PROJECT REPORT

(PROJECT TERM JANUARY-MAY2024)

Predicting student performance in education

ACKNOWLEDGEMENT
It is with my immense gratitude that I acknowledge the support and help of my Professor, Dr. Premananda Sahu,
who has always encouraged me into this research. Without his continuous guidance and persistent help, this project
would not have been a success for me. I am grateful to the Lovely Professional University, Punjab and the
department of Computer Science without which this project would have not been an achievement. I also thank my
family and friends for their endless love and support throughout my life.

TABLE OF CONTENTS Page No

1. Title Page 1

2. Declaration 2

3. Certification 3

4. Acknowledgement 4

5. Abstract 6

6. Introduction 7

7. Related work 8

8. Methodology 8

9. Result and Discussion 9

10. Conclusion 11

11. References 12
 ABSTRACT

The field of study on predicting student performance in educational institutions has gained significant importance
and holds the potential to transform teaching and learning methodologies. The usefulness and viability of using
data-driven methods to forecast student outcomes are examined in this study. The abstract examines the state of
education today, emphasizing the difficulties teachers have in spotting pupils who might perform below
expectations. The idea of feature creation is then explored in depth, with a focus on the procedure of gathering
pertinent data points from several sources, such as academic records, demographic data, and metrics related to
student involvement.

The use of statistical modeling and machine learning approaches to forecast student performance is explained in
this abstract. It talks on the benefits of using several algorithms to find hidden patterns in educational data, such as
support vector machines, random forests, and linear regression. The significance of model assessment is further
explored in the abstract, which also outlines techniques for evaluating the built prediction models' generalizability,
accuracy, and precision.

The abstract also highlights the possible advantages of using these prediction models in educational institutions.
The timely identification of kids who are considered to be at-risk enables educators to customize interventions and
offer focused assistance, thus promoting enhanced academic performance and overall student success. It recognizes
the possibility for bias in the models as well as the ethical issues surrounding the use of educational data.

The abstract ends by summarizing the potential directions for this field of study. It investigates the possibilities of
combining these models with tailored learning environments and adaptive learning systems. Furthermore, the
abstract delves into the necessity of ongoing research and development to enhance current models and investigate
innovative methods for forecasting student success in the always changing field of education.
 INTRODUCTION

The ever-increasing availability of data and the potential for data-driven techniques to improve teaching and learning are
driving a substantial shift of the educational environment. Predicting student success is a critical area of concentration in
this shift. When teachers are able to forecast with precision how a student will do in a given course, semester, or even
throughout their academic career, it allows them to make well-informed decisions and tailor instruction to maximize the
performance of their students.

This study explores the viability and efficacy of using data-driven techniques to forecast student achievement. Subjective
evaluations and prior academic achievement have been major components of the conventional method for identifying
pupils who are at risk of performing below expectations. Although useful, these techniques frequently lack the precision
and predictive ability required to proactively intervene and support struggling students.

By investigating the potential of data analytics and machine learning approaches in forecasting student performance, this
research aims to overcome these constraints. The main premise is that we may find important patterns and insights that
can be utilized to create reliable prediction models by examining a large collection of student data points. Teachers may
then use these models to identify pupils who are at danger of falling behind and provide interventions that are specifically
tailored to meet their needs.

The vital work of feature creation is where the project starts. Finding and gathering pertinent data points that could have
an impact on student performance is required for this. It is probable that the data sources will comprise a wide variety of
information, such as:

● Academic History: A student's past grades, performance on standardized tests, course preferences, and completion of
required courses can all offer important clues about their strengths and shortcomings in the classroom.
● Demographic Information: A student's academic experiences and results can occasionally be correlated with their age,
socioeconomic background, and parental education level. To prevent prejudices from being reinforced, it is essential to
recognize the limits of this type of data.
● Engagement Metrics: Information on involvement in class discussions, online activity within learning management
systems, attendance patterns, and assignment completion rates can provide important hints about a student's degree of
engagement as well as possible problems.
● Cognitive Abilities: The results of assessments or standardized tests intended to gauge pertinent cognitive abilities related to
the topic matter may offer further information.
● Learning Styles: Knowing a student's chosen learning style—visual, auditory, or kinesthetic—helps teachers create resources
and tactics that are specifically tailored to their needs.

It is crucial to remember that the precise features selected will probably change based on the educational setting, the particular
learning objectives being assessed, and data privacy concerns.

The research will move into the domain of modeling and assessment when the pertinent data points have been located and
gathered. At this point, data analysis and statistical modeling approaches are used to identify any hidden correlations between
the attributes and student performance. For this problem, a number of algorithms work well, including:

● Linear Regression: A linear relationship between the characteristics and the anticipated result (such as the final course grade)
is established via the frequently used approach known as linear regression. Even while linear regression is easy to understand
and apply, it might not be able to fully represent the complexity of real-world data.
● Random Forests: Using a random selection of characteristics and data points, this ensemble learning technique builds many
decision trees. Compared to single decision trees, the final forecast is more accurate and flexible since it is based on the
combined votes of these separate trees.
● Support Vector Machines (SVM): This method uses the largest margin to generate a hyperplane that divides data points into
distinct groups (e.g., high vs. low performers). SVMs are capable of handling non-linear data connections, making them
effective tools for classification problems.

Furthermore, curriculum design and resource allocation might benefit from the insights gained from these prediction models.
Teachers may adjust their curricula and methods of instruction to optimize learning outcomes for every student by having a
thorough understanding of the elements that contribute to student success
RELATED WORK:
There has been a notable increase in research over the last five years that examines data-driven methods for forecasting student
performance. Here, we examine some significant contributions made by scholars.

In a study by Chassignol et.al [1], it's noted that technology significantly influences both individuals and company activities,
including education. Although artificial intelligence won't completely replace traditional schooling, it's reshaping the
educational landscape.

Al-Sudani and Palaniappan [2] conducted research using a dataset of 470 students to predict their performance and classify their
degree into either good or basic. Among the students, 330 received excellent degrees, while 140 received basic degrees. The
dataset included attributes like institutional details, academic performance, demographics, psychological factors, and economic
background. They used a Neural Network with a 100-node forward network trained using the Levenberg-Marquardt learning
algorithm. The results showed an average classification accuracy of 83.7%, with sensitivity at 77.37%, specificity at 85.16%,
Positive Predictive Value at 94.04%, and Negative Predictive Value at 50.93%. This emphasizes the importance of training
hidden layers in Neural Networks.

In another study by Creswell et.al[3], they searched for relevant literature using keywords related to student profiles,
performance prediction, career guidance systems, and the role of machine learning (ML) and artificial intelligence (AI) in
education. They employed logical operators to match relevant literature for their topic.

Khalfan Al Mayahi and Dr. Mahmood Al-Bahri[4] conducted research using supervised learning to predict how well students
would perform at the University of Nizwa. They analyzed data from 530 students spanning the academic years 2017/2018 to
2018/2019. The dataset included student details like names, IDs, and grades from previous schools in subjects like Mathematics,
Arabic, Social Studies, and English. They developed a highly accurate model using the Pandas library, achieving an 87%
accuracy rate. This model predicted whether students would pass or fail exams based on their academic backgrounds. Among
the various machine learning techniques they tried, Support Vector Classification and Elastic Net were the most effective.
Shah Hussain and Muhammad Qasim Khan[5] highlighted how technology can benefit education, focusing on improved
communication between students and instructors, access to diverse data, identification of disengaged students, and facilitating
feedback.

Marbouti et.al[6] aimed to identify students at risk of academic struggle using data from 1650 first-year engineering students at
a Midwestern US university during 2013-2014. They divided the data for training, comparison, and testing. Among the models
tested, the Naive Bayes Classifier and Ensemble model performed best in predicting student outcomes. Although less than 10%
of students failed the course, accurately identifying them was crucial. The K-Nearest Neighbor algorithm correctly identified
94.9% of all students, while the Naive Bayes Classifier achieved an 86.2% accuracy in identifying at-risk students.

kiselev et.al[7]The dataset obtained from Digital Human comprises 1252 social network users aged 18-22, with 65% females
and 35% males, who completed a 30-question survey with weights ranging from 1 to 10. The study identifies key factors
(family background, culture, gender, and social conditions) affecting career development, and evaluates career orientation using
AUC-ROC with 5-fold cross-validation. The authors highlight that AI and machine learning can assist individuals and
counselors in navigating a constantly evolving job market and personalized career paths.

Another study D. Kabakchieva [8] focused on predicting student performance based on preliminary data from high school. They
utilized classification methods to identify which students would excel academically. Data were collected from the University of
National and World Economy (UNWE), stored in two datasets: administrative and faculty. These datasets included personal
information, high school results, entrance exam scores, and final classification data for 10,330 students over three years (2007-
2009). Variables such as age at admission and evaluation criteria were considered. The authors transformed some numerical
variables into nominal for better insights. Various data mining techniques were employed, with the decision tree classifier (J48)
showing the highest accuracy, followed by JRip and kNN classifiers. Despite efforts, the average accuracy remained at 70%,
suggesting high error rates and unreliable predictions.

Saa et.al[9] In this paper , The United Arab Emirates (UAE) compiled a dataset spanning five years (2013-2018), containing
records of 56,000 students with 34 variables. These students represent 100 different countries, reflecting a diverse background.
The dataset includes information on students' origin, course details, personal information, and early performance indicators. Key
factors such as origin school, student program, nationality, and gender were identified as crucial in determining academic
success. The Random Forest (RF) algorithm emerged as the most effective tool for predicting overall student performance.
Furthermore, the study highlighted four major categories influencing student performance: demographics, student
characteristics, educational changes, and early performance indicators.
While Arsad et.al[10]data from three student cohorts (2005-2007) comprising 391 students and 7 courses were analyzed. The
dataset included gender, subjects, and grade point averages (GPAs). Researchers developed an Artificial Neural Network
(ANN) model to predict Cumulative Grade Point Average (CGPA) using initial semester outcomes from basic courses as
predictors. This model aids educational advisors in enhancing student academic achievement.

Tarika et.al[11] In this, researchers analyzed a dataset from Guelmim Oued Noun in central Morocco, comprising information
on 72,010 students enrolled between 2000 and 2015. The dataset included student grades across three levels: Common Core,
first year baccalaureate, and second year baccalaureate. They employed algorithms such as linear regression, Decision tree, and
Random Forest. Results showed that Random Forest outperformed other algorithms in predicting the baccalaureate average.

Anal Acharya, and Devadatta Sinha[12] utilized a dataset from Kolkata College's computer science program, spanning 6
semesters and sourced from questionnaires, literature, and faculty input. The dataset comprises 297 records from 2006-10 and
92 records from 2011-12, including attributes such as gender, religion, caste, attendance, academic achievements, income status,
and exam performance. The aim was to identify scholarship students and potential risks to others. Attributes like academic
achievements, income status, and exam performance influenced predictions. Gender, origin, urban or rural background, and
family size also played a role. Decision Trees, Bayesian Networks, ANN, and SVM algorithms were employed, with Decision
Tree (C4.5) proving most effective.

Shahiria et.al[13] reviews literature to pinpoint crucial factors in forecasting student performance, emphasizing the importance
of internal assignments and cumulative grade points. Internal assignments encompass tasks, quizzes, and participation.
Demographics, notably gender, are also explored, with females often showing higher success rates and discipline. The study
employs classification, regression, and categorization methods, with classification being the primary approach. Techniques
utilized include decision trees, artificial neural networks, naive Bayes, k-nearest neighbor, and support vector machines.

Sotiris Kotsiantis et al. [14] propose a method to assess students' performance in a computer science course. They collected data
from 498 students enrolled in INF10 class at the Hellenic Open University during the 2000-2001 academic years. The dataset
includes demographic variables like age, gender, and domicile. Evaluation forms were used for the Tutor attribute, specifically
if a student received 100 or more hours of instruction in informatics. Six supervised algorithms were compared, with the Naive
Bayes algorithm achieving the highest accuracy at 72.48% and overall sensitivity at 78.00%. Naive Bayes was also deemed the
easiest algorithm to implement.

Athanasios S. Drigas and Rodi-Eleni Ioannidou[15],The methods of artificial intelligence (AI) and machine learning (ML)
techniques are now regarded as among the most influential implementations for specific educational needs. The main purpose of
such technologies is to improve how the young generation communicate with their surroundings, and also to encourage and
enforce education processes and improve their general lifestyle .

These are but a handful of the many studies on student performance prediction being carried out in the US and India. This
succinct summary highlights the variety of methods and the expanding corpus of research in this area. Our study can help create
even more reliable and efficient prediction models to enhance educational results for all students by expanding on our previous
work.

Methodology:

We will examine three fundamental approaches, including their features and how
the prediction models were constructed using them.

1. Feature Engineering: Constructing the Foundation

This project's cornerstone is feature engineering. It entails painstakingly locating, gathering, and preparing pertinent data
points (features) that might have an impact on student performance. The information will come from institutional
databases and a range of educational technology platforms, including:
● Academic History: To provide a complete picture of a student's academic past, information on
their prior grades, test results, course preferences, and performance in required courses will be
gathered.
● Engagement Metrics: Information on involvement in class discussions, attendance, online activity
in learning management systems, and assignment completion rates will all be gathered to
determine the degree of student engagement.
● Demographic Information: In order to prevent prejudices from being reinforced, this data will be
handled extremely sensitively and utilized sparingly. Parents' ages and educational attainment may
be taken into account, but only if they are anonymised and can be shown to be pertinent to the
particular educational setting.

To deal with missing values, outliers, and inconsistencies in the data, preprocessing and data cleaning procedures will be
used. The most influential elements for model construction will be found using feature selection techniques like
 correlation analysis and feature importance scores.

2. Unveiling Patterns: Machine Learning Techniques

Upon finalizing the feature collection, the research will utilize machine learning techniques to unearth the intricate
connections between these variables and student performance. The following three well-known methods will be
examined:
● Random Forests: Because of its stability and capacity for handling huge datasets, this ensemble
learning technique will be used. A random subset of characteristics and data points will be used to
train each of the several decision trees that will be built. When compared to single decision trees,
the final prediction will be more accurate and flexible since it will be based on the combined votes
of these separate trees.
● Support Vector Machines (SVM): This approach is highly effective in classification tasks and
will be taken into consideration in situations where it is necessary to classify student performance
(e.g., high vs. low performers). The hyperplane that divides data points from various classes with
the greatest margin will be made using SVMs.
● Logistic Regression: This technique will be investigated for certain scenarios in which a student's
performance is anticipated as a probability, such as the possibility of passing a course. Using this
method, a mathematical link is established between the traits and the likelihood of a particular
result.

The best method will be chosen based on a number of criteria, including the model's interpretability requirements, the
desired output (for example, predicting grades as opposed to identifying at-risk children), and the characteristics of the
data.

Student's

Pre-

Training Testing

Applying regression

Prediction

Evaluation

Display

Figure 1: Research Workflow Block diagram

3. Evaluating the Crystal Ball: Model Validation

Ensuring prediction models are accurate and generalizable is just as critical as developing robust
models. In order to determine if the selected algorithms are successful, this research will make use
of strict model assessment methodologies. Cross-validation is a crucial technique. Here, the data
will be split into two sets: a testing set for assessing the model's performance on untested data, and a
training set for building the model. By doing this, overfitting is reduced and the model's ability to
adapt well to fresh student data is guaranteed.

In addition, performance measures including recall, accuracy, precision, and F1-score will be computed to evaluate how
well the model predicts student performance. We can pinpoint areas for development and fine-tune the model for the best
outcomes by examining these indicators.

This research aims to build dependable and efficient models for forecasting student performance by putting these three
fundamental strategies into practice: feature engineering, machine learning algorithms, and model assessment. With the
use of these models, teachers will be able to tailor lessons and intervene proactively, which will eventually result in a
more encouraging and productive learning environment for all students.

Results and Discussion:

This section explores the outcomes of using the described methods to forecast student performance. At this point,
particular statistics and numbers cannot be shared due to the continuing nature of the research. But, once the project is
over, we may talk about the expected results, graphical representations, and assessment techniques that will be used.

Expected Results:
A collection of prediction models with a high degree of accuracy in projecting student performance will be the main
product of this research. Various machine learning techniques will probably be used by these models, depending on the
particular job and the properties of the data. For example, a Support Vector Machine may be used to categorize students
as strong achievers vs low performers, and a Random Forest model can be used to predict a student's final course grade.

Evaluation Metrics:
The effectiveness of the developed models will be assessed using various evaluation metrics. Here are some key
examples presented in a table for better understanding:

Table 1
SELECTING 9 CLASS LABELS REGARDING TO STUDENTS' GRADES
Class Grade Student Percentage

1 0.0 2 0.9%

2 0.5 0 0.0%

3 1.0 10 4.4%

4 1.5 28 12.4%

5 2.0 23 10.1%

6 2.5 43 18.9%

7 3.0 52 22.9%

8 3.5 41 18.0%

9 4.0 28 12.4%

Table 2
SELECTING 3 CLASS LABELS REGARDING TO STUDENTS' GRADES
Class Grade Student Percentage

High Grade >= 3.5 69 30.40%

Middle 2.0 < Grade < 3.5 95 41.80%

 Low Grade <= 2.0 63 27.80%

Table 3
SELECTING 2 CLASS LABELS REGARDING TO STUDENTS' GRADES
Class Grade Student Percentage

Passed Grade > 2.0 164 72.2%

Failed Grade <=2.0 63 27.80%

Graphical Representations
To help with interpretation and display the data, the project will make use of a variety of graphical representations. Here
are a few possible instances:
1. Line Charts: Line charts may be used to show how a certain attribute, like average attendance, relates to student
performance indicators, such as final grade.
2. Confusion Chart: The performance of a classification model is graphically represented by this table-like
depiction. For each class, it shows the quantity of cases that are properly and wrongly categorized.
3. Feature Importance: These charts help illustrate the relative weights assigned to various features in the model's
predictions. This makes it easier to determine which characteristics have the most effects on how well students
do.

Mathematical Formulas
Here are some often used examples, while the exact formulae utilized will depend on the machine learning methods
selected:

1. Random Forest: This technique blends forecasts from several decision trees. Every decision tree divides the
data at each node using a variant of the following formula:

Information Gain(Split Attribute) = Entropy(Parent) - Σ ( |Child i| / |Parent| ) * Entropy(Child i)

This formula determines how much information is gained by dividing the data according to a specific property.
For the split, the attribute with the most information gain is selected.

2. Support Vector Machine (SVM): The goal of a support vector machine (SVM) is to identify a hyperplane that
optimizes the margin between data points from various classifications. The hyperplane equation can be written
as follows:

w^T * x + b = 0

In this case, b is the bias term, x is a data point, and w is a weight vector. The SVM method looks for values of
w and b that maximize the margin between the nearest data points from each class and the hyperplane.

Discussion
The expected outcomes will undergo a thorough analysis to see how well the selected approach predicts student
achievement. The following topics will be covered in detail:

● Accuracy and Generalizability of the Models: To evaluate the model's capacity to accurately forecast student
performance, the obtained accuracy scores, precision, recall, and F1-scores will be examined. We shall assess
generalizability by gauging the model's output on hypothetical data.
● Effects of Various Features: Which characteristics have the most effects on student performance will be
determined by analyzing feature importance plots. This can help educational institutions develop focused
interventions and resource allocation plans.
● Restrictions & Upcoming Projects: The project's shortcomings, such as possible data biases or the difficulties
 of generalizing models to other educational environments, will be discussed. There will be an outline of future
research areas that will look at how to handle ethical issues, add more data sources, and increase model
accuracy.

This part will offer a thorough grasp of the project's findings and their implications for educational practice by clearly and
succinctly presenting the data, supplemented where necessary by pertinent graphical representations and mathematical
formulas.

Conclusion
The goal of this study is to investigate the viability and efficacy of data-driven approaches for forecasting student
achievement. Through the application of feature engineering, machine learning algorithms, and rigorous assessment
approaches, this study aims to create reliable models that enable educators to tailor learning experiences and intervene
proactively.

The expected outcomes have great potential to revolutionize schooling. When teachers are able to forecast student
performance accurately, they can:
Determine which kids are at danger of falling behind. In order to minimize academic difficulties and promote student
achievement, early intervention can be very important.
● Customize instructional strategies: Teachers can enhance learning outcomes for all students by customizing
their teaching techniques based on their awareness of each student's unique requirements and learning
preferences.
● Distribute resources wisely: Insights derived from data may guide the distribution of resources, guaranteeing
that the students who require the most focused assistance receive it.
● Enhance the creation of curricula: A curriculum that is more interesting and successful may be developed
with guidance from an understanding of the elements impacting student performance.

But it's important to recognize that using educational data raises ethical questions. Careful attention is needed for privacy
issues, model biases, and the appropriate use of these tools.

Future Goals and Directions

Building on the groundwork this study provided, further investigation can focus on a number of areas:

● Adding More Data Sources: Investigating the addition of non-conventional data sources, including sentiment
analysis from student conversations or social network analysis, might improve the models' ability to predict
outcomes.
● Managing Algorithmic Bias: It is critical to create strategies to lessen bias in the algorithms and make sure that
every student receives results that are just and equal.
● Explainable AI: To enable educators to make well-informed judgments, models that can both forecast and offer
interpretable insights into the factors driving student performance must be developed.
Studies that monitor student performance over time are known as longitudinal studies, and they can offer important
insights into the efficacy of treatments and the changing circumstances that affect academic success.

We have only just begun our quest to use data-driven prediction to transform education. It is possible to fully fulfill the
promise of this strategy to enhance educational results for every student by promoting collaboration among researchers,
educators, and policymakers. Establishing a learning environment where each student has the chance to realize their full
potential is the ultimate objective. This initiative is a first step toward accomplishing this challenging but crucial goal.

References

[1] Maud Chassignol, Aleksandr Khoroshavin, Alexandra Klimova, and Anna Bilyatdinova, "Artificial
Intelligence trends in education: a narrative overview," Procedia Computer Science, vol. 136, pp. 16-- 24,
2018.
[2] Sahar Al-Sudani, and Ramaswamy Palaniappan, "Predicting students’ final degree classification using an
extended profile," Education and Information Technologies, vol. 24, no. 4, p. 2357– 2369 , 2019.
[3] J. W. Creswell, Educational Research: Planning, Conducting and Evaluation Quantitative and Qualitative
Research, 4th Ed., Boston, Massachusetts, USA: Pearson, 2012.
[4] Khalfan Al Mayahi, and Dr. Mahmood Al-Bahri, "Machine learning based predicting student academic
success," in 2020 12th International Congress on Ultra Modern Telecommunications and Control Systems
and Workshops (ICUMT), Brno, Czech Republic, IEEE, 2020, pp. 264--268.
[5]Shah Hussain, and Muhammad Qasim Khan, "Student-performulator: predicting students’ academic
performance at secondary and intermediate level using machine learning," Annals of data science, pp. 1--
19, 2021.
[6]Farshid Marbouti, Heidi A. Diefes-Dux, and Krishna Madhavan, "Models for early prediction of at-risk
students in a course using standards-based grading," Computers & Education, vol. 103, pp. 1-- 15, 2016.
[7] Pavel Kiselev, Boris Kiselev, Valeriya Matsuta, Artem Feshchenko, Irina Bogdanovskaya, and Alexandra
Kosheleva, "Career guidance based on machine learning: social networks in professional identity
construction," Procedia Computer Science, vol. 169, pp. 158--163, 2020.
[8] D. Kabakchieva, "Predicting student performance by using data mining methods for classification,"
Cybernetics and information technologies, vol. 13, no. 1, pp. 61--72, 2013.
[9]Amjad Abu Saa, Mostafa Al-Emran, and Khaled Shaalan, "Mining student information system records to
predict students’ academic performance," in International conference on advanced machine learning
technologies and applications, Springer International Publishing, 2019,
[10] Pauziah Mohd Arsad, Norlida Buniyamin, and Jamalul-lail Ab Manan, "Neural network model to predict
electrical students' academic performance," in 2012 4th International Congress on Engineering Education,
Georgetown, Malaysia, IEEE, 2012
[11] Ahajjam Tarika, Haidar Aissab, and Farhaoui Yousef, "Artificial intelligence and machine learning to
predict student performance during the COVID-19," Procedia Computer Science, vol. 184, pp. 835--840,
2021.
[12] Anal Acharya, and Devadatta Sinha, "Early prediction of students performance using machine learning
techniques," International Journal of Computer Applications, vol. 107, no. 1, pp. 37--43, 2014.
[13] Amirah Mohamed Shahiria, Wahidah Husaina, and Nur’aini Abdul Rashid, "A Review on Predicting
Student’s Performance using Data Mining Techniques," Procedia Computer Science, vol. 72, pp. 414-- 422,
2015.
[14] Sotiris Kotsiantis, Christos Pierrakeas, and Panagiotis Pintelas, "Predicting students' performance in
distance learning using machine learning techniques," Applied Artificial Intelligence, vol. 18, no. 5, pp.
411--426, 2010.
[15]Athanasios S. Drigas and Rodi-Eleni Ioannidou, "Artificial intelligence in special education: A decade
review," International Journal of Engineering Education, vol. 28, p. 1366, 2012.