin, C.: Automatic speech emotion recognition using
auditory models with binary decision tree and SVM. In: 2014 22nd International Conference
on Pattern Recognition, pp. 773–778. IEEE (2014)
27. Brigham, K., Kumar, B.V.: Imagined speech classification with EEG signals for silent com-
munication: a preliminary investigation into synthetic telepathy. In: 2010 4th International
Conference on Bioinformatics and Biomedical Engineering, pp. 1–4. IEEE (2010)
28. He, F., Zhou, J., Feng, Z.K., Liu, G., Yang, Y.: A hybrid short-term load forecasting model
based on variational mode decomposition and long short-term memory networks considering
relevant factors with Bayesian optimization algorithm. Appl. Energy 1(237), 103–116 (2019)
29. Leathwick, J.R., Elith, J., Hastie, T.: Comparative performance of generalized additive models
and multivariate adaptive regression splines for statistical modelling of species distributions.
Ecol. Modell. 199(2), 188–196 (2006)
30. Tena, A., Clarià, F., Solsona, F.: Automated detection of COVID-19 cough. Biomed. Signal
Process. Control 1(71), 103175 (2022)
31. Pahar, M., Klopper, M., Warren, R., Niesler, T.: COVID-19 cough classification using machine
learning and global smartphone recordings. Comput. Biol. Med. 1(135), 104572 (2021)
 Classification of High-Resolution Satellite Image
with Content Based Image Retrieval and Local
Binary Pattern

Rajalaxmi Padhy, Laxmipriya Samantaray, Sanjit Kumar Dash(B) , and Jibitesh Mishra

Odisha University of Technology & Research, Bhubaneswar, Odisha, India

[email protected]

Abstract. The framework of content-based image retrieval (CBIR) is based on

the visual interpretation of contents that are present in the query image, which
can deal with problems like image texture classification and image texture visu-
alization. This study uses texture features to create a CBIR system for satellite
image datebase exclusively for high-resolution satellite pictures. The discussed
technique makes use of a block-based scheme and the local binary pattern tex-
ture feature. In order to extract LBP histograms, the query and database pictures
are separated into equal-sized blocks. The Chi-square distance is then applied to
compare the block histograms. According to experimental data, the local binary
pattern (LBP) formulation is a potent tool for retrieving high-resolution satellite
images.

Keywords: Local binary pattern · Texture feature extraction · Content-based

image retrieval · High resolution satellite image

1 Introduction
Image-processing applications potentially benefit from using image retrieval techniques.
A CBIR system is based on comparing the query over full images. Color, texture, and
shape are common methods of retrieving images. These methods are used to retrieve
an image from an image dataset. The multiple image resolutions and sizes and spatial
color distribution are irrelevant to them. Therefore, none of these methods are suitable
for retrieving art images. Furthermore, shape-based retrievals are only helpful in a small
number of scenarios. With the development of imaging technology in recent years,
satellites with extremely high-resolution spatial imaging systems, such as IKONOS,
GeoEye-1, World-View-1, and Quick Bird, have been launched and are now able to
make more precise observations of the world. The issue of managing image databases
has arisen as high-resolution photos have become more widely available. A CBIR system
returns images from an image database based on a query image. Images are provided by
the content and metadata-based systems, utilizing an efficient image retrieval method.
Global features including color, shape, and texture are employed by various new picture
retrieval algorithms. However, the earlier outcomes indicate that using those global
attributes to look for related photos produces too many false positives.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022

M. Panda et al. (Eds.): ICIICC 2022, CCIS 1737, pp. 409–416, 2022.
https://doi.org/10.1007/978-3-031-23233-6_31
410 R. Padhy et al.

In this paper, we present a novel interpretation of image retrieval employing both

content and metadata. The user-provided query image is used to extract related pictures
from the picture library. Using “content-based” research implies that rather than exam-
ining the image’s metadata, such as its categories and keywords, it will examine the
image’s actual contents. The features of query image along with the images stored in
database are extracted and appropriately indexed in a typical CBIR. The system lists all
those in the database. The database is searched for relevant pictures using the closest
similarity distance. It calculates the distance between the features of the requested image
and all those in the database. The database is searched for relevant pictures using the
closest similarity distance. We use the local binary pattern (LBP) operator to evaluate
texture performance for extracting high resolution satellite images.
The remainder section of the paper is structured as follow. Second segment describes
the background and related work. The third segment gives the description of the system
model and methodology. The fourth segment shows the results and discussions of the
respective data set, and the fifth segment gives a conclusion to our paper.

2 Background and Related Work

The method of visual perception for humans heavily depends on color information. In
addition, color is independent of the size and position of objects in an image. Because
of these factors, the color characteristic is frequently used in image retrieval to describe
a picture’s contents. Creating a color space histogram is the easiest way to get color
information. Color, texture, and intensity signals have been shown to be effective for
identification of low-resolution satellite images [1, 2]. In order to identify the type of
clouds in weather images, feature extraction has been exploited extensively [3, 4]. In
collaboration with color, shape, and texture features, effective features produced by the
combining of channels such as the vegetation indexes (NDVI), brightness indexes (BI), or
urban indexes (ISU) are employed with multi-spectral and hyper-spectral low resolution
satellite pictures (Noaa, Landsat). [5] Textural and structural features have long been
known to be highly discriminative in mid-resolution images [6, 7]. High resolution
satellite picture indexing and retrieval, however, is still a relatively unexplored field of
study. Recent research [8, 9], in which the texture descriptor is typically paired with other
features, has demonstrated the texture descriptor’s capability in dividing and extracting.
Images’ color information is displayed using the General Color Histogram (GCH),
and the distance between two images’ color histograms indicates how similar they are.
This technique is susceptible to changes in intensity, color shifts, and cropping. Images
are divided into blocks by local color histograms (LCH), which then provide a different
histogram for each block. Thus, a mixture of these histograms serves as an image rep-
resentation. Every block in one picture is compared to an equivalent block in the other
image to compare the two images [1].
A content-based color picture retrieval system based on color features of the visual
features and their spatial relationship was proposed by Jaehyun An et al. in [3–5]. The
proposed scheme discovers numerous dominant colors for each region after extracting
the salient regions using a color contrast method. Then, a binary map is used to charac-
terize the spatial patterns of each dominant hue. Hamed Qazanfari et al. used the colour
 Classification of High-Resolution Satellite Image 411

difference histogram (CDH) in the HSV colour space [20]. The perceptually different
color between two adjacent pixels in terms of colors and edge orientations is included
in CDH [17–19]. Utilizing entropy and correlation criteria, efficient features are chosen
from the extracted features. Feature selection, feature extraction, and similarity mea-
surement are the three steps in this methodology. It’s also proposed in [24, 25] to use
the primitives of colored instant as the framework for a color image retrieval system. A
block-by-block division of the image comes first. All blocks’ color moments are then
extracted and grouped into several classes. An image’s primal is assumed to be the aver-
age moments of each class. There are features for every primitive. The YIQ color model
is employed in this article for images that are in color [13].
It gives details on how surfaces and objects are arranged structurally within an image.
The distribution of light intensity inside an image is characterized by its texture. A pic-
ture’s texture information is crucial for defining the image’s contents. A new technique
for content-based picture retrieval using the color difference histogram (CDH) was pro-
posed by Guang-Hai Liu et al. [14]. The distinguishing feature of CDHs is that, for
colors & edge directions in L*a*b* color space, they count the perceptually uniform
color distinction between 2 points on different backgrounds [21]. Two different kinds of
special histograms used in the proposed approach work together to calculate background
colors and orientations simultaneously. A similar perceptual color change between adja-
cent pixels’ edge orientations may be seen in the first histogram. The second histogram
displays the consistent perceived color difference between the edge orientation data and
the color indices of nearby pixels. Pictures are covered by a number of elements in an
application for image categorization and retrieval. Both spatial and wavelet-changed
input images are used to extract relevant features. The wavelet transformation separates
smooth and sharp information into separate channels, degrading images into various
resolutions.to provide further context for the image.

3 System Model and Methodology

Fig. 1. System model for image retrieval

Figure 1 shows the framework of the system model for image retrieval. Here we are
using several samples from the database of high-resolution satellite images. Firstly, we
load the satellite picture databases in the MATLAB workspace. Resize the picture and
convert it from RGB to grayscale. Then perform pre-processing exercises and register
the local vector design with a reason for each pixel. Following that, extract the LBP
features and retrieve the image with the help of CBIR.
412 R. Padhy et al.

3.1 Color Feature Extraction

In this case, we extracted the colour feature from each channel and merged them. Then
the colour will effectively change if the location of the image changes. To locate the
feature, color histograms are built in various color spaces. For color spaces like RGB,
LAB, HSV, HUE, and OPP, 3-D histograms are built. Red, green, and blue make up the
RGB color space. These three colors’ combination coefficients function as a color com-
ponent. The limitations of the color spaces are not visually constant for the acquisition.
These three-color channels have a strong association. To solve issues, many color rep-
resentations have been suggested. For example, the edges are determined in inspiration
by biological color spaces like the opposing color space, hue saturation & brightness
are indeed the color spaces connected to how people describe color, and CIE La*b* &
L*uv are perceptually accurate color spaces. These two-color spaces are dependent on
the device. Such histograms, which are classified as a single feature vector, were created
for 8 × 8 × 8 = 512 color bins.
The LBP texture generator has gained popularity as a strategy in many applications
because of its discriminating capabilities & mathematical flexibility. The robustness
of a LBP operator towards periodic grey-scale changes that are based on illumination
changes is its most special feature in real-world applications. LBP’s operator used the
center pixel as a threshold when it was first operational, operating with a 3 × 3 rectangle
neighbor. In order to determine the LBP code associated with the center pixel, threshold
values were multiplied by weights whose powers of 2 then were added, as shown in (1):
⎧
⎨p−1)

LBPP,R(x) = s(gp − gc )2p
⎩
p=0 (1)

1, x ≥ 0
Where s(X) =
0, x ≤ 0

where gp is the value of the neighbouring pixel, gc is the grey value of the central pixel,
P is the number of neighbours, and R is the neighbourhood’s radius.

3.2 Image Pre-processing

RGB colour image is a collection of pixels in an image e.g.- red, green, & blue color
schemes, respectively. The 3-D RGB satellite images are transformed into 1-D grey scale
image by retaining texture features in the images.

IGS = 0.3R + 0.6 V + 0.1 B (2)

The LBP operator has the problem of modifying the output if the image size changes
even slightly. For complex photos, LBP might not perform as expected. The operator’s
thresholding strategy is too responsible for this. We suggest reducing the image’s colour
depth in order to increase the LBP’s resistance to these small variations in pixel values.
Various quantification levels were used in a set of experiments (q = 16, q = 32, q = 64,
q = 128), and the quantization also applied to both the image database as well as the
test image.
 Classification of High-Resolution Satellite Image 413

3.3 Block Based CBIR

For complete images, the CBIR systems compute textures. It keeps the database at a man-
ageable size. If the feature for the image were extracted worldwide, the textures inside
the image would not be represented properly. It is preferable to extract local features
from the proposed feature rather than use feature descriptors. This can be accomplished
by using photo segmentation or division. The segmentation procedure can be difficult
to create a good segment, especially towards the edge of the textured region that might
affect the retrieval outcome later. So, the CBIR approach is used to divide the image into
sections without applying any segmentation at all.
Every 512 × 512-pixel database image separated into 16 unique fixed-size blocks
that don’t cross. Every block’s LBP pattern is computed. The query image is treated
identically throughout the query phase. For demonstrated consistent using their LBP
attributes, a distance measure is required in the mutual information. There are a variety
of heterogeneity measurements available. Due to its efficiency in terms of the both
quickness and high retrieval rates, the Chi-square value is utilized in this work to assess
the similarity of 2 LBPV graphs, H1 and H2 , , referring to (3):


B
D1,2 = (H1 (b) − H2 (b))2 /H1 (b) − H2 (b) (3)
b=1

B is the total of histogram bins, which is determined by the parameters P & R that
were utilised. By computing the Chi-square distances between both ith the query as well
as the jth database blocks histogram, the query and dataset images can be contrasted.
The total of the lowest distances shown by (4) is the final picture similarity distance D
for retrieval:

16
 
D= minj Di,j j = 1, 2 . . . 16 (4)
i=1

The CBIR system’s performance could be assessed using a variety of objective

criteria. Return (Re), a commonly used performance indicator, is described as.

Re = N/Mt (5)

Here N represents the no. of relevant objects, & Mt is the overall number of relevant
objects that were retrieved.

4 Result and Discussions

In this part, we show how our High-Resolution Satellite Image retrieval technique can
recognize objects. Google Earth data has been used to create the database. It has been
chosen to encompass the low textural, highly textural, low structural and highly structural
area using various illumination conditions and geographic resolution. It consists of 12
categories, each comprising 50 samples. To create 16 blocks of 128 × 128 pixels each,
here we enlarged the pictures to 512 × 512 pixels.
414 R. Padhy et al.

We calculated the LBPP,R riu2 in the investigations using (P, R) parameters of (8,
1), (16, 2), and (24, 3), accordingly. The Interactive Data Language is used to create
the system for retrieving images. In this paper, we acquire a set of satellite pictures
exported from Google Earth2, which gives higher resolution satellite pictures up to
0.5 m to evaluate the current approach for indexing satellite pictures. In Fig. 2, some
samples from the database are shown (https://captain-whu.github.io/BED4RS/). There
are 50 samples of 12 types of relevant scenes, which include (a) airport, (b) forest, and
(c) parking, in high-resolution satellite images. It’s important to note that image samples
from the same class may have different scales, directions, and illuminations because
they are gathered from various regions in satellite picture resolutions.

Fig. 2. Testing samples of high-resolution satellite

All 600 pictures inside the database are deployed as query images to plot the recall
curves. The number of retrieved pictures was then multiplied by the standard recall.
The curves in Fig. 3 show that even more over 71% of the pictures are successfully
retrieved for the first 180 comparisons. Additionally, the difficulty of the LBP algorithm
for extracting features is substantially lower than that of the Gabor technique.
 Classification of High-Resolution Satellite Image 415

Fig. 3. Image retrieval result

5 Conclusion
In this paper, we discuss a content-based image retrieval system based on Local Binary
Pattern as well as a block division technique to categories high-resolution satellite pic-
tures. It means there are so many complicated images in satellite images that this is the
medium to give clarity to the messy picture. so that we can identify the whole region
using this method. The experiments demonstrate that with High Resolution Satellite
Images being compressed, Local Binary Pattern texture features can still deliver effec-
tive picture retrieval. Future research and the fact that High Resolution Satellite Image
has both frameworks and texturing imply that the addition of some structural features
would most likely result in improved results.

References
1. Maheswary, P., Srivastava, N.: Retrieval of remote sensing images using colour and texture
attribute. arXiv preprint arXiv:0908.4074 (2009)
2. Tebourbi, R., Belhadj, Z.: A texture based multispectral images indexing. In: 2005 12th IEEE
International Conference on Electronics, Circuits and Systems, pp. 1–4. IEEE (2005)
3. Ma, A.L.: Indexing and retrieval of satellite images. Doctoral dissertation, Oakland University
(2005)
4. Upreti, D., Saran, D.S., Hamm, D.N.:. Content-based satellite cloud image retrieval.
University of Twente Faculty of Geo-Information and Earth Observation (ITC) (2011)
5. Bouteldja, S., Kourgli, A.: Retrieval of high-resolution satellite images using texture features.
J. Electron. Sci. Tech 12(2), 211–215 (2014)
6. Maître, H.: Indexing and retrieval in large satellite image databases. In: MIPPR 2007: Remote
Sensing and GIS Data Processing and Applications; and Innovative Multispectral Technology
and Applications, vol. 6790, pp. 24–38). SPIE (2007)
7. Bhattacharya, A., Roux, M., Maitre, H., Jermyn, I.H., Descombes, X., Zerubia, J.: Index-
ing of mid-resolution satellite images with structural attributes. International Society for
Photogrammetry and Remote Sensing (2008)
8. Wang, S., Wang, A.: Segmentation of high-resolution satellite imagery based on feature
combination. Int. Arch. Photogramm. Remote. Sens. Spat. Inf. Sci. 37, 1223–1227 (2008)
416 R. Padhy et al.

9. Wan, Q. M., Wang, M., Zhang, X. Y., & Zhang, D. Q. (2009, October). Two-stage high
resolution remote sensing image retrieval combining semantic and visual features. In MIPPR
2009: Automatic Target Recognition and Image Analysis (Vol. 7495, pp. 1291–1301). SPIE
10. Nisia, T.G., Rajesh, S.: Classification of high-resolution Images with local binary pattern and
convolutional neural network: an advanced study. New Appr. Eng. Res. 3, 1–6 (2021)
11. Kavitha, P.K., Saraswathi, P.V.: Machine learning paradigm towards content-based image
retrieval on high-resolution satellite images. Int. J. Innov. Technol. Explor. Eng. 9, 2278–3075
(2019)
12. Liu, G.H., Yang, J.Y.: Content-based image retrieval using color difference histogram. Pattern
Recogn. 46(1), 188–198 (2013)
13. Singh, H., Kumar, A., Balyan, L.K., Singh, G.K.: A novel optimally weighted framework
of piecewise gamma corrected fractional order masking for satellite image enhancement.
Comput. Electr. Eng. 75, 245–261 (2019)
14. Asokan, A., Anitha, J.: Change detection techniques for remote sensing applications: a survey.
Earth Sci. Inf. 12(2), 143–160 (2019). https://doi.org/10.1007/s12145-019-00380-5
15. Afifi, A.J., Ashour, W.M.: Image retrieval based on content using color feature. International
Scholarly Research Notices (2012)
16. An, J., Lee, S.H., Cho, N.I.: Content-based image retrieval using color features of salient
regions. In: 2014 IEEE International Conference on Image Processing (ICIP), pp. 3042–3046.
IEEE (2014)
17. Degerickx, J., Roberts, D.A., Somers, B.: Enhancing the performance of Multiple Endmember
Spectral Mixture Analysis (MESMA) for urban land cover mapping using airborne Lidar data
and band selection. Remote Sens. Environ. 221, 260–273 (2019)
18. Dhivya, R., Prakash, R.: Edge detection of satellite image using fuzzy logic. Clust. Comput.
22(5), 11891–11898 (2017). https://doi.org/10.1007/s10586-017-1508-x
19. Qazanfari, H., Hassanpour, H., Qazanfari, K.: Content-based image retrieval using HSV color
space features. Int. J. Comput. Inf. Eng. 13(10), 533–541 (2019)
20. Bu, X., Wu, Y., Gao, Z., Jia, Y.: Deep convolutional network with locality and sparsity
constraints for texture classification. Pattern Recogn. 91, 34–46 (2019)
21. Wang, Q., Wan, J., Li, X.: Robust hierarchical deep learning for vehicular management. IEEE
Trans. Veh. Technol. 68(5), 4148–4156 (2018)
22. Li, X., Yuan, Z., Wang, Q.: Unsupervised deep noise modeling for hyperspectral image change
detection. Remote Sens. 11(3), 258 (2019)
23. Pavithra, L.K., Sree Sharmila, T., Subbulakshmi, P.: Texture image classification and retrieval
using multi-resolution radial gradient binary pattern. Appl. Artif. Intell. 35(15), 2298–2326
(2021)
24. Shih, J.L., Chen, L.H.: Colour image retrieval based on primitives of colour moments. IEE
Proc.-Vis. Image Sig. Process. 149(6), 370–376 (2002)
25. Ruichek, Y.: Attractive-and-repulsive center-symmetric local binary patterns for texture
classification. Eng. Appl. Artif. Intell. 78, 158–172 (2019)
 Author Index

Adhikari, Nibedita 204 Nanda, Arabinda 184

Akshaj, Nannapaneni 82 Nayak, Narayan 157
Alam, Ashraf 17, 47 Nayak, Sarat Chandra 299
Neogi, Biswarup 137
Badajena, J. Chandrakant 3
Balabantaray, Rakesh Chandra 273 Padhy, Rajalaxmi 409
Behera, Prafulla Kumar 222, 333 Panda, Banoj Kumar 175
Benala, Tirimula Rao 62 Panda, Ganapati 401
Bhale, Pradeep Kumar 204 Panda, Madhumita 289
Panda, Mrutyunjaya 348
Chakraborty, Soham 249
Panda, Ravi Narayan 175
Cho, Sung-Bae 299
Panigrahi, Narayan 95
Choudhury, Dwarika Nath 157
Pant, Shivam 95
Chowdhury, Manas Ranjan 184
Patra, Manas Ranjan 261
Daga, Hitesh 113 Pattnayak, Partha Sarathi 389
Das, Amardeep 204 Phaneendra, Kolloju 194
Dash, Sanjit Kumar 3, 409 Priyadarshinee, Sudipta 289
Dash, Sweta 3
Dash, Tusar Kanti 401 Rath, S. K. 204
Dehuri, Satchidananda 62, 299, 313 Rout, Chinmayee 3
Devanapalli, Suresh 194 Rout, Manas Prasad 389
Rout, Saroja Kumar 175
Ghosh, Sourajit 137 Routray, Preeti 273
Goel, Ashish 236 Routray, Sidheswar 30
Roy, Arpita 137
Habtemariam, Getachew Mekuria 148
Sahu, Sambit Subhasish 3
Jain, Aviral 113
Samantaray, Laxmipriya 409
Jayasudha, J. 365
Sarangi, Sanjaya Kumar 184
Jena, Om Prakash 389
Sarma, Pankaj Kumar Deva 313
Khetriwal, Vidipt 113 Satapathy, Pritimayee 333
Kumar, Amit 401 Seid, Hussien Worku 148
Singh, Mohit Kumar 236
Mallick, Sumitra 348 Singh, Sushant Kumar 401
Mandal, Akshaya Kumar 313 Sinha, Aritra 137
Meda, Anudeep 175 Srivastava, Jyoti 30
Mishra, Jibitesh 409 Swain, Chinmaya Kumar 273
Mishra, Niyati 379 Swain, Satyananda 261
Mishra, Soumya 401
Mishra, Subhadra 184 Thilagu, M. 365
Mishra, Sushruta 249, 379 Tripathy, Alok Ranjan 389
Mohanty, Atasi 17, 47 Tripathy, Asesh Kumar 125
Mohapatra, Pradyumna Kumar 175 Tripathy, B. K. 82, 113
Mohapatra, Seli 222 Tripathy, Hrudaya Kumar 249, 379
Mohapatra, Sudhir Kumar 148 Tripathy, Sukanta Kumar 125