Received 20 November 2024, accepted 6 December 2024, date of publication 13 December 2024,

date of current version 24 December 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3516816

Early Prediction of Student Performance in

Face-to-Face Education Environments: A Hybrid
Deep Learning Approach With XAI Techniques
AHMET KALA 1,2 , ORHAN TORKUL 1, TUGBA TUNACAN YILDIZ 3,

AND IHSAN HAKAN SELVI 4

1 Industrial
Engineering Department, Sakarya University, 54050 Sakarya, Türkiye
2 Department of Information Technologies, Sakarya University of Applied Sciences, 54050 Sakarya, Türkiye
3 Industrial
Engineering Department, Bolu Abant Izzet Baysal University, 14030 Bolu, Türkiye
4 Department of Information Systems Engineering, Sakarya University, 54050 Sakarya, Türkiye

Corresponding author: Ahmet Kala ([email protected])

ABSTRACT A community is only as strong as its weakest link; this principle also applies to student
communities in the educational field. The quality of learning achieved by students in a course is directly
related to the performance of the weakest student in that course. Therefore, a high number of students failing a
course and the necessity of repeating it are undesirable in terms of learning quality. This study aims to predict
students’ performance early during their coursework to identify those at risk of failing, thus improving the
quality of the course and determining the necessary resources to achieve this goal. To this end, we proposed
a conceptual model based on a hybrid method combining Particle Swarm Optimization (PSO) and Deep
Neural Networks (DNN). To evaluate the classification performance of the model, comparisons were made
with classical machine learning and deep learning models. SHAP (SHapley Additive exPlanations) and
LIME (Local Interpretable Model-agnostic Explanations) methods were used to determine the contribution
of different features to the model’s predictions. Additionally, to assess the generalizability and applicability
of the model, the widely used xAPI-Edu-Data dataset, which covers various courses, was employed, and the
accuracy results of the model were compared with early prediction studies published in the literature. As a
result, it was found that our prediction model performed 6% better than the classical models and achieved
better results than most of the models, except for two models in the literature with similar results. Moreover,
important performance features that can be used to evaluate students earlier in the course were identified.

INDEX TERMS Early prediction of student performance, deep learning (DL), explainable artificial
intelligence, local interpretable model-agnostic explanations (LIME), and shapley additive explanations
(SHAP).

I. INTRODUCTION tools, and by whom is planned with specific objectives in

Education is a tool that facilitates the support and develop- mind. This teaching plan and method are implemented with
ment of individuals in specific areas to enhance society’s minimal changes throughout their education. The quality of
creativity and productivity. Conversely, teaching refers to the the education and teaching plan received can be measured
planned, programmed, and professional instruction provided in various ways depending on the stage the individual is at
in schools by qualified personnel. When an individual begins in the process. There are notable differences between the
their education, whether at the secondary or higher level, secondary and higher education stages, particularly in Turkey.
what will be taught, when, where, for how long, with what The effectiveness of education at the end of the 4+4 education
system can be assessed through the high school placement
The associate editor coordinating the review of this manuscript and exam, which measures how effectiveprimary and secondary
approving it for publication was Li Zhang . school education has been. The outcome is the rate at which
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 191635
 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

students are accepted into science high schools, Anatolian learning models can pose challenges in building trust among
high schools, and traditional high schools in Türkiye. The educators. Interpreting and trusting model predictions can
success of high school education is then measured through become difficult without sufficient transparency, which may
the higher education entrance exam. The outcome here is hinder the adoption of these techniques in educational
the number of students who succeed in entering university contexts [2].
and their scores in mathematics, science, Turkish, and In this study, we focused on early prediction of the
social sciences. The success of the university’s education academic performance of students receiving face-to-face
and teaching plan can be measured by the number of education at the university. We proposed a conceptual
graduates and how many of these graduates contribute to the model for this purpose. The evaluation criteria used in our
workforce. All these measurements are general assessments. model include the academic skills acquired during high
For instance, a low number of graduates from a specific school, the results of the higher education entrance exam,
program at a university does not necessarily indicate a university academic performance data, and demographic
low-quality education and teaching plan. However, it does characteristics. Our output value is the student’s pass-fail
indicate the need for further investigation. It can be predicted status in the course. We used a hybrid deep learning model
that measuring the performance of courses offered within the with particle swarm optimization and deep neural networks
education-teaching plan of the relevant program instead of (PSO-DNN) to create the early prediction model of academic
the number of graduates at the final stage would increase performance using these features. Since the performance in
the number of graduates. This can be achieved by assessing the higher education entrance exam is particularly important
students’ academic performance early on. for building this model, we limited the study to a single
The early prediction of students’ academic performance institution and engineering program to easily obtain these
will facilitate assistance for students who are performing values. To determine student performance using the variables
below the class average or struggling to keep up with the in the model, we focused on the ‘‘Programming’’ course,
learning process. Students’ academic performance varies which is common across different programs, taught by
depending on the academic knowledge they have acquired different instructors, and has a low course success rate due
during their education or their individual needs. Especially to the high number of failing students. We measured the
for university students, their performance in different courses general applicability of our model using the xAPI-Edu-Data
may vary depending on the skills they acquired during dataset and compared it with five previous studies that used
their previous education. For example, if the academic this dataset. To explain the results of the proposed hybrid
performance of an engineering student is to be assessed, the PSO-DNN model and understand why it outperforms other
successes they demonstrated in subjects like mathematics, models, we used SHAP and LIME explainable artificial
physics, and chemistry during secondary education can intelligence (XAI) methods. Fig. 1 illustrates the interrelation
be used as indicators. This is because many university among deep learning, explainable artificial intelligence, and
engineering education programs emphasize the importance of early prediction of student performance, which constitute the
subjects like mathematics, physics, and chemistry throughout primary focus of this study.
high school, and the student’s success in these areas in the
higher education entrance exam leads to their admission to
the engineering program. In this case, a student’s academic
performance can be measured using factors such as the
numerical placement score obtained in the higher education
entrance exam, the number of correct and incorrect answers
in subjects like mathematics, physics, chemistry, and even
history and geography from social sciences, and the high
school diploma average upon graduation. However, these
factors have not been used in the early prediction of student
performance in the literature. At the same time, studies
focusing on early prediction in the context of face-to-
face education are quite limited. This is because previous
studies have primarily focused on online education and
measured performance based on the effort shown during
the course. In other words, performance is determined not
FIGURE 1. The relationship between deep learning, explainable artificial
during the course but based on the activities demonstrated intelligence, and early student performance prediction.
during the course. While the deep learning models used in
these studies have the potential to reveal complex patterns This study is organized as follows: Section II will present
in educational data, their effectiveness often depends on the literature on deep learning and explainable artificial
hyperparameter settings, and the quality and quantity of intelligence (XAI) for the early prediction of students’
the data used [1]. Additionally, the complexity of deep performance. Chapter 3 will outline the methodology applied

191636 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

in this research, detailing the data collection process, model BLSTM and CRF [7]. Al-azazi and Ghurab introduced a
development, and evaluation techniques. Sections IV and V hybrid model merging Artificial Neural Network (ANN)
will present the findings and discussions based on applying and Long Short-Term Memory (LSTM) models to predict
the hybrid deep learning model incorporating XAI for early students’ performance on a day-wise multi-class basis [9].
student performance prediction. Finally, Chapter 6 will Lastly, Du et al. proposed the LVAEPre framework, combin-
conclude the study by summarizing the key results and ing LVAE with deep neural networks to address imbalanced
implications for educational practice and outlining potential dataset distributions and offer early warnings for at-risk
avenues for future research in educational data mining. students [8].
Studies employing EDM for early student performance
prediction have yielded valuable insights, yet further research
II. LITERATURE REVIEW is warranted. There is a need to delve into the predictive power
A. DEEP LEARNING FOR EARLY PREDICTION OF of different features, develop interpretable models, explore
STUDENT PERFORMANCE the potential of deep learning techniques, and address data-
Educational Data Mining (EDM) has indeed experienced related limitations. Moreover, extending the research scope
a surge in studies focusing on early predicting student to encompass diverse educational levels and countries would
performance since 2017, with researchers from various enhance the generalizability of findings.
countries making significant contributions. This focus has
primarily centered on e-learning platforms like MOOCs,
VLEs, LMSs, and ITSs. While the literature predominantly B. EXPLAINABLE ARTIFICIAL INTELLIGENCE FOR EARLY
covers undergraduate-level higher education students, more PREDICTION OF STUDENT PERFORMANCE
research is needed concerning traditional face-to-face and The application of Explainable Artificial Intelligence (XAI)
hybrid educational settings. in Educational Data Mining (EDM) has gained significant
Datasets commonly used in these studies include student attention recently due to its ability to provide insights into
demographic information, assessment results, activity data, the decision-making process of complex models, particularly
daily LMS data, and behavioral data. Advanced machine in predicting student performance. Several studies have
learning and deep learning techniques such as LSTM, explored using XAI to enhance the interpretability and
CNN, DFFNN, BLSTM, and Deep ANN have been heavily transparency of predictive models in educational settings.
adopted, with LSTM being the most prevalent. Hybrid DL Pereira et al. investigated the interpretation of a black-box
techniques, particularly CNN-LSTM models [3], [4], [5], [6], predictive model to explain individual and collective
have gained traction for effectively handling complex data programming students’ behavior. The study focused on
distributions. leveraging statistics to interpret the predictive model, shed-
Recent studies have highlighted the efficacy of hybrid ding light on the factors influencing student performance
DL techniques in early student performance prediction. in programming courses [10]. Chen et al. [5] proposed
Innovative models like Sequential Prediction Based on an intelligent framework for explainable student perfor-
Deep Network (SPDN) [6], BLSTM + Condition Random mance prediction (ESPP) in virtual learning environments.
Field (CRF) [7], DNN Integrated Framework Based on By incorporating deep XAI techniques, the study aimed
Latent Variational Autoencoder (LVAEPre) [8], and LSTM- to provide interpretable predictions of student performance
ANNs [9] have been developed. These models leverage weekly, enhancing the understanding of factors contributing
combinations of deep learning architectures to analyze to academic success [5]. Singh et al. [11] developed a
diverse data sources and accurately predict student outcomes. framework for suggesting corrective actions to support
In these studies, Chen et al. proposed a hybrid intelligent students at risk of low performance based on an experimental
framework combining CNN and LSTM models to enhance study of college students using an explainable machine
prediction accuracy in Virtual Learning Environments (VLE) learning model. The research emphasized the importance of
[5]. Li et al. introduced the SPDN model, integrating actionable insights derived from XAI models to intervene and
CNN and LSTM DL models to forecast students’ course improve student outcomes [11]. Adnan et al. [12] introduced
performance using online learning records and blog data [6]. an XAI model for the earliest possible global and local
Li et al. presented an end-to-end hybrid DL model merging interpretation of students’ performance in virtual learning
CNN and LSTM models to extract features from multi-source environments. By leveraging XAI, the study aimed to provide
heterogeneous behavioral data, showcasing its effectiveness timely insights into student performance throughout the
in educational contexts [3]. Venkatachalam and Sivanraju course, enabling educators to offer targeted support when
developed the SADDL framework, incorporating LSTM, needed [12]. Embarak and Hawarna [13] presented an
CNN, and MLP modules, which outperforms other models automated AI-driven system for the early detection of at-risk
in predicting student outcomes by leveraging physiological, students. The study highlighted the role of XAI in identifying
academic, and demographic data [4]. students who may be struggling academically, allowing
Uliyan et al. reported high accuracy in assessing students’ for timely interventions to prevent adverse outcomes [13].
retention status through a hybrid DL technique comprising Jang et al. [14] focused on practical early prediction of

VOLUME 12, 2024 191637

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

students’ performance using machine learning and XAI. TABLE 1. The features of the dataset.
By integrating XAI techniques into predictive models, the
study aimed to provide transparent and interpretable pre-
dictions to support educators in making informed decisions
about student progress [14]. Abdalkareem and Min-Allah
[15] explored explainable models for predicting academic
pathways for high school students in Saudi Arabia. The
study emphasized the importance of XAI in guiding students
toward suitable academic trajectories by providing transpar-
ent and understandable predictions [15].
Overall, integrating XAI in EDM offers promising
opportunities to enhance the transparency, interpretability,
and effectiveness of predictive models for early student
performance prediction, ultimately supporting educators in
improving educational outcomes.

III. METHODOLOGY
This section introduces an innovative approach for early
predicting student performance in face-to-face educational
environments by employing a hybrid deep learning model
that integrates explainable artificial intelligence within the
domain of Educational Data Mining (EDM). Fig. 2 visually
illustrates the workflow adopted in this study. The method-
ology comprises a series of crucial stages: data collection,
data preprocessing, feature selection, data splitting, model
development, evaluation, and model interpretation. Each of
these stages will be expounded upon in the subsequent
sections.

A. DATA COLLECTION
The dataset utilized in this study encompasses information
gathered from the Programming course, a fundamental
compulsory course within engineering programs, conducted
in the conventional face-to-face educational environment of
a state university in Turkey during 2019, 2021, and 2022.
The dataset comprises a comprehensive set of 55 features,
including three nominal, 51 numerical, and one numerical
output value. A detailed enumeration of these features is
provided in Table 1. xAPI-Edu-Data consists of 480 data instances and 16 fea-
The dataset comprises information from three distinct tures. The features are categorized into three sets: Demo-
programs, instructed by seven instructors, and involving graphic, Academic, and Behavioral features. The target
1268 students. The student distribution reveals that 11.1% are variable is the ‘‘Class’’ feature, where students’ grades can
female, while 88.9% are male. Among these students, 49.4% be classified into three categories: Low, Medium, and High.
are enrolled in the daytime program, with the remaining
50.6% in the evening program. Regarding programs, 50.5% B. DATA PREPROCESSING
of the enrolled students are pursuing Mechatronics Engineer- At this stage, data from various sources were merged, and
ing, 30.4% are in Mechanical Engineering, and 19.1% are in features with single values and those with over 80% missing
Electrical-Electronics Engineering. data were removed. Additionally, for features with less than
Additionally, to evaluate the applicability of the proposed 80% missing data, the missing values were imputed using
model to different courses or educational institutions and to the mean of the respective column. Categorical data were
address dataset biases, the xAPI-Edu-Data general dataset, converted into numerical values using LabelEncoder methods
which is widely used in academic research and covers to make them suitable for models. Numerical data were
various courses, has been utilized. This dataset, obtained normalized using mean and standard deviation to ensure
from https://www.kaggle.com/datasets/aljarah/xAPI-Edu- proportional scaling of the data distribution. The normalized
Data, comprises educational data from the online Learning values were then scaled to a range of 0-1 using min-max
Management System Kalboard 360, used worldwide. scaling.

191638 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

FIGURE 2. Workflow of the hybrid deep learning model with XAI for early prediction of student
performance in face-to-face education.

C. FEATURE SELECTION
Feature selection is a critical process that involves
choosing essential features by eliminating irrelevant or
low-information features before model training.
In this study, the Pearson correlation filter method was
employed to identify unnecessary or highly similar features
by analyzing linear relationships between features. Feature
pairs with high correlation are detected using a threshold
value (0.95). More informative features are selected, while
others are removed. This method reduces the complexity of
the model, improves performance, and reduces the risk of
overfitting.
FIGURE 3. The number of students passing and failing the course over
the years.
D. DATA SPLITTING
The dataset is divided into two subsets: training and testing using the GPA (Grade Point Average) was established at a
sets. The training set, which constitutes 70% of the data in minimum success value of 50. According to this threshold,
this case, is used to train the machine learning model on out of the 1268 students, 60% (757) failed the course, while
the patterns and relationships present in the data. The testing 40% (511) passed. The distribution of students who passed
set, comprising 30% of the data, is kept separate from the and failed the course across the years is depicted in Fig. 3.
training set and is used to evaluate the model’s performance
on unseen data. By splitting the data into training and testing 1) PARTICLE SWARM OPTIMIZATION (PSO)
sets, we can assess how well the model generalizes to new, Particle Swarm Optimization (PSO), is a widely used search
unseen data and identify potential issues such as overfitting method involving agents moving through a search space to
or underfitting. find the global minimum of an objective function [16], [17].
PSO is particularly effective in optimizing hyperparameters
E. MODEL DEVELOPMENT for neural networks. PSO has been applied to various neural
In this research, machine learning (ML) and deep learning network architectures, such as CNNs, LSTMs, and MLPs,
(DL) models were employed for binary classification to enhancing their performance [18], [19], [20]. This study used
predict students’ outcomes as either pass or fail. For binary it to optimize the hyperparameters of the deep neural network
classification, the threshold for identifying failing students model.

VOLUME 12, 2024 191639

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

2) DEEP NEURAL NETWORK (DNN) TABLE 2. The parameters used in the PSO-DNN model.

A deep neural network (DNN) is an artificial neural network

(ANN) that contains multiple hidden layers between the
input and output layers. These hidden layers allow DNNs
to learn hierarchical representations of data, making them
powerful tools for image and speech recognition, natural lan-
guage processing, and more. Hyperparameters are essential
parameters set before the learning process begins, controlling
the network’s behavior during training and significantly
influencing the model’s performance. Key hyperparameters
in deep learning include the learning rate, the number of
layers in the neural network, the number of neurons per layer,
activation functions, batch size, regularization techniques,
dropout, batch normalization, optimization algorithms, and
learning rate schedules [18].

3) PROPOSED MODEL (PSO-DNN)

The proposed Particle Swarm Optimization-based Deep
Neural Network model (PSO-DNN) utilizes the iterative and computational burden commonly associated with deep
optimization capabilities of PSO to enhance the training learning models.
process of deep neural networks. By optimizing the network’s We use a global-best particle swarm optimization
structure, weights, and hyperparameters, PSO aids in finding algorithm implemented in the PySwarms Python library [21].
near-optimal solutions for complex optimization problems. The parameters used in the model are detailed in Table 2.
The proposed PSO-DNN model architecture, illustrating the Additionally, the PSO process for optimizing DNN is
layers and neurons of the deep neural network (DNN) model, presented in Algorithm 1.
is presented in Fig. 4.
Algorithm 1 The PSO Process for Optimizing DNN
Require: α: number of particles in a swarm
Require: c: cognitive coefficients
Require: w: inertia weight
Require: bounds: a tuple of size two where the first entry is
the minimum and the second entry is the maximum
Require: V : initial velocity
Require: P: initial the nodes of double hidden layers
Require: pfit: initial individual best fitness value
Require: pibest: individual best value
Require: gfit: initial global best fitness value calculated by
DNN
Require: gbest: global best value
1: while gfit > 0 do
2: for i = 1 to α do
3: Calculate update: Vi ← wVi−1 + c1 rand1 (pibest −
FIGURE 4. The proposed PSO-DNN model.
Pi−1 ) + c2 rand2 (gbest − Pi−1 )
4: Calculate update: Pi ← Pi−1 + Vi
This model automatically adjusts the hyperparameters of
5: Calculate pfit(i) through DNN using Pi
a Deep Neural Network (DNN), unlike traditional Deep
6: if pfit(i) < pfit(i − 1) then
Neural Networks (DNNs) that are trained with fixed archi-
7: Update global best value: pibest ← Pi
tectures and hyperparameters—often determined through
8: if pfit(i) < gfit then
time-consuming trial and error processes. While existing
9: Update global best fitness value: gfit ← pfit(i)
PSO and DNN studies optimize structural parameters such
10: Update global best value: gbest ← pibest
as the number of hidden layers and neurons per layer, our
11: end if
model automates the optimization of not only structural
12: end if
parameters but also important DNN model hyperparame-
13: end for
ters such as activation functions, optimization algorithms,
14: end while
learning rates, training epoch numbers, and batch size. This
15: return gfit, gbest
automated hyperparameter tuning alleviates the complexity

191640 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

TABLE 3. The compared ML/DL models. • Kappa Statistic: Reflects the agreement between predic-
tions and actual labels beyond chance, with values closer
to 1 indicating stronger agreement.
• AUC (Area Under Curve): Evaluates the model’s ability
to differentiate between classes, especially in binary
classifications, with higher values indicating better
performance.
These metrics offer valuable insights into the efficacy and
reliability of the models in accurately predicting student
outcomes.

G. MODEL EXPLANATION
After completing the training and testing phases of the
model, widely used explainable artificial intelligence (XAI)
4) COMPARED DL AND ML MODELS techniques, namely Local Interpretable Model-agnostic
Explanations (LIME) and Shapley Additive exPlanations
The hybrid deep learning model proposed in this study was
(SHAP), were employed to interpret the model’s predictions.
compared DNN and against conventional machine learn-
ing models, including the k-nearest Neighbors Algorithm
(KNN), Logistic Regression (LR), Support Vector Machines 1) SHAP (SHAPLEY ADDITIVE exPlanations)
(SVM), Random Forests (RF), Decision Trees (DT), and SHAP is a technique within XAI that leverages Shapley
Artificial Neural Networks (ANN). The parameters of the values from game theory to explain and understand the
compared models are detailed in Table 3. predictions of complex models. SHAP values quantify the
contribution of each feature to a prediction, providing insights
into the role of features in shaping the model’s output.
5) HARDWARE AND SOFTWARE ENVIRONMENT USED FOR
By calculating Shapley values for each feature in a specific
MODEL TRAINING AND TESTING PROCESSES
order, SHAP determines feature weights and showcases how
Proposed PSO-DNN model and other comparative models individual features influence the model’s predictions. The
were trained and tested on a computer with the following cumulative sum of SHAP values offers a comprehensive view
specifications: 11th Gen Intel® Core™ i5-11400H processor of the factors impacting the model’s predictions, enhancing
(2.70 GHz), Windows 11 Pro operating system, 16 GB the interpretability of the model [22], [23]. This study used
RAM, NVIDIA GeForce GTX 1650 graphics card, and 64-bit the SHAP DeepExplainer tool in Python to interpret the
architecture, utilizing CPU resources. proposed hybrid deep learning model. The DeepExplainer
The classification models were developed in Python using tool in the SHAP Python package is handy for approximating
the Pandas, Keras, and Numpy libraries within Anaconda and aiding the interpretation of SHAP values in deep neural
Navigator. The training dataset was utilized to train both
network models [24].
the proposed Hybrid DL model and the compared ML/DL
models. Subsequently, all models were evaluated using a
2) LIME (LOCAL INTERPRETABLE MODEL-AGNOSTIC
separate test dataset.
EXPLANATIONS)
LIME is a model-agnostic XAI technique designed to
F. EVALUATION elucidate the decision-making processes of complex models
Evaluating deep learning models involves using various on specific instances. Its versatility allows it to be applied
parameters to gauge their performance accurately. In this across various model types, making model explanations
study, the evaluation process incorporates commonly used accessible to a broader audience. By providing meaningful
classification metrics, including: and understandable explanations, LIME aids in demystifying
• Accuracy: The ratio of correctly predicted instances to the functioning of black-box models, enhancing transparency
the total. It provides an overall performance snapshot but and interpretability [23], [25]. This study utilized the Python
may fall short with imbalanced datasets. package ‘‘lime’’ to implement the LIME-related functions.
• Precision: Measures the proportion of true positive
predictions among predicted positives, highlighting the IV. EXPERIMENTAL RESULTS
reliability of positive classifications. A. FEATURE SELECTION
• Recall (Sensitivity): Assesses the model’s ability to In the feature selection process, features with high correlation
identify actual positives, crucial in scenarios where values were identified based on a threshold of 0.95. The cor-
missing positive cases is costly. relation matrix for these features and the target variable GPA
• F1-Score: The harmonic mean of precision and recall, is presented in Fig. 5. Diploma Grade, High School Achieve-
balancing both metrics for imbalanced datasets. ment Score, Placement Score, SAY Placement Score, SAY

VOLUME 12, 2024 191641

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

FIGURE 5. The correlation matrix for features with high correlation values (threshold: 0.95) and the target variable GPA.

Placement Success Ranking Percentage, SAY Score, SAY TABLE 4. The best parameter combination of PSO-DNN models, based on
the number of hidden layers.
Success Ranking Percentage, TYT Placement Score, TYT
Placement Success Ranking Percentage, TYT Score, and
TYT Success Ranking Percentage were features that have
high correlation according to the specified threshold. Among
these features, Diploma Grade, Placement Score, TYT Score,
SAY Score, TYT Success Rank Percentage, and SAY Success
Rank Percentage were deemed unnecessary and removed due
to their low correlation with GPA. Following the feature
selection process, the dataset was reduced from 54 features
to 48 features.

B. THE BEST PARAMETER COMBINATION OF PSO-DNN

MODELS
learning (DL) and machine learning (ML) models. The
The initialization parameters of the PSO-DNN proposed in hybrid deep learning model proposed in this study, PSO-
this study are previously given in Table 2. The best parameter DNN, was compared to conventional machine learning and
combinations of the models created with these parameters, DNN DL models. The proposed PSO-DNN model and the
based on the number of hidden layers, are presented in compared models were evaluated by testing on unseen test
Table 4. According to Table 4, the best accuracy value was data. The classification performance metrics for each model
obtained with the 2nd model consisting of 3 hidden layers. are presented in Fig. 6.
According to Fig. 6, the PSO-DNN model achieved the
C. EARLY PREDICTION OF STUDENT PERFORMANCE highest accuracy (0.633), F1-score (0.561), precision (0.638),
The early prediction of student performance in face-to- and recall (0.633). Additionally, while its AUC (0.555) value
face educational settings at the beginning of the semester was moderate, the Kappa value (0.125) was relatively low,
was conducted using demographic, academic performance, indicating room for improvement in terms of agreement
high school, and higher education entrance data with deep with true classifications. The Random Forest (RF) model

191642 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

FIGURE 6. Evaluation metrics of ML and DL models.

followed PSO-DNN with competitive accuracy (0.627) and

relatively strong performance across other metrics such as
AUC (0.572) and Kappa (0.156). Similarly, the ANN model
showed balanced performance with an accuracy of 0.622 and
a slightly higher AUC (0.586), making it a strong contender
in terms of predictive capability. The Logistic Regression
(LR), KNN, and DNN models achieved moderate accuracy,
with values ranging from 0.562 to 0.614, showing varying
strengths in specific metrics. However, the Decision Tree
(DT) model exhibited the weakest performance across all
metrics, with the lowest accuracy (0.562) and Kappa value
(0.087). FIGURE 7. Confusion matrix of PSO-DNN and RF models.
The confusion matrix of the PSO-DNN and RF models are
shown in Fig. 7. The confusion matrix allows us to evaluate
the performance of two classification scenarios from different
perspectives. It provides a more detailed understanding of
which metrics are more critical and which classes each
model predicts better. The PSO-DNN model exhibits better
performance than the RF model. It provides higher precision
and recall for ‘‘fail’’ class predictions.
For the proposed PSO-DNN model, Fig. 8 (a) illustrates
the trade-off between false positives and false negatives
FIGURE 8. ROC curve and precision-recall for PSO-DNN model.
through the ROC-AUC curve. Likewise, Fig. 8 (b) depicts
the trade-off between precision and recall for predicting the
Pass class. When evaluated together, the proposed model D. THE RESULT OF THE GENERALIZABILITY OF THE
exhibits a high area under the Precision-Recall curve (AUC- PROPOSED MODEL
PR) and a high area under the ROC curve (AUC-ROC). This To assess the generalizability of the proposed PSO-DNN
indicates that the model performs well in correctly classifying model, the widely used general dataset in academic research,
rare classes and achieving good discrimination in a balanced xAPI-Edu-Data, which classifies students’ grades into three
classification problem. Therefore, the overall performance of categories: Low, Medium, and High, was utilized. The
this model is high. results of the three-class classification conducted using the

VOLUME 12, 2024 191643

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

TABLE 5. The best parameter combination of PSO-DNN models, based on Achievement Score.’’ ‘‘Course Repeat Count’’ was the
the number of hidden layers using xAPI-Edu-Data dataset.
third most impactful feature on the prediction, while
‘‘TYT Placement Score’’ was the fourth most compelling
feature. Other significant contributing features include the
‘‘TYT Placement Success Ranking Percentage,’’ ‘‘Num-
ber of Incorrect Answers in Mathematics,’’ ‘‘Number of
Incorrect Answers in Social Sciences,’’ ‘‘SAY Placement
Success Ranking Percentage,’’ ‘‘Number of Courses Taken,’’
‘‘EA Score,’’ ‘‘Number of Correct Answers in Chemistry,’’
‘‘Number of Incorrect Answers in Science 1,’’ ‘‘Number of
Correct Answers in Geography 1,’’ ‘‘SAY Placement Score,’’
‘‘Number of Incorrect Answers in Physics,’’ ‘‘Preference
Order,’’’ ‘‘Number of Incorrect Answers in History 1,’’
TABLE 6. The results of previous academic studies and proposed model
‘‘Number of Incorrect Answers in Biology,’’ ‘‘Number
on the xAPI-Edu-Data. of Correct Answers in Mathematics,’’ and ‘‘Number of
Semesters in Preparatory Program.’’

proposed model and xAPI-Edu-Data dataset are presented in

Table 5. The highest accuracy value of 0.806 was achieved by
Model 2.
The models employed in previous studies using the
xAPIEdu-Data dataset for three-class classification, along
with their corresponding accuracy values, are presented
in Table 6. According to Table 6, the highest accuracy
(0.840) was achieved with the DNN model in the study
by Vijayalakshmi and Venkatachalapathy [26]. Despite the
constraints of swarm size, maximum iterations, and the
number of hidden layers, the proposed model demonstrates
promising results, achieving accuracy close to the two
highest-performing models and outperforming the others.

E. EXPLAINABLE ARTIFICIAL INTELLIGENCE RESULTS OF

THE MODEL
This study utilized SHAP and LIME XAI methods to FIGURE 9. SHAP global bar plots.
explain the proposed hybrid PSO-DNN model’s results and
understand why it outperformed other models.
2) SHAP LOCAL EXPLANATION RESULTS
1) RESULT OF SHAP GLOBAL EXPLANATION SHAP also provides local explanations for each student’s
Fig. 9 displays the SHAP graph of the PSO-DNN model for classification. For the local explanation, two students from
all test data. The top 20 features with the highest impact on the ‘‘fail’’ class were selected: one whose most impactful
the prediction are sorted from highest to lowest. This graph features align with the global features, and another whose
illustrates which features have a more significant influence most impactful features do not.
on the predictions. Thus, it enables us to understand better Fig. 10 shows the SHAP decision plot providing a local
which features the model assigns more weight to and how explanation for the classification of the first student as ‘‘fail.’’
they affect the results. The horizontal axis shows SHAP values, representing each
When Fig. 9 is examined, the ‘‘Number of Students’’ feature’s impact on the prediction; positive values increase
was an important feature contributing to the prediction. the prediction, while negative values decrease it. The vertical
The second most effective feature was the ‘‘High School axis lists the features, ordered by their impact. In a SHAP

191644 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

decision plot, the values provided in parentheses next to negatively impact the classification of the second student as
the features represent the values of those features for the ‘‘fail,’’ the student is still classified as ‘‘fail.’’
specific data point being analyzed. The overall effect in If educators intervened solely based on global attributes
the graph is determined by summing the contributions of such as ‘‘Number of Students,’’ ‘‘High School Achievement
all features. According to this analysis, it was determined Score,’’ and ‘‘Course Repeat Count’’ to prevent failure for
that the ‘‘Course Repeat Count,’’ ‘‘Number of Students,’’ all ‘‘fail’’ class students, this approach could be deemed
and ‘‘Number of Courses Taken’’ features had the most effective for the first student. However, it might be perceived
significant impact on the student being classified as ‘‘fail.’’ as relatively less effective in improving the performance of
However, the ‘‘TYT Placement Score’’ negatively affected the second student.
this student’s classification as a ‘‘fail.’’ As a result, local explanations enable educators to provide
students with more personalized education and respond
more effectively to individual needs. This approach allows
for education programs to be organized more effectively,
considering global features and each student’s specific
characteristics and performance.

3) LIME LOCAL EXPLANATION RESULTS

The LIME method allows us to obtain probability values
for each class by selecting a specific instance from the
test dataset. This method graphically explains the reason
for the probability assigned to each class. On the left side
of this graph, the probability of each class is shown, the
middle column displays the most influential features, and the
right column shows the values of these features. The orange
color indicates the positive impact of the feature, while blue
indicates the negative impact.
Fig. 14 presents the results obtained for the first student
FIGURE 10. Decision plots of the first student. using the LIME method. When looking at the probability
values on the left side for the first student, it is noted that
Fig. 11 presents a local explanation of the SHAP Force there is a 93% probability of being classified as a ‘‘fail.’’
Plot for the classification of the first student, same as the The most influential features were determined to be ‘‘Course
decision plot discussed earlier in Fig. 10. The red part of Repeat Count,’’ ‘‘Number of Students,’’ ‘‘TYT Placement
this graph represents the features that positively contribute Score,’’ ‘‘SAY Placement Success Ranking Percentage,’’ and
to the student being classified as ‘‘fail,’’ while the blue part ‘‘Number of Incorrect Answers in Mathematics.’’ Among
indicates adverse effects. Additionally, the length of each bar these features, the ‘‘TYT Placement Score’’ is highlighted
for a feature indicates the degree of impact. Particularly, the in orange to signify positive impact, while the remaining
bars for ‘‘Course Repeat Count,’’ ‘‘Number of Students,’’ and influential features are highlighted in blue to indicate negative
‘‘Number of Courses Taken’’ are more extended, indicating a impact.
higher degree of impact for these features. Fig. 15 presents the results obtained for the second student
Local explanations provide more information for educators using the LIME method. When examining the probability
to deliver personalized education effectively to students. values on the left side for the second student, it is indicated
While providing the same feedback to all fail-class students that there is a 74% probability of being classified as ‘‘fail.’’
based on global features can be effective, this approach The most influential features were determined to be ‘‘Number
may have a limited educational impact for some students. of Students,’’ ‘‘Course Repeat Count,’’ ‘‘Number of Incorrect
For instance, in the case of the first student, the features Answers in History 1’’, ‘‘Number of Correct Answers in
‘‘Course Repeat Count’’ and ‘‘Number of Students’’ had Geography 1’’, and ‘‘Number of Courses Taken.’’ Among
the most significant impact, two of the top three essential these features, ‘‘Number of Incorrect Answers in History 1’’
features according to the global explanation. However, the and ‘‘Number of Correct Answers in Geography 1’’ were
situation is different for the second student; the most highlighted in blue to indicate negative impact.
compelling features in classifying this student as a ‘‘fail,’’ When comparing the SHAP and LIME graphs, it is
as seen in Fig. 12 and Fig. 13, were ‘‘Number of Incorrect observed that the top 5 features in the LIME explanations
Answers in Physics,’’ ‘‘EA Score,’’ and ‘‘Number of Correct are among the top 20 features in the SHAP explanations.
Answers in Chemistry’’ which do not rank among the top However, as in example of the second student, LIME’s 3rd,
three features in the global explanation. Additionally, while fourth, and fifth most essential features are the 16th, 19th,
‘‘Course Repeat Count’’ and ‘‘Number of Students’’ are and 17th most important features in the SHAP explanations.
two of the top three features in the global explanation, they SHAP and LIME can produce different results because

VOLUME 12, 2024 191645

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

FIGURE 11. Force plot of the first student.

educational stages when they took the course. To build

our prediction model, we analyzed 1268 students pursuing
university education in engineering fields. The data used in
our study were obtained from the ‘‘Programming’’ course,
which is commonly offered across different engineering
programs. Due to permission issues regarding data collection,
the academic performance data of students from only one
university were used. The collected data were categorized
into demographic features, academic performance (univer-
sity), high school, and higher education entrance exams.
While four predictive features reflected the skills students
gained during their university education, the remaining
50 features represented the academic skills they acquired
before entering university (e.g., High School Achievement
Score, TYT Placement Score, Number of Correct Answers
FIGURE 12. Decision plots of the second student. in Mathematics, etc.). The academic success classified as
‘‘passed’’ or ‘‘failed’’ for the course was used as the output.
their methods and focus areas differ. SHAP uses a game To make predictions, we used a hybrid model combining
theory-based approach to provide global explanations by Particle Swarm Optimization (PSO) and Deep Neural Net-
considering the overall behavior of the model and feature works (DNN). Features such as ‘‘Course Repeat Count,’’
interactions. In contrast, LIME builds a local model around ‘‘Number of Students,’’ ‘‘Number of Courses Taken,’’
a specific prediction and evaluates the importance of features ‘‘Number of Correct Answers in Chemistry,’’ ‘‘Number
within that local region. These differences make SHAP more of Correct Answers in Basic Mathematics,’’ ‘‘Number of
computationally intensive but comprehensive, while LIME Correct Answers in Biology,’’ and ‘‘Number of Semesters in
offers faster but localized explanations. Preparatory Program’’ had the highest predictive accuracy for
early academic success prediction. When these features are
V. CONCLUSION AND FUTURE STUDIES used, the potential of students can be identified as soon as they
Education is an important factor for socio-economic devel- take the course, and different academic interventions can be
opment in every country and society around the world. preemptively defined for those predicted to fail, helping them
It helps individuals improve their knowledge and skills in pass the course.
areas of interest, enhance their quality of life, contribute to The proposed PSO-DNN model demonstrates outper-
the workforce by providing added value, manage societies formed performance in early prediction of student per-
better, and even better understand the world. They can achieve formance compared to traditional machine learning and
this through the professions they acquire upon graduation, deep learning models, achieving the highest accuracy, F1-
whether in secondary or higher education. However, if they score, precision, and recall. SHAP and LIME facilitated
fail the courses they take during their studies, they not only an understanding of why the model performs better than
fail to graduate but also cannot acquire the professions that others. Identifying key features in classifying each student
will add value. In this context, the academic success of has provided insights into the model’s decision-making
individuals classified as students in society, and the prediction process and improved our interpretation of its predictions.
of this success, becomes an important issue. The need to This model interpretability helps educators understand the
conduct studies to identify students with a high potential fundamental factors affecting student performance and sup-
for failure and who may even drop out in the future arises. ports the development of personalized strategies. In the XAI
When we examine the literature, it is observed that students’ dataset, features such as ‘‘Number of Students,’’ ‘‘High
academic success is mostly predicted based on the academic School Achievement Score,’’ ‘‘Course Repeat Count,’’ ‘‘TYT
efforts they demonstrate in the classroom environment. Placement Score,’’ ‘‘TYT Placement Success Ranking Per-
In this study, we aimed to predict the success of students centage,’’ ‘‘Number of Incorrect Answers in Mathematics,’’
based on the academic skills they acquired from previous ‘‘Number of Incorrect Answers in Social Sciences,’’ ‘‘SAY

191646 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

FIGURE 13. Force plot of the second student.

FIGURE 14. LIME explanation of dataset instances of the first student.

FIGURE 15. LIME explanation of dataset instances of the second student.

Placement Success Ranking Percentage,’’ and ‘‘Number of applied to the model’s hyperparameters, such as swarm
Courses Taken’’ were used to predict the success of 72% of size, maximum iterations, and the number of hidden layers.
the students, while the academic success of the remaining These constraints may affect the model’s performance and
28% was predicted using all the features in the dataset. its effectiveness with complex datasets. While loosening
In our study, the academic performance indicators pro- hyperparameter constraints can yield improvements in model
duced as outputs in the Turkish education environment performance, it introduces significant challenges, including
were used. At the same time, the study was limited to a prolonged training times, elevated computational costs, and a
single university and a single course, which constrains the greater likelihood of overfitting.
generalizability of our findings. To mitigate this limitation, In light of the findings and limitations of this study,
we conducted a performance evaluation study using the several opportunities for future research emerge. First, further
xAPI-Edu-Data dataset, which is widely used in academic validation of the proposed model across different courses,
research and covers various courses. The same dataset has universities, educational settings, and countries such as the
been used in five different studies in the literature with UK, Iran, France, Germany, and Austria where similar
various machine learning methods. When compared to these academic performance metrics are employed during the
studies, our proposed prediction model demonstrated better transition to higher education could enhance its general-
accuracy in predicting students’ academic performance. izability and robustness. Second, comprehensive research
Only the SVM method from Sixhaxa [29] and the DNN on hybrid deep learning models created with different
method from Vijayalakshmi and Venkatachalapathy [26] had optimization techniques and deep learning algorithms could
accuracy rates close to the one we achieved. This success is improve model performance. Lastly, exploring additional
attributed to the PSO method’s effective optimization of the XAI methods and techniques could deepen our understanding
hyperparameters of the DNN deep learning model. Moreover, of model interpretability and provide more comprehensive
our proposed model has been shown to be applicable in insights into model predictions.
different educational environments and courses, representing To evaluate these opportunities in the future, we aim to
an important step toward generalizability. extend the PSO-DNN model to support early prediction
Another limitation is that due to technical infrastructure of student performance and facilitate its integration within
requirements and long processing times, constraints were educational institutions. To enhance accessibility, we plan

VOLUME 12, 2024 191647

 A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

to develop API and cloud-based web services that allow [8] X. Du, J. Yang, and J.-L. Hung, ‘‘An integrated framework based
the model to receive student data, generate predictions, and on latent variational autoencoder for providing early warning of at-
risk students,’’ IEEE Access, vol. 8, pp. 10110–10122, 2020. [Online].
deliver outputs to systems such as Learning Management Available: https://ieeexplore.ieee.org/document/8952699/
Systems (LMS) and Student Information Systems (SIS). [9] F. A. Al-azazi and M. Ghurab, ‘‘ANN-LSTM: A deep learning
Leveraging cloud-based infrastructure will increase the model for early student performance prediction in MOOC,’’ Heliyon,
vol. 9, no. 4, Apr. 2023, Art. no. e15382. [Online]. Available:
model’s scalability, enabling the flexible management of https://linkinghub.elsevier.com/retrieve/pii/S2405844023025896
high-processing-power tasks and facilitating deployment and [10] F. D. Pereira, S. C. Fonseca, E. H. T. Oliveira, A. I. Cristea,
use even in institutions with limited technological infras- H. Bellhäuser, L. Rodrigues, D. B. F. Oliveira, S. Isotani, and
L. S. G. Carvalho, ‘‘Explaining individual and collective programming
tructure. To accelerate adaptation across various courses and students’ behavior by interpreting a black-box predictive model,’’
institutions, we plan to employ transfer learning techniques. IEEE Access, vol. 9, pp. 117097–117119, 2021. [Online]. Available:
This approach will allow a model trained on one institution or https://ieeexplore.ieee.org/document/9517104/
course to be efficiently adapted to similar settings with mini- [11] H. Singh, B. Kaur, A. Sharma, and A. Singh, ‘‘Framework for suggesting
corrective actions to help students intended at risk of low performance
mal data and processing power requirements, ensuring a rapid based on experimental study of college students using explainable machine
adaptation process without performance loss. To enhance learning model,’’ Edu. Inf. Technol., vol. 29, no. 7, pp. 7997–8034,
usability for educators and students, we aim to develop May 2024. [Online]. Available: https://link.springer.com/10.1007/s10639-
023-12072-1
user-friendly interfaces. These will provide detailed analysis [12] M. Adnan, M. I. Uddin, E. Khan, F. S. Alharithi, S. Amin, and
reports of prediction results, supported by Explainable AI A. A. Alzahrani, ‘‘Earliest possible global and local interpretation of
(XAI) visualizations that will help users better understand students’ performance in virtual learning environment by leveraging
explainable AI,’’ IEEE Access, vol. 10, pp. 129843–129864, 2022.
the model’s decision-making process. Additionally, protocols [Online]. Available: https://ieeexplore.ieee.org/document/9970711/
for data security and privacy will be established to ensure [13] O. H. Embarak and S. Hawarna, ‘‘Automated AI-driven
the protection of student data. Finally, mechanisms will be system for early detection of at-risk students,’’ Proc. Comput.
Sci., vol. 231, pp. 151–160, Jan. 2024. [Online]. Available:
put in place to enable continuous learning and re-training https://linkinghub.elsevier.com/retrieve/pii/S1877050923021993
of the model with new data, enhancing its adaptability over [14] Y. Jang, S. Choi, H. Jung, and H. Kim, ‘‘Practical early prediction of
time. These developments will support data-driven decision- students’ performance using machine learning and eXplainable AI,’’ Edu.
Inf. Technol., vol. 27, no. 9, pp. 12855–12889, Nov. 2022. [Online].
making in education and contribute to improving student Available: https://link.springer.com/10.1007/s10639-022-11120-6
success. [15] M. Abdalkareem and N. Min-Allah, ‘‘Explainable models for predict-
ing academic pathways for high school students in Saudi Arabia,’’
ACKNOWLEDGMENT IEEE Access, vol. 12, pp. 30604–30626, 2024. [Online]. Available:
https://ieeexplore.ieee.org/document/10444523/
Not applicable. [16] K. Kameyama, ‘‘Particle swarm optimization-A survey,’’ IEICE Trans.
Inf. Syst., vol. 92, no. 7, pp. 1354–1361, Jul. 2009. [Online]. Available:
REFERENCES http://www.jstage.jst.go.jp/article/transinf/E92.D/7/E92.D_7_1354/_article
[1] A. Hernández-Blanco, B. Herrera-Flores, D. Tomás, and B. Navarro- [17] M. Clerc and J. Kennedy, ‘‘The particle swarm–explosion, stability,
Colorado, ‘‘A systematic review of deep learning approaches to educational and convergence in a multidimensional complex space,’’ IEEE Trans.
data mining,’’ Complexity, vol. 2019, no. 1, Jan. 2019, Art. no. 1306039, Evol. Comput., vol. 6, no. 1, pp. 58–73, Feb. 2002. [Online]. Available:
doi: 10.1155/2019/1306039. http://ieeexplore.ieee.org/document/985692/
[2] D. A. Shafiq, M. Marjani, R. A. A. Habeeb, and D. Asirvatham, ‘‘Student [18] A. B. Putra Utama, A. P. Wibawa, M. Muladi, and A. Nafalski, ‘‘PSO
retention using educational data mining and predictive analytics: A based hyperparameter tuning of CNN multivariate time-series analysis,’’
systematic literature review,’’ IEEE Access, vol. 10, pp. 72480–72503, Jurnal Online Informatika, vol. 7, no. 2, pp. 193–202, Dec. 2022. [Online].
2022. [Online]. Available: https://ieeexplore.ieee.org/document/9815588/ Available: https://join.if.uinsgd.ac.id/index.php/join/article/view/858
[3] X. Li, Y. Zhang, H. Cheng, M. Li, and B. Yin, ‘‘Student achievement [19] D. Sarkar, T. Khan, and F. A. Talukdar, ‘‘Hyperparameters opti-
prediction using deep neural network from multi-source campus data,’’ mization of neural network using improved particle swarm opti-
Complex Intell. Syst., vol. 8, no. 6, pp. 5143–5156, Dec. 2022. [Online]. mization for modeling of electromagnetic inverse problems,’’ Int. J.
Available: https://link.springer.com/10.1007/s40747-022-00731-8 Microw. Wireless Technol., vol. 14, no. 10, pp. 1326–1337, Dec. 2022.
[4] B. Venkatachalam and K. Sivanraju, ‘‘Predicting Student performance [Online]. Available: https://www.cambridge.org/core/product/identifier/
using mental health and linguistic attributes with deep learning,’’ Revue S1759078721001690/type/journal_article
d’Intell. Artificielle, vol. 37, no. 4, pp. 889–899, Aug. 2023. [Online]. [20] A. Pranolo, Y. Mao, A. P. Wibawa, A. B. P. Utama, and
Available: https://www.iieta.org/journals/ria/paper/10.18280/ria.370408 F. A. Dwiyanto, ‘‘Optimized three deep learning models based-PSO
[5] H.-C. Chen, E. Prasetyo, S.-S. Tseng, K. T. Putra, S. S. Kusumawardani, hyperparameters for Beijing PM2.5 prediction,’’ Knowl. Eng.
and C.-E. Weng, ‘‘Week-wise Student performance early prediction Data Sci., vol. 5, no. 1, p. 53, Jun. 2022. [Online]. Available:
in virtual learning environment using a deep explainable artificial http://journal2.um.ac.id/index.php/keds/article/view/29489
intelligence,’’ Appl. Sci., vol. 12, no. 4, p. 1885, Feb. 2022. [Online]. [21] L. James and V. Miranda, ‘‘PySwarms: A research toolkit
Available: https://www.mdpi.com/2076-3417/12/4/1885 for particle swarm optimization in Python,’’ J. Open Source
[6] X. Li, X. Zhu, X. Zhu, Y. Ji, and X. Tang, ‘‘Student academic Softw., vol. 3, no. 21, p. 433, Jan. 2018. [Online]. Available:
performance prediction using deep multi-source behavior sequential http://joss.theoj.org/papers/10.21105/joss.00433
network,’’ in Advances in Knowledge Discovery and Data Mining [22] F. Mei, X. Li, J. Zheng, H. Sha, and D. Li, ‘‘A data-driven
(Lecture Notes in Computer Science), vol. 12084, H. W. Lauw, approach to state assessment of the converter valve based on over-
R. C.-W. Wong, A. Ntoulas, E.-P. Lim, S.-K. Ng, and S. J. Pan, Eds., sampling and shapley additive explanations,’’ IET Gener., Transmiss.
Cham, Switzerland: Springer, 2020, pp. 567–579. [Online]. Available: Distrib., vol. 16, no. 8, pp. 1607–1619, Apr. 2022. [Online]. Available:
http://link.springer.com/10.1007/978-3-030-47426-3_44 https://onlinelibrary.wiley.com/doi/10.1049/gtd2.12385
[7] D. Uliyan, A. S. Aljaloud, A. Alkhalil, H. S. A. Amer, [23] A. R. Rajpurkar, L. J. Mateo, S. E. Murphy, and A. N. Boettiger,
M. A. E. A. Mohamed, and A. F. M. Alogali, ‘‘Deep learning ‘‘Deep learning connects DNA traces to transcription to reveal
model to predict students retention using BLSTM and CRF,’’ predictive features beyond enhancer–promoter contact,’’ Nature
IEEE Access, vol. 9, pp. 135550–135558, 2021. [Online]. Available: Commun., vol. 12, no. 1, p. 3423, Jun. 2021. [Online]. Available:
https://ieeexplore.ieee.org/document/9555608/ https://www.nature.com/articles/s41467-021-23831-4

191648 VOLUME 12, 2024

A. Kala et al.: Early Prediction of Student Performance in Face-to-Face Education Environments

[24] L. Fu, K. Luo, J. Lv, X. Wang, S. Qin, Z. Zhang, S. Sun, X. Wang, B. Yun, Industrial Engineering, from 2011 to 2014, and the Dean of the Faculty
Y. He, W. He, W. Li, and L. Chen, ‘‘Integrating expression data-based of Engineering, from 2014 to 2018. He was the Vice Rector of Yalova
deep neural network models with biological networks to identify University, from 2018 to 2020. He conducts research and studies in the
regulatory modules for lung adenocarcinoma,’’ Biology, vol. 11, no. 9, fields of manufacturing planning and control, management information
p. 1291, Aug. 2022. [Online]. Available: https://www.mdpi.com/2079- systems, distance education, artificial intelligence, digital transformation,
7737/11/9/1291 and industry 4.0. He has articles in many international peer-reviewed
[25] P. Linardatos, V. Papastefanopoulos, and S. Kotsiantis, ‘‘Explainable AI: journals, papers at national and international conferences, and book chapters.
A review of machine learning interpretability methods,’’ Entropy, vol. 23,
He took part in many national and international conference and symposium
no. 1, p. 18, Dec. 2020. [Online]. Available: https://www.mdpi.com/1099-
organizing committees and participated in many conferences, scientific
4300/23/1/18
[26] V. Vijayalakshmi and K. Venkatachalapathy, ‘‘Comparison of predicting meetings, and workshops as an invited speaker. He teaches various courses
student’s performance using machine learning algorithms,’’ Int. J. Intell. on management information systems, information systems security and
Syst. Appl., vol. 11, no. 12, pp. 34–45, Dec. 2019. [Online]. Available: control, industry 4.0 and digital transformation, and advanced manufacturing
http://www.mecs-press.org/ijisa/ijisa-v11-n12/v11n12-4.html planning and control systems. He is the Founding Partner and the Chairman
[27] S. Tuaha, I. F. Siddiqui, and Q. Ali Arain, ‘‘Analyzing students’ of the Board of Directors at Arvasis Information and Consultancy Company.
academic performance through educational data mining,’’
3C Tecnología_Glosas de innovación aplicadas a la pyme,
vol. 8, no. 1, pp. 402–421, May 2019. [Online]. Available:
https://www.3ciencias.com/articulos/articulo/analyzing-students-
academic-performance-through-educational-data-mining
[28] M. S. Abubakari, F. Arifin, and G. G. Hungilo, ‘‘Predicting students’
academic performance in educational data mining based on deep learning
using TensorFlow,’’ Int. J. Edu. Manage. Eng., vol. 10, no. 6, pp. 27–33,
Dec. 2020. [Online]. Available: http://www.mecs-press.org/ijeme/ijeme-
v10-n6/v10n6-4.html
[29] K. Sixhaxa, A. Jadhav, and R. Ajoodha, ‘‘Predicting students performance
in exams using machine learning techniques,’’ in Proc. 12th Int. Conf. TUGBA TUNACAN YILDIZ was born in Sakarya,
Cloud Comput., Data Sci. Eng. (Confluence), Noida, India, Jan. 2022, in 1981. She received the degree from the
pp. 635–640. [Online]. Available: https://ieeexplore.ieee.org/document/ Department of Industrial Engineering, Faculty
9734218/ of Engineering, Sakarya University, in 2001,
[30] U. Pujianto, W. A. Prasetyo, and A. R. Taufani, ‘‘Students academic and the master’s and Ph.D. degrees from the
performance prediction with K-Nearest neighbor and C4.5 on SMOTE- Department of Industrial Engineering, Institute
balanced data,’’ in Proc. 3rd Int. Seminar Res. Inf. Technol. Intell. of Science, Sakarya University, in 2004 and
Syst. (ISRITI), Yogyakarta, Indonesia, Dec. 2020, pp. 348–353. [Online]. 2010, respectively. From 2011 to 2017, she was
Available: https://ieeexplore.ieee.org/document/9315439/
an Assistant Professor with the Department of
Industrial Engineering, Faculty of Engineering,
Karabük University. Since 2017, she has been an Assistant Professor with
the Department of Industrial Engineering, Faculty of Engineering and
Architecture, Bolu Abant Izzet Baysal University.
AHMET KALA was born in Sakarya, in 1981.
He received the degree from the Department of
Industrial Engineering, Faculty of Engineering,
Sakarya University, in 2003, and the master’s
degree from the Department of Mechanical and
Manufacturing Engineering, Institute of Science,
Bilecik Şeyh Edebali University, in 2014. He is
currently pursuing the Ph.D. degree with the
Department of Industrial Engineering, Institute of
Science, Sakarya University. From 2010 to 2021,
he was a Lecturer with Bilecik Şeyh Edebali University. Since 2021, he has IHSAN HAKAN SELVI received the bachelor’s,
been a Software Engineer with the Information Technology Department, master’s, and the Ph.D. degrees in industrial
Sakarya Applied Sciences University. engineering from Sakarya University, in 2001,
2004, and 2011, respectively. He was a Research
Assistant with the Informatics Department and
Distance Education Research and Application
Center, Sakarya University, from 2001 to 2010.
ORHAN TORKUL received the bachelor’s degree From August 2013 to July 2014, he was a Visiting
from Sakarya University, Sakarya Engineering Researcher with the University of Missouri Sci-
Faculty, Department of Industrial Engineering, ence and Technology, USA. He is currently with
in 1982, the master’s degree from Yıldız Technical the Department of Information Systems Engineering, Sakarya University.
University, Institute of Science and Technology, He is a project coordinator in the 2244 Ph.D. Program and a Researcher
Department of Industrial Engineering, in 1987, in 1001 research projects supported by the Scientific and Technical
and the Ph.D. degree from Cranfield Institute Research Council of Turkey (TUBITAK). He conducts research and
of Technology, England, in 1993. He was with lectures on production and service systems, digital transformation, machine
the Industrial Engineering Department, Sakarya learning, deep learning, and heuristic optimization. He has articles in
University, Faculty of Engineering, an Assistant many international peer-reviewed journals, book editorship, book chapter
Professor, from 1993 to 1998, an Associate Professor, from 1998 to 2003, authorship, and national/international conference presentations. He took part
and a Professor, since 2003. He was the Vice Dean of the Faculty of in many national and international conferences and symposium organizing
Engineering, from 1993 to 1995, the Head of the Department of Informatics, committees. He is on the Editorial Board of Sakarya University Journal of
from 1997 to 2011, the Director of the Distance Education Research and Science.
Application Center, from 2005 to 2011, the Head of the Department of

VOLUME 12, 2024 191649