International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Contents lists available at ScienceDirect

International Journal of Cognitive Computing in Engineering

journal homepage: http://www.keaipublishing.com/en/journals/international-journal-of-
cognitive-computing-in-engineering/

Enhancing personalized learning: AI-driven identification of learning styles

and content modification strategies
Md. Kabin Hasan Kanchon, Mahir Sadman, Kaniz Fatema Nabila, Ramisa Tarannum,
Riasat Khan ∗
Electrical and Computer Engineering, North South University, Dhaka, Bangladesh

ARTICLE INFO ABSTRACT

Keywords: In the rapidly advancing era of educational technology, customized learning materials have the potential
Blending ensemble technique to enhance individuals’ learning capacities. This research endeavors to devise an effective method for
Content modification detecting a learner’s preferred learning style and subsequently adapting the learning content to align with
Learning style detection
that style, utilizing artificial intelligence AI techniques. Our investigation finds that analyzing learners’ web
Named entity recognition
tracking logs for activity classification and categorizing individual responses for feedback classification are
Text-to-text transfer transformer (T5) model
highly effective methods for identifying a learner’s learning styles, such as visual, auditory, and kinesthetic.
A custom dataset has been constructed in this research comprising approximately 506 samples and 22
features utilizing the Moodle learning management system (LMS), successfully categorizing students into
their respective learning styles. Furthermore, decision tree, random forest, support vector machine (SVM),
logistic regression, XGBoost, blending ensemble, and convolutional neural network (CNN) algorithms with
corresponding optimized hyperparameters and synthetic minority oversampling technique (SMOTE) have been
applied for learning behavior classification. The blending ensemble technique with the XGBoost meta-learning
model accomplished the best performance for learning style detection with an accuracy of 97.56%. Next, the
text content of the electronic documents is modified by employing different natural language processing (NLP)
techniques, including named entity recognition of spaCy, knowledge graph, generative pre-trained transformer
3 (GPT-3), and text-to-text transfer transformer (T5) model, to accommodate diverse learning styles. Various
approaches, such as color coding, audio scripts, mind maps, flashcards, etc., are implemented to adapt the
content effectively for the detected categories of learners. The spaCy NLP-based named entity recognition
(NER) model demonstrates a 94.16% F1 score and 0.92 exact match ratio for color coding text generation of
ten electronic documents comprising 790 distinct individual words. These modifications aim to cater to the
unique preferences of learners, fostering a more personalized and engaging educational experience. To the best
of our knowledge, this is the first time an integrated learning style detection and content modification system
has been developed in this work utilizing efficient AI techniques and a private dataset.

1. Introduction videos. Auditory learners derive benefits from listening to lectures and
engaging in discussions. Kinesthetic learners, on the other hand, thrive
Learning is a complex and multifaceted process, and comprehending in hands-on activities and movements, such as experiments and role-
individual differences in learning styles is imperative for effective edu- playing. Various studies indicate a distribution ratio among visual,
cation. The term ‘‘learning style’’ pertains to an individual’s preferred auditory, and kinesthetic learners of 65:30:5 (Rogowsky, Calhoun, &
method of processing information and acquiring knowledge (Whitman, Tallal, 2020).
2023). It involves discerning how a person learns best through visual, Our motivation for undertaking this research stems from a keen
auditory, read/write, or kinesthetic (hands-on) approaches. Learning interest in understanding individual students’ diverse adaptation capa-
style theory posits that individuals exhibit distinct learning prefer- bilities and learning outcomes in response to various content and study
ences, and understanding these preferences can significantly enhance materials. Consequently, we aim to identify each student’s learning
the efficacy of teaching and learning (Newton & Salvi, 2020). Visual style and tailor the learning materials accordingly. This customiza-
tion is envisioned to render the learning process more enjoyable,
learners favor learning through visual aids such as images, charts, and

∗ Corresponding author.
E-mail address: [email protected] (R. Khan).

https://doi.org/10.1016/j.ijcce.2024.06.002
Received 24 February 2024; Received in revised form 10 June 2024; Accepted 11 June 2024
Available online 5 July 2024
2666-3074/© 2024 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

enhance engagement with educational content, and inspire motivation El Alami, Oughdir, and El Allioui (2018) proposed an approach for
to acquire knowledge in alignment with their preferred learning style. detecting learning styles using machine learning algorithms to imple-
Consequently, our objective is to detect an individual’s learning style ment adaptive e-learning systems. The authors used a private dataset
and facilitate the optimization of their learning experiences, ultimately and an unsupervised method (K-means clustering) based on FSLSM. The
contributing to improved academic performance. This work aims to naive Bayes tree technique estimated a new learner’s preferred learning
develop a learning style detection and content modification system, method with 89.06% accuracy.
offering the following contributions: Kuttattu, Gokul, Prasad, Murali, and Nair (2019) applied the visual-
auditory kinesthetic (VAK) model, focusing on the simultaneous iden-
• The work introduces a novel learning style detection system tification of multiple learning styles. The authors conducted a sur-
through a dataset comprising 506 samples with 22 attributes. vey with fifteen multiple-choice questions, gathering data from 1,123
The dataset is meticulously gathered from diverse undergraduate students. Implementing the VAK model, they utilized the k-means
and graduate students at North South University, Bangladesh. algorithm to cluster the dataset, resulting in 51 clusters. Achieving
Participants engaged in a Moodle-based course, allowing com- an accuracy rate of 92.87%, the SVM algorithm was employed to
prehensive tracking and analysis of various factors related to predict learning styles. Subsequently, they integrated the predictions
learner interaction, duration of course completion, assessment using the decision tree algorithm, incorporating inputs from cluster
of diverse content types, surveys, and feedback, facilitating the values, individual learning styles, and combinations via the Rand in-
detection of their learning styles. Employing synthetic minority dex approach. Wibirama, Sidhawara, Pritalia, and Adji (2020) dis-
oversampling technique (SMOTE) data balancing and six machine cussed the strengths and limitations of three main approaches to distin-
learning (ML) algorithms — decision tree, random forest, support guishing learning styles: the conventional, artificial, and sensor-based
vector machine (SVM), logistic regression, XGBoost and blending approaches. They used web-based learning programs, object-oriented
ensemble, and convolutional neural network (CNN) deep learn- modeling lectures for Bayesian network and NBTree algorithms, an
ing. The curated dataset is classified into three categories: visual,
accelerometer, and an eye-tracker to measure different approaches to
auditory, and kinesthetic.
learning style detection. Furthermore, they introduced a theoretical
• For the content modification, we have worked on multiple modi- way to detect learning styles utilizing SVM classification algorithms.
fication processes for the learning materials, which include color Vaishnav and Chirayu (2013) investigated three prominent learning
code, audio content, mind maps, flashcards, rhymes, etc. The
styles, visual, auditory, and kinesthetic and their influence on students’
methods we used for content modification are classification
academic performance. A sample of 200 students was employed for
method, text summarization, knowledge graph, and question–
the study. Analysis of the sample reveals that the percentage of vi-
answer techniques with the help of multiple natural language
sual, auditory, and kinesthetic learners is 33.50%, 28.50%, and 38%,
processing (NLP) models and libraries such as named entity
respectively. The study also proposes distinct learning methods for
recognition (NER) with spaCy, text-to-text transfer transformer
VAK learners. Visual learners are recommended to watch videos, draw
(T5), and generative pre-trained transformer 3 (GPT-3).
maps, take notes, and use highlights. Auditory learners are advised to
• This work aims to identify learning styles and modify contents
participate in group discussions, record lectures, and listen to videos.
based on particular learning styles. Through this, we aim to
Kinesthetic learners are encouraged to study in brief sessions and attend
enhance personalized learning experiences by accurately detect-
lab classes. Thushara, Mownika, and Mangamuru (2019) conducted a
ing learning styles and modifying content to cater to individual
comparative study on keyword and keyphrase extraction algorithms
learners’ needs.
without utilizing a corpus. The authors implemented TextRank, Rapid
While learner-type detection and content modification exist sepa- Automatic Keyword Extraction (RAKE), and PositionRank algorithms
rately, the novelty of this work is to develop an integrated system for the study. While the RANK algorithm encompasses candidate and
with specialized features like mind maps, audio content, flashcards, keyword selection, TextRank and PositionRank involve graph-building,
etc., enhancing personalized learning experiences based on advanced sorting, and forming keyphrases. The authors conducted experiments
machine learning and NLP techniques. with the algorithms on research documents. The applied PositionRank
We present the related works concerning learning style classification model provided the most relevant keyphrases in less time than Tex-
and didactic style in Section 2 and discuss the proposed system in tRank and RAKE. TextRank requires more time for large documents,
Section 3, where we have presented the datasets and models used and the keyphrases extracted from RAKE tend to be lengthy and less
for this research. Section 4 is dedicated to the significant results of pertinent.
the proposed learning style detection and content modification system. Obeng (2023) implemented machine learning techniques for auto-
Finally, Section 5 concludes the article with some directions for the matic learning style classification of computer science undergraduate
future improvement of this work. students. The applied OneR technique attained approximately 80%
accuracy for classifying 198 instances and five categories of VARK
2. Related works didactic styles. Ahmad Muhammad, Qi, Wu, and Kabir Ahmad (2022)
assigned a graph representation-based embedded learning approach for
Learning style detection has long captivated the interest of re- learning style detection in an online class. The model demonstrated
searchers in the field of education, as educators endeavor to craft 88.25% accuracy and a 0.87 F1-score for the KDDCup dataset. Nguyen
personalized and engaging learning experiences for their students. The et al. (2023) demonstrated the XGBoost ensemble approach produced
traditional available approaches of detecting learning styles and con- 96% accuracy for an automatic learning style identification custom
verting content to individual preferences are implemented by various dataset of 550 volunteers.
independent tools. This section demonstrates a comprehensive review Naseer et al. (2021) discussed different NER techniques and the
of relevant studies to explore different methodologies employed by advantages and disadvantages of various libraries and compared them
recent artificial intelligence (AI)-based articles on automatic learning based on their performances. The NER techniques they discussed in-
style classification and content modification. clude rule-based, learning-based, and hybrid approaches. Among the
Rasheed and Wahid (2021) applied the Felder–Silverman learning libraries and NER tools, spaCy, StanfordNLP, TensorFlow, and Apache
style model (FSLSM) and a diverse questionnaire of 498 samples to OpenNLP are the subject of the discussion. The NER tools show that
classify learning styles. The authors used various ML algorithms, where spaCy has the highest F1-score, followed by TensorFlow, OpenNLP, and
the SVM algorithm attained the highest accuracy of 75.55% and preci- StanfordNLP. spaCy provides the best result with 100% training accu-
sion and recall of 73% and 76%, respectively. El Aissaoui, El Madani racy, F1 score, and prediction probability and with the least training

270
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

loss compared to all the tools. Ranganathan and Abuka (2022) proposed 3.2. Dataset preprocessing
a text summarization technique using the T5 model. To perform the
theoretical text summarization, the authors fine-tuned the T5 model. The curated dataset consists of four sections, each represented
The authors used the University of California, Irvine’s drug reviews and by an Excel file containing grades, participation, activity logs, event
BBC news datasets. The efficacy of the model was measured using the completion, and event types (ground truth). We mapped each section
ROUGE metrics with ROUGE-1 values of 45.62 and 69.05 for the UCI using the ‘‘User full name’’ column from the activity logs and concate-
and BBC News datasets, respectively. nated all sections into a single data frame. The resulting data frame
Popescu-Belis et al. (2023) implemented a novel solution to gener- comprises 506 samples and 22 columns. We have carefully examined
ate synthetic data based on a rhyming algorithm that helps learn to
the data for null values and duplicates, subsequently removing any
rhyme. The authors used the GPT-2 model and successfully obtained
rows exhibiting either. The ‘‘Time’’ column contains timestamps for
60% of rhyming lines of (AABB) consecutive rhymes. However, it
each event completion or initiation by each student. Next, this column
failed to work with alternative versus (ABAB) as it has longer-range
is converted to the ‘‘Datetime’’ datatype and chronologically sorted the
dependencies. Chen, Jia, and Xiang (2020) represented the knowledge
DataFrame based on both ‘‘User full name’’ and ‘‘Time’’. Subsequently,
graph, which works according to knowledge reasoning. Pirnau et al.
we generated a new column, ‘‘Time difference’’, by calculating the
(2024) initiated the development of NLP-based automatic techniques
time difference between consecutive events for each user using the
for content analysis in scientific research articles. The authors used over
1,000 COVID-19-related articles to classify its contents into three cate- diff() method with user grouping. The initial time difference is set
gories. The Bidirectional Encoder Representations from Transformers to 0. The label encoding approach (Talib, Albahri, & Edoh, 2023) is
(BERT) model demonstrated the most efficient results with a cosine applied to convert categorical columns into numerical representations.
similarity of 94%. In a similar article, Tunca, Sezen, and Balcioglu We assessed the data balance and identified the presence of outliers. We
(2023) applied various NLP approaches to investigate the content of applied the SMOTE technique to achieve balanced data representation.
newspaper articles. Among the applied models, the BERT technique Finally, the preprocessed dataset has been split into training and testing
obtained the maximum F1 score of 73%. sets using an 8:2 ratio. Hyperparameters optimization (Tasin, Nabil,
Significant work on learning style detection and didactic content Islam, & Khan, 2023) of the applied models is performed on the training
modification has been completed individually in various articles. How- samples. This comprehensive data preprocessing pipeline ensures our
ever, existing solutions relying on features such as user surveys and data is well-suited for subsequent machine learning analysis.
eye movements to identify learning styles have proven ineffective and
are often constrained in their implementation settings. In contrast, 3.2.1. Preprocessing techniques for content modification
the proposed solution presented herein is designed to seamlessly inte- To prepare the raw text data of the electronic document for further
grate with standard learning management systems (LMSs), eliminating analysis, we employed a comprehensive preprocessing pipeline, i.e., to-
the need for additional devices or dedicated time for learning be- kenization, segmenting the text into individual units, lemmatization
havior detection. While some content converters exist independently, to normalize words to their base forms, and grouping words with
centralized systems are limited to specific models or conversions. Con-
similar meanings. Additionally, the texts are converted to lowercase
sequently, this work introduces a comprehensive and generalized learn-
for consistency, and stop words, which hold minimal semantic value,
ing style classification and content modification system to address these
are removed. Finally, we eliminated extraneous symbols and condensed
limitations.
multiple spaces into single spaces using regular expressions within a
Python library, effectively cleaning and standardizing the text data for
3. Proposed system
downstream processing.
The proposed AI-based learning style detection and content modifier
system is divided into two major parts, as shown in Fig. 1. The proposed 3.3. Learning style detection
system identifies a user’s learning style from their LMS activity and
modifies content accordingly. A custom dataset has been created to To identify a student’s learning style, we input the student’s data
train various AI models for learning style detection. Different NLP from the LMS and utilize the corresponding parameters to the identify-
techniques were used to modify the content based on the detected ing methods. In this method, we identify the student’s style by his/her
learning style. The system can convert text content into audio and activities in the Moodle Learning Management System, such as average
visual formats to cater to different learning preferences. The descrip- video-watching time, audio listing time, quiz scores, lesson completion
tions of individual components of the proposed system are presented time, etc. Given the absence of a suitable existing dataset for activity
in subsequent paragraphs. classification, a customized dataset has been constructed in this work.
The actual label (ground truth) of a student’s learning style has been
3.1. Dataset
determined from the corresponding student’s comments and feedback
in text form as input.
In this work, a customized dataset comprising 506 samples and 22
features has been created by introducing a course to bachelor’s and
master’s students of North South University, Bangladesh, utilizing the 3.3.1. Machine learning models
Moodle LMS platform. This course included videos, audio, and text, and Various machine learning models, decision tree, random forest,
students engaged with the content by taking quizzes, participating in SVM, logistic regression, and XGBoost techniques have been used for
a survey, and providing feedback. Moodle allowed us to track time, learning approach classification of the customized dataset.
assess marks, and give feedback. Treating this information as data, we
discern the learner types. Students aged 19 to 27 years participated • Decision Tree (DT): The DT uses a flowchart-like tree structure to
in this study, with the gender distribution being 57.91% male and learn simple decision rules and predict outcomes. It begins with a
42.09% female. Fig. 2 summarizes the demographic distribution of root node, splits into branches based on the decision rules, travels
the custom dataset. The dataset was collected from July to October through child nodes, and ends with the leaf nodes that denote
2023. We considered students’ survey responses the ground truth for the result of the algorithm (Lee, Lee, Mun, & Kim, 2022). The
identifying their learning styles. From the dataset, we found that 65% uncertainty in the dataset decides whether a feature should be the
of the students who took the course were visual learners, 30% were root node, child node, or leaf node. Although DT is a supervised
auditory learners, and 5% were kinesthetic learners, as shown in Fig. 3. algorithm, it can also solve regression problems.

271
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Fig. 1. Working sequences of the proposed learning style detection and content modifier system.

• Random Forest: A random forest classifier combines multiple

decision trees generated randomly. Then, each tree generates a
prediction, and the prediction that receives the most votes is the
outcome (Soliman, 2020). Using a bootstrap method, a classifier
is built randomly selecting different features. When compared to
other classifiers, it gives more accurate predictions.
• SVM: Binary classification is carried out by SVM on the mapped
data points. It plots a hyperplane by separating the two distinct
classes of data points (Ben Seghier, Carvalho, de Faria, Correia, &
Fakury, 2023). It maximizes the distance between the hyperplane
and the closest data points to the hyperplane. Solving multi-
classification problems with SVM requires splitting the problem
into minor binary classification problems.
Fig. 2. Demographic distribution of the custom dataset. • Logistic Regression: Logistic regression uses a statistical method
to predict dichotomous outcomes but can be extended to solve
multi-classification problems. It does not face the shift in thresh-
old value issue as it uses the sigmoid function, an S-shaped
curve. To represent two discrete classes, it predicts probabilities
to maintain the prediction range between 0 and 1 (Shah, Patel,
Sanghvi, & Shah, 2020).
• XGBoost: XGBoost is a decision tree-based algorithm that utilizes
gradient boosting to improve model accuracy. It has been demon-
strated to be highly effective for various machine learning tasks
and is among the most popular machine learning libraries. The
hyperparameters, e.g., number of decision trees, learning rate,
maximum features, etc., tuning has a significant impact on the
model’s performance.
• Blending ensemble technique: In this study, we have developed
a blending ensemble technique that integrates learning from
three base models. Specifically, we employed three heterogeneous
models—random forest, SVM, and decision tree with the corre-
sponding optimized hyperparameters—to generate an XGBoost-
Fig. 3. Percentage of various learning style classes in the custom dataset. based meta-learner. The intelligence and predictive capabilities
of these approaches were utilized for training the meta-model
(Naveen Venkatesh, Divya, Sugumaran, & Mohammadreza, 2024).
Subsequently, both the input and outputs of the trained XGBoost

272
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Fig. 4. Working sequences of the proposed blending ensemble model for learning style detection.

Fig. 5. Architecture of the proposed CNN technique for learning style detection.

meta-model were considered as the final predictions. Working 3.4. Content modifier
sequences of the proposed blending ensemble model for learning
style detection are illustrated in Fig. 4 and Algorithm 1. The text of an electronic document (PDF, text, Word document,
etc.) is considered learning content and converted into various learning
Algorithm 1 Algorithm of the proposed blending ensemble model for contents for three different learning styles.
learning style detection
• Visual: Three contents can be created for visual learners.
Start
Step 1: Split the custom dataset into training, validation, and test. – Color Code: Combining the Named Entity Recognition and
Step 2: Train the random forest, SVM, and decision tree base models Key Phase Extraction techniques, we are identifying the
with the training samples. words, which are persons’ names, locations, year, time, etc.
Step 3: Generate the new dataset to train the XGBoost meta-model. These words will be colored according to the type so that a
Step 4: Train the meta-model with the validation samples. virtual learner can easily notice the necessary information.
Step 5: Inference using the test samples with the trained meta-model. – Mind Map: Mind maps are helpful to visual learners to
End understand a topic better. We use Key Phase Extraction and
Knowledge Graph techniques to extract the key points and
• Convolutional Neural Network (CNN): CNN is an integral com- the relations between the elements.
ponent of artificial neural networks and a widely utilized deep – Flashcards: Using Knowledge Graph, Text-to-image, and
Question-answering techniques, we enlist the most essential
learning algorithm. It consists of layers that use convolving filters
information and represent that. It helps visual learners to
to extract local features. This study employs two one-dimensional
memorize better.
convolutional layers with a kernel size of three and 32 filters.
Furthermore, a dropout value of 0.2 was set, and a MaxPooling1D • Auditory: Three contents can be created for auditory learners.
layer was added. The activation function utilized in the convo-
– Audio Script: Auditory learners grasp the study matters
lutional part was ‘‘ReLU’’, while the final dense layer employed
best by listening, so we are creating a script for the audio
the softmax function. KerasTuner framework has been used to
lesson. The lesson is designed with some questions and men-
select the values of various hyperparameters of the applied CNN tions what to expect in this lesson at the beginning. After
model. The proposed CNN technique’s architecture employed in that, the lesson will continue following the primary learn-
this work is demonstrated in Fig. 5. ing content. Text summarization and question-generation
• Synthetic Minority Oversampling Technique (SMOTE): SMOTE is techniques have been applied to construct the audio scripts.
a popular machine learning preprocessing approach for unbal- – Summary: After an audio lesson, an overview of the lesson
anced data (Solayman, Aumi, Mery, Mubassir, & Khan, 2023). is best for this learner to take out the key points. We use the
To adjust imbalances in datasets, it creates synthetic samples text summarization technique to get an abstract summary of
focusing on the minority class. the learning content.

273
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

– Rhyme: Rhymes are the best ways for auditory learners

to memorize something. Using the text generation tech-
nique and appropriate prompts, we generate short rhymes
containing the learning content’s key points.

The text contents can be converted into audio content using the
text-to-speech technique, as audio is the best way for auditory
learners to learn.
• Kinesthetic: For auditory learners, efficient content can be devel-
oped.

– Activity List: As kinesthetic learners excel in learning through

hands-on activities, we devised a methodology involving
formulating a set of questions derived from the learning con-
tent. Subsequently, students engage in self-directed learning Fig. 6. Balancing the training data using SMOTE.
by seeking answers through various sources, such as text-
books or the Internet. Following this preparatory phase, the
Table 1
actual lesson commences, and once again, we employ the
Hyperparameters of the applied blending ensemble technique.
question-generation technique.
Estimator 1 Estimator 2 Estimator 3 Meta estimator Cross- Pass-
Finally, the selected converted contents are provided to the stu- validation through

dent according to the detected learning style. Random SVM Decision XGBoost None False
Forest Tree (Blending) (0)
3.4.1. Applied NLP techniques
We used various techniques for different types of modification of
the contents. A brief discussion of these methods is stated below.
4.1. Learning style detection
• SpaCy: SpaCy is a library for NLP built to extract information (Hu,
Gong, & He, 2022). The text is initially tokenized using spaCy, The learning styles within the custom dataset have been classified
which then undergoes a processing pipeline that includes an using various machine learning techniques. Due to the kinesthetic
entity recognizer, lemmatizer, tagger, and parser. category containing fewer instances compared to visual and auditory
• Named Entity Recognition: Named Entity Recognition (NER), an learners, the SMOTE data balancing approach has been employed.
element of natural language processing (NLP), also known as Optimization of different hyperparameters of the applied models has
entity chunking or entity extraction, recognizes preset categories resulted in improved performance.
of objects inside a text (Jehangir, Radhakrishnan, & Agarwal, Initially, the visual (label: 2), auditory (label: 0) and kinesthetic
2023). In this work, the NER model of the spaCy library has been (label: 1) categories constitute 122, 20, and 263 training instances,
used. respectively. Following the application of the SMOTE method, each
• Key Phase Extraction: In natural language processing (NLP), category comprises 300 samples. Fig. 6 shows the imbalanced and
keyphrase or keyword extraction is a text analysis approach that balanced number of samples concerning various classes of the training
pulls keywords and phrases out of the input text. Numerous
dataset.
activities can benefit from using these key terms, including con-
Table 1 presents the hyperparameters used in the proposed blending
tent categorization, document summarization, and information
ensemble technique, where Random Forest, SVM, and Decision Tree
retrieval.
• Knowledge Graph: A Knowledge Graph is a knowledge base that are employed as base estimators, and XGBoost is used as the meta
employs a graph-structured data model (Kejriwal, 2022). It is estimator. The blending process does not utilize cross-validation, and
a type of network graph showing qualitative connections be- the pass-through parameter is set to false, indicating that the original
tween facts, concepts, and events of real-world entities. The mind features are not passed to the meta-estimator.
map for the visual learner has been generated employing this Table 2 describes various metrics of the applied models for learning
approach. style classification. According to this table, the blending ensemble
• Text-to-Text Transfer Transformer (T5): T5 restructures different model with its optimized hyperparameters achieves the maximum ac-
NLP actions into a single text-to-text format, with text strings as curacy of 97.56%, with the highest number of correct predictions out of
both the input and the output (Raffel et al., 2023). This tech- the total predictions. The blending model’s F1 score of 96.76% indicates
nique considers the model text as input and trains it to generate that it maintains an outstanding trade-off between correctly identifying
target text. Translation, question-answering, summarization, and positive instances and overall correct predictions across various classes
categorization are comprehensively produced using a text-to-text within the dataset.
methodology. In this study, this transformer has been employed The total training time in seconds of the applied models is illustrated
for question–answer generation and text summarization. in Table 3. According to this table, the CNN model comprises the
• Generative Pre-trained Transformer 3 (GPT-3): To carry out a longest training time of 600.37 s, followed by the blending ensemble
range of natural language activities, GPT-3 processes text input. technique. The decision tree and SVM techniques have significantly
It combines natural language processing with natural language shorter training times of 70.99 s and 79.41 s, respectively, indicating
generation to comprehend and produce text written in real human faster training processes than more complex models like CNN and
language (Brown et al., 2020). For generating automated text of
blending ensemble.
rhyme for auditory learners, we used a GPT-3 transformer model
Fig. 7 shows that both training and test accuracy rapidly increase
in this study.
and stabilize around 96.8% after a few epochs for the applied CNN
4. Results and discussion deep learning model. The training and test loss sharply decrease and
converge, indicating effective learning and minimal overfitting. It is
The performance of the proposed AI-based automatic learning ap- worth mentioning that, different hyperparameters of the proposed CNN
proach detection and content modifier system has been measured uti- model are optimized utilizing the KerasTuner framework, leading to
lizing various metrics, which have been discussed in this section. improved performance.

274
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Table 2
Performance metrics of the applied models for learning style classification.
Model Accuracy F1 score Precision Recall/Sensitivity Specificity AUC-ROC
Decision Tree 92.63% 91.14% 91.26% 91.02% 91.84% 0.91
Random Forest 96.34% 94.78% 94.92% 94.65% 94.95% 0.94
SVM 95.57% 94.07% 94.13% 94.01% 94.79% 0.94
Logistic Regression 91.48% 89.71% 89.84% 89.58% 90.74% 0.90
CNN 96.82% 96.13% 96.24% 96.03% 96.45% 0.96
XGBoost 97.03% 96.39% 96.63% 96.17% 96.66% 0.96
Blending 97.56% 96.76% 96.94% 96.59% 96.41% 0.96

Table 3 auditory, and kinesthetic. In this research, we meticulously utilized

Total training time in seconds of the applied models.
various evaluation metrics to discern the effectiveness and performance
Decision Random SVM Logistic CNN XGBoost Blending of the applied models for content modification.
tree forest regression ensemble
Hyperparameters used in tokenizer and output generation of textual
70.99 212.98 79.41 84.69 600.37 244.82 612.52
contents of various electronic documents for three categories of learners
are shown in Tables 4 and 5, respectively. According to this table,
the early stopping method is used for rhyme generation for auditory
learners.
Figs. 9 and 10 depict samples of textual electronic documents
and their converted contents tailored for visual, auditory, and kines-
thetic categories of learners, respectively. The auditory learner content
comprises summarized text, audio, and rhyme generated by the T5
transformer and GPT-3 models. Texts with distinct colors for visual
learners are created using the spaCy-based NER model, flashcards, and
mind maps generated with T5 and knowledge graphs. For kinesthetic
learners, broad questions and corresponding activities are developed
employing the T5 model. Notably, during the generation of summaries
and questions, the ‘‘output_scores’’ parameter was set to ‘‘True’’ to
identify the sequence with the highest value.
In this work, ten original electronic documents comprising 790
distinct individual words are converted to various content for specific
learners. The performance of the spaCy library for generating color-
coding NER texts for visual learning is illustrated in Table 6. The
spaCy framework efficiently transformed the original text tokens into
seven categories: person, date, NORP, ORG, event, GPE, and plain. The
Fig. 7. Training and test accuracy and loss vs. epoch of the proposed CNN model.
confusion matrix of the spaCy NER library is shown in Fig. 11. The
spaCy NLP-based NER model also performs well in identifying date,
NORP, event, and GPE entities with minimal errors. However, some
misclassifications include confusing plain text with person, date, NORP,
ORG, and event.
Finally, 48 volunteers (29 male and 21 female) provided their
review of the proposed automatic content modification system. The
ratings of the feedback of various content modifications, e.g., summa-
rized text, audio output, rhyme, colored test, flashcard, mind map, and
kinesthetic activity generations, are illustrated in Fig. 12. The average
rating of 4.66 for various features demonstrates the effectiveness of the
implemented system.
A brief comparison of the proposed learning style detection and
content modification system with related articles is illustrated in Ta-
ble 7. According to this table, the major contribution of this work is
the development of an integrated system with a custom dataset and
outstanding efficiency.
Fig. 8. Confusion matrix of test samples for the XGBoost model of learning style
detection.
4.3. Limitations

The following are the limitations of the proposed AI-based auto-

The confusion matrix of 101 test samples for the XGBoost model
matic learning style detection and content modification work:
of learning style detection is illustrated in Fig. 8. This figure summa-
rizes the performance of the applied XGBoost model by providing a • The Moodle course employed to generate the private dataset
detailed snapshot of a model’s effectiveness in the proposed multiclass includes videos, audio, and text, and students engaged with the
classification of didactic genres. content by taking quizzes, participating in a survey, and providing
feedback. A significant limitation of this study is that it needs to
4.2. Content modifier incorporate video recordings and audio input from students.
• The proposed study has not been assessed to determine the ef-
Textual contents of various electronic documents are modified using fectiveness of personalized content in improving the learning
different AI-based NLP techniques tailored to the learner’s style, visual, outcomes of the participating students.

275
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Fig. 9. Converted content for visual learners employing colored texts, flashcards, and mindmap.

Table 4
Hyperparameters used in tokenizer.
Hyperparameters/Tasks Max length Pad to max length Truncation Return tensor
Summarization 512 False True ‘pt’
Short Question Generation 512 False True ‘pt’
Short Answer Generation X False False ‘pt’
Broad Question Generation 512 True True ‘pt’
Rhyme Generation X False False ‘pt’

Table 5
Hyperparameters used in output generation.
Hyperparameters/ Num beams Early Num return No repeat Min Max Output Return dict
Task stopping sequences NGram size length length scores in generate
Summarization 3 False 3 2 75 200 True True
Short Question Generation 3 False 3 3 20 50 True True
Short Answer Generation 1 False 1 0 10 X False False
Broad Question Generation 10 False 3 2 10 40 True True
Rhyme Generation 5 True 1 2 10 3000 False False

Table 6 instances and 22 features was curated, encompassing three classes:

Performance metrics of the spaCy NER library for color coding text generation.
visual, auditory, and kinesthetic. The applied blending ensemble tech-
Precision Recall Accuracy F1 score Exact match ratio
nique with the XGBoost meta learner and optimized hyperparame-
94.55% 93.78% 94.67% 94.16% 0.92 ters achieved the best performance in learning behavior classification,
yielding accuracy and F1 score of 97.56% and 96.76%, respectively.
Subsequently, textual electronic documents were transformed into sum-
• Currently, the Moodle-based LMS framework is only used to marized texts, activities, mind maps, and flashcards tailored to diverse
create the dataset. Integrations of AI systems within the LMS learners, leveraging GPT-3, T5 transformer, spaCy NER, and knowledge
platform to track students’ progress in real-time, affecting their graph NLP techniques. The spaCy NLP-based NER model demonstrates
academic performances, are yet to be implemented. a 94.16% F1 score and 0.92 exact match ratio for color coding text
generation of ten electronic documents comprising 790 distinct indi-
5. Conclusion vidual words. While learner-type detection and content modification
exist separately, this work uniquely develops an integrated system with
In this study, we present the development of an automated sys- specialized features, e.g., mind maps, audio content, flashcards, etc., en-
tem for detecting learning styles and modifying content using ma- hancing personalized learning experiences based on advanced machine
chine learning and NLP techniques. A custom dataset comprising 506 learning and NLP techniques. The proposed system offers valuable

276
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

Table 7
Comparison of the proposed learning style detection and content modification system with other works.
Ref. Dataset Learning style detection Content Accuracy F1 score
modification
Rasheed and Wahid (2021) Custom Decision tree N/A 92% 92%
Samples: 498
Kuttattu et al. (2019) Custom SVM N/A 92.87% N/A
Samples: 1,123
Lokare and Jadhav (2024) Custom EEG Random forest N/A 87.5% 86.8%
Samples: 52
Obeng (2023) Custom OneR N/A 80% 87.5%
Samples: 198
Ahmad Muhammad et al. (2022) Custom Graph representation N/A 88.25% 78.50%
Samples: 198 learning
Nguyen et al. (2023) KDDCup XGBoost N/A 96% 96%
Public
Proposed Custom Blending ensemble with GPT-3, T5, 97.56% 96.76%
Samples: 506 XGBoost meta learner spaCy, etc.

Fig. 10. Converted content for (a) auditory and (b) kinesthetic learners.

Fig. 12. User feedback ratings of the proposed content modification system.

input from students can facilitate the creation of a multimodal dataset,

Fig. 11. Confusion matrix of the spaCy NER library.
thereby enabling more personalized learning opportunities and poten-
tially improving academic outcomes across educational institutions. In
the future, A/B testing with control groups or long-term studies can be
insights into understanding optimal learning modalities and improv- performed to observe changes in the student’s academic performances.
ing learning experiences by personalizing academic content based on
individual learning styles.
Future endeavors may involve the expansion of the dataset to Declaration of competing interest
include more diverse instances, aiming to gather additional insights.
The adoption of a uniform tokenizer can be applied to streamline The authors declare that they have no known competing finan-
and enhance the efficiency of model development and refinement cial interests or personal relationships that could have appeared to
processes. Furthermore, the integration of video recordings and audio influence the work reported in this paper.

277
M.K.H. Kanchon et al. International Journal of Cognitive Computing in Engineering 5 (2024) 269–278

References Obeng, A. (2023). Consequential effects of using competing perspectives to predict

learning style in e-learning systems. Cogent Education, 10.
Ahmad Muhammad, B., Qi, C., Wu, Z., & Kabir Ahmad, H. (2022). GRL-LS: A learning Pirnau, M., Botezatu, M. A., Priescu, I., Hosszu, A., Tabusca, A., Coculescu, C., et al.
style detection in online education using graph representation learning. Expert (2024). Content analysis using specific natural language processing methods for big
Systems with Applications, 201, URL https://www.sciencedirect.com/science/article/ data. Electronics, 13, URL https://www.mdpi.com/2079-9292/13/3/584.
pii/S0957417422005346. Popescu-Belis, A., Atrio, À. R., Bernath, B., Boisson, É., Ferrari, T., Theimer-
Ben Seghier, M. E. A., Carvalho, H., de Faria, C. C., Correia, J. A., & Fakury, R. H. Lienhardt, X., et al. (2023). GPoeT: A language model trained for rhyme generation
(2023). Numerical analysis and prediction of lateral-torsional buckling resistance of on synthetic data. In Workshop on computational linguistics for cultural heritage, social
cellular steel beams using FEM and least square support vector machine optimized sciences, humanities and literature.
by metaheuristic algorithms. Alexandria Engineering Journal, 67, 489–502, URL Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., et al. (2023).
https://www.sciencedirect.com/science/article/pii/S111001682200847X. Exploring the limits of transfer learning with a unified text-to-text transformer.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al. arXiv:1910.10683.
(2020). Language models are few-shot learners. In Conference on neural information Ranganathan, J., & Abuka, G. (2022). Text summarization using transformer model.
processing systems. In International conference on social networks analysis, management and security (pp.
Chen, X., Jia, S., & Xiang, Y. (2020). A review: Knowledge reasoning over knowledge 1–5).
graph. Expert Systems with Applications, 141, Article 112948. Rasheed, F., & Wahid, A. (2021). Learning style detection in E-learning systems using
El Aissaoui, O., El Madani El Alami, Y., Oughdir, L., & El Allioui, Y. (2018). A hybrid machine learning techniques. Expert Systems with Applications, 174.
machine learning approach to predict learning styles in adaptive E-learning system. Rogowsky, B., Calhoun, B., & Tallal, P. (2020). Providing instruction based on students’
In International conference on advanced intelligent systems for sustainable development learning style preferences does not improve learning. Frontiers in Psychology, 11.
(pp. 772–786). Shah, K., Patel, H., Sanghvi, D., & Shah, M. (2020). A comparative analysis of logistic
Hu, C., Gong, H., & He, Y. (2022). Data driven identification of international cutting regression, random forest and KNN models for the text classification. Augmented
edge science and technologies using SpaCy. PLoS One, 17, 1–24. http://dx.doi.org/ Human Research, 5, 1–16.
10.1371/journal.pone.0275872. Solayman, S., Aumi, S. A., Mery, C. S., Mubassir, M., & Khan, R. (2023). Automatic
Jehangir, B., Radhakrishnan, S., & Agarwal, R. (2023). A survey on named entity COVID-19 prediction using explainable machine learning techniques. Interna-
recognition — datasets, tools, and methodologies. Natural Language Processing tional Journal of Cognitive Computing in Engineering, 4, 36–46, URL https://www.
Journal, 3, Article 100017, URL https://www.sciencedirect.com/science/article/pii/ sciencedirect.com/science/article/pii/S2666307423000037.
S2949719123000146. Soliman, H. (2020). Random forest based searching approach for RDF. IEEE Access, 8,
Kejriwal, M. (2022). Knowledge graphs: A practical review of the research landscape. 50367–50376.
Information, 13, URL https://www.mdpi.com/2078-2489/13/4/161. Talib, H., Albahri, A., & Edoh, T. (2023). Fuzzy decision-making framework for
Kuttattu, A. S., Gokul, G., Prasad, H., Murali, J., & Nair, L. S. (2019). Analysing the sensitively prioritizing autism patients with moderate emergency level. Applied Data
learning style of an individual and suggesting field of study using machine learning Science and Analysis, 2023, 16–41.
techniques. In International conference on communication and electronics systems (pp. Tasin, I., Nabil, T. U., Islam, S., & Khan, R. (2023). Diabetes prediction using machine
1671–1675). learning and explainable AI techniques. Healthcare Technology Letters, 10, 1–10, URL
Lee, S., Lee, C., Mun, K. G., & Kim, D. (2022). Decision tree algorithm considering https://ietresearch.onlinelibrary.wiley.com/doi/abs/10.1049/htl2.12039.
distances between classes. IEEE Access, 10, 69750–69756. Thushara, M., Mownika, T., & Mangamuru, R. (2019). A comparative study on
Lokare, V. T., & Jadhav, P. M. (2024). An AI-based learning style prediction model different keyword extraction algorithms. In International conference on computing
for personalized and effective learning. Thinking Skills and Creativity, 51, URL methodologies and communication (pp. 969–973).
https://www.sciencedirect.com/science/article/pii/S187118712300189X. Tunca, S., Sezen, B., & Balcioglu, Y. S. (2023). Content and sentiment analysis of the
Naseer, S., Ghafoor, M. M., bin Khalid Alvi, S., Kiran, A., Rahmand, S. U., Murtazae, G., New York Times coronavirus (2019-nCOV) articles with natural language processing
et al. (2021). Named entity recognition (NER) in NLP techniques, tools accuracy (NLP) and Leximancer. Electronics, 12, URL https://www.mdpi.com/2079-9292/12/
and performance. Pakistan Journal of Multidisciplinary Research, 2, 293–308. 9/1964.
Naveen Venkatesh, S., Divya, S., Sugumaran, V., & Mohammadreza, A. (2024). Vaishnav, R. S., & Chirayu, K. (2013). Learning style and academic achievement of
Detection of visual faults in photovoltaic modules using a stacking ensem- secondary school students. Voice of Research, 1, 1–4.
ble approach. Heliyon, 10, URL https://www.sciencedirect.com/science/article/pii/ Whitman, G. M. (2023). Learning styles: Lack of research-based evidence. The Clearing
S2405844024039252. House: A Journal of Educational Strategies, Issues and Ideas, 96, 111–115. http:
Newton, P. M., & Salvi, A. (2020). How common is belief in the learning styles neu- //dx.doi.org/10.1080/00098655.2023.2203891.
romyth, and does it matter? A pragmatic systematic review. Frontiers in Education, Wibirama, S., Sidhawara, A. P., Pritalia, G. L., & Adji, T. B. (2020). A survey of learning
5, URL https://www.frontiersin.org/articles/10.3389/feduc.2020.602451. style detection method using eye-tracking and machine learning in multimedia
Nguyen, H.-H., Nguyen Duc, L., Do Trung, K., Dang Hoang, L., Vu, T., Vu, T. V., et al. learning. In International symposium on community-centric systems (pp. 1–6).
(2023). Applying machine learning techniques to detect student’s learning styles. In
International conference on education technology and computers (pp. 456–462). New
York, NY, USA: Association for Computing Machinery, http://dx.doi.org/10.1145/
3572549.3572622.

278