A Deep Learning Approach Towards Student

Performance Prediction in Online Courses:

Challenges Based on a Global Perspective
Abdallah Moubayed∗†§ , MohammadNoor Injadat‡ , Nouh Alhindawi∗¶ , Ghassan Samara∥ ,
Sara Abuasal∗∗ , and Raed Alazaidah‡
∗ School of Computing and Augmented Intelligence, Arizona State University, Gilbert, Arizona, USA;
e-mails: {amoubaye, nalhinda}@asu.edu
arXiv:2402.01655v1 [cs.CY] 10 Jan 2024

† Electrical and Computer Engineering Department, Lebanese American University, Byblos, Lebanon;

e-mail: [email protected]
‡ Data Science and Artificial Intelligence Department, Faculty of Information Technology, Zarqa University, Zarqa, Jordan;

e-mails: {minjadat, razaidah}@zu.edu.jo

§ Electrical and Computer Engineering Department, Western University, London, Ontario, Canada;

e-mail: [email protected]
¶ Faculty of Science and Information Technology, Jadara University, Irbid, Jordan;

e-mail: [email protected]
∥ Computer Science Department, Faculty of Information Technology, Zarqa University, Zarqa, Jordan;

e-mail: [email protected]
∗∗ Software Engineering Department, Faculty of Information Technology, Zarqa University, Zarqa, Jordan;

e-mail: [email protected]

Abstract—Analyzing and evaluating students’ [3], [4], [5]. The ongoing researches in this area
progress in any learning environment is stressful and are trying to address certain objectives that enhance
time consuming if done using traditional analysis the e-Learning environment from a certain point of
methods. This is further exasperated by the increasing
number of students due to the shift of focus toward view. One of those important aspects is offering
integrating the Internet technologies in education and personalized learning [6]. The researchers focus on
the focus of academic institutions on moving toward personalized learning because of the importance of
e-Learning, blended, or online learning models. As a this aspect in improving the e-Learning contents and
result, the topic of student performance prediction delivery methods to satisfy the learner’s needs [7].
has become a vibrant research area in recent years.
To address this, machine learning and data mining Another important aspect is the adaptive learning,
techniques have emerged as a viable solution. To that which concentrates on the learning style of the learner
end, this work proposes the use of deep learning [8]. Both aspects will encourage the e-Learning en-
techniques (CNN and RNN-LSTM) to predict the vironment’s parties to integrate the required features
students’ performance at the midpoint stage of the and technologies in the e-Learning environment in
online course delivery using three distinct datasets
collected from three different regions of the world. order to satisfy the individual learner’s needs [9].
Experimental results show that deep learning models Consequently, this requires research in learning ana-
have promising performance as they outperform other lytics to analyze the learners’ needs. Thus, conducting
optimized traditional ML models in two of the three traditional analytical research with the large number
considered datasets while also having comparable of learners nowadays is a complex and time consum-
performance for the third dataset.
Index Terms—Deep Learning, e-Learning, Online ing process.
Courses, Student Performance Prediction As a solution, this paper proposes a predictive
model to predict the learner’s final grade in the course
I. I NTRODUCTION at earlier stages during the course. The predictive
model implements deep learning (DL) models to clas-
With the continued growth of Internet users, sify the students and predicts their final grades at the
the demand for new learning paradigms that rely on midway point of the online course. More specifically,
e-Learning environments is increasing rapidly [1], a convolutional neural network (CNN) and a recurrent
[2]. Different researches have been conducted in neural network with long short term memory (RNN-
the e-Learning environment in order to improve the LSTM) models are proposed as they have the poten-
beneficial advantages of e-Learning courses’ delivery
tial to accurately predict student performance given course program across 6 years.
their promising results in other applications such as Similarly, the authors of [13] proposed using a
pancreas and breast tumor detection [10], [11]. As a decision tree algorithm, namely the J48 (also known
result, this will help the instructors at earlier stages of as C4.5) algorithm, to predict student performance.
the course delivery to take care of the students who Their results emphasized the potential of such algo-
may need help. rithms for student performance prediction as they had
To evaluate the performance of the proposed deep high accuracy and speed.
learning model, three distinct datasets from three The authors of [14] also proposed using the J48
different universities located in three different regions algorithm by comparing its performance with that
in the world are used. More specifically, the first of the k-means algorithm. The results reiterated the
dataset is for a first year engineering course at a superiority of the the J48 algorithm in predicting
European University. The second dataset is for a student performance.
third year engineering course at a North American In contrast, the authors of [16] investigated the
University. Finally, the third dataset is for a first year performance of different clustering algorithms such
undergraduate course in Information Technology (IT) as k-means and hierarchical clustering in accurately
at a Middle Eastern university. This provides us with predicting student performance. The experimental re-
a more global model that is applicable to different sults showed that k-means algorithm outperformed
courses and student demographics. other algorithms as it had the better performance and
The rest of the paper is organized as follows: Sec- the faster building time.
tion II describes the previous researches and related The authors in [17] proposed the use of deep neural
work. Section III represents the proposed approach. network model for an e-learning recommendation
Section IV describes the datasets used. Section V framework. Experimental results illustrated that the
evaluates the performance of the proposed framework proposed model improved the learning experience of
in comparison with other works from the literature. the students.
Finally, Section VI concludes the paper and discusses In a similar fashion, the authors of [18] also pro-
the future work. posed using deep learning to predict student perfor-
mance. Their distinctive characteristic was that they
II. R ELATED W ORKS
used both academic and non-academic subjects. Ex-
Due to the fact that this research area has been perimental results showed that deep learning models
gaining significant attention and growing in popu- are a viable potential solution for such problems as
larity, several research works have investigated the it can offer high accuracy.
use of machine learning and data mining techniques On the other hand, the authors in [15] compared the
in educational settings. As shown in Figure 1, these performance of four tree-based models in accurately
works can be categorized as either being year-to-year predicting students’ performance. Using a Portuguese
prediction frameworks (such as [12], [13], [14], [15]) high school dataset, their experimental results showed
or course-to-course prediction frameworks (such as that tree-based models achieved high accuracy and
[16], [17]). Year-to-year prediction frameworks aim fast execution time.
to predict the performance of the students in the Finally, the authors in [19], [20] used unsupervised
courses/classes of a year based on their performance learning models to cluster users in various engage-
in the previous year. Alternatively, course-to-course ment levels. Based on these levels, they used Apriori
prediction frameworks aim to predict the performance association rules to related academic performance
of students in a course based on their performance in with student engagement. Their experimental results
previous similar courses. showed that there exists a positive correlation be-
tween students’ engagement level and their academic
performance in an e-learning environment.
Despite the promising results presented in previous
related works, their main limitation lies in the fact that
they rely on courses/classes that have already been
completed to predict the performance in subsequent
courses/classes. As mentioned earlier, this is done
either on a course-to-course basis or a year-to-year
Figure 1. Student Performance Prediction Frameworks’ Catego- basis. However, very few previous works aimed at
rization predicting the student performance in a course during
its delivery. Hence, this work aims at filling this gap
For example, the authors in [12] proposed the use by trying to accurately predict the performance of the
of classification models to predict the final grades of students at the midpoint stage of the course delivery
students. This was done by analyzing data from a
to ensure satisfactory completion of the course itself from three different universities located in three sepa-
rather than in subsequent courses. rate regions are used. In what follows, a brief descrip-
tion and visualization of these datasets is provided.
III. P ROPOSED F RAMEWORK
This work proposes using deep learning-based A. Dataset 1: First Year Engineering Course at a
models to predict the student performance at the European University
midpoint stage of the course. As is shown in Figure 2, The first dataset was collected for a first year
the proposed framework includes two main portions. engineering course at a European University [22].
The first portion is the data pre-processing and It consists of a set of tasks of different weights
and difficulty levels that were completed using a
dedicated simulation environment. Note that the data
was collected for 115 students originally. However,
only 52 students completed the course and thus were
used as part of this dataset. Figure 3 plots the first
and second principal components for Dataset 1.

Figure 2. Proposed Deep Learning-based Framework

Figure 3. Dataset 1 visualization
scaling portion. Specifically, the proposed framework
uses z-score normalization as per the following equa-
B. Dataset 2: Third Year Engineering Course at a
tion:
x − µx North American University
xscaled = (1)
σx The second dataset was collected for a third year
where µx is the mean vector and σx is the standard engineering course at a North American university
deviation vector of the training samples. As a result, which followed an online delivery mode. It contains
all features have a unified scale with zero mean and the records for 162 students and consists of their
unit standard deviation. It is worth noting that z-score performance in various evaluation tasks including as-
normalization is used in this case since educational signments, a midterm exam, and a final exam. Figure
data representing grades often follow Gaussian dis- 4 plots the first and second principal components for
tribution [21]. Dataset 2.
After splitting the data into training and testing
datasets using 80%-20% ratio, the data is fed to
the deep learning model. Two different models are
considered, namely the CNN and the RNN-LSTM
models. The CNN model consists of a 1D convolu-
tional layer, a max pooling layer, a flatten layer, and
finally a dense layer made up of 128 units with the
adam optimizer. The RNN-LSTM model consists of
the LSTM layer made up of 64 units and a dense
layer made up of 3 units with the ReLu and Softmax
activation functions.

IV. DATASETS ’ D ESCRIPTION

To evaluate the performance of the proposed deep Figure 4. Dataset 2 visualization
learning framework, three distinct datasets collected
C. Dataset 3: First Year IT Course at a Middle recall, and F-score which can be calculated using the
Eastern University following equations [23], [24]:
The third dataset was collected for a first year (T P + T N )
IT course at a Middle Eastern University that was Accuracy = (2)
(T P + T N + F P + F N )
also delivered in an online manner. It contains the
records for a group of 222 students and consists of the TP
P recision = (3)
students’ performance in the different assignments, (T P + F P )
one midterm exam, and one final exam. Figure 5 plots TP
the first and second principal components for Dataset Recall = (4)
(T P + F N )
3.
(P recision × Recall)
F − score = 2 × (5)
(P recision + Recall)
C. Performance Results & Discussion
The performance of the proposed deep learning-
based models is compared to four optimized tradi-
tional machine learning (ML) models, namely Sup-
port Vector Machines (SVM), K-nearest neighbors
(K-NN), Naive Bayes (NB), and Random Forest
(RF). These models are optimized using grid search
algorithm.
Tables I, II, and III provide the performance results
of the different ML and DL models for Dataset 1,
Dataset 2, and Dataset 3 respectively. The following
Figure 5. Dataset 3 visualization
observations can be made. The results from Table
I show that the RNN-LSTM model clearly outper-
D. Datasets’ Processing: formed the other models. In a similar fashion, the
For all three datasets, the data was further pro- results from Table III show that the CNN model
cessed by replacing empty marks with a 0 and outperforms the other models. Additionally, although
rounding all decimal value grades to the nearest 1. the optimized K-NN model had the best performance
Moreover, the students’ final course grade was used to for Dataset 2 as is shown in Table II, the CNN
construct the target classification variable as follows: model achieved comparable performance. The slight
degradation of the deep learning-based models for
• Good (G): Given to students that finished the
this dataset compared to that of the optimized K-NN
course with a final grade in the range between
is attributed to the extremely low number of weak
70-100%.
students to be classified (4 students out of 162) which
• Fair (F): Given to students that finished the
resulted in these models incorrectly classifying them
course with a final grade in the range between
as good students.
51-69%.
• Weak (W): Given to students that finished the
Table I
course with a final grade in the range between P ERFORMANCE E VALUATION OF T RADITIONAL ML AND DL
0-50%. M ODELS F OR DATASET 1
Our goal is to accurately identify the students classi- Algorithm Accuracy Precision Recall F-score
fied as W since these are the ones that are predicted Optimized SVM 0.45 0.46 0.48 0.46
Optimized K-NN 0.72 0.59 0.72 0.63
to potentially fail the course. Optimized RF 0.63 0.56 0.63 0.59
Optimized NB 0.64 0.71 0.64 0.64
CNN 0.45 0.56 0.45 0.47
V. P ERFORMANCE E VALUATION RNN-LSTM 0.82 0.89 0.82 0.82

A. Experiment Setup
MATLAB 2022a is used to plot the PCA figures
Table II
for the different considered datasets. Additionally, P ERFORMANCE E VALUATION OF T RADITIONAL ML AND DL
Python is used to develop and evaluate the perfor- M ODELS F OR DATASET 2
mance of the proposed framework. Algorithm Accuracy Precision Recall F-score
Optimized SVM 0.87 0.88 0.88 0.87
B. Performance Metrics Optimized K-NN 0.94 0.95 0.94 0.94
Optimized RF 0.91 0.91 0.91 0.91
Four main metrics are used to evaluate the perfor- Optimized NB 0.91 0.91 0.92 0.91
CNN 0.91 0.91 0.91 0.91
mance of the proposed deep learning-based frame- RNN-LSTM 0.88 0.88 0.88 0.88

work. These are the prediction accuracy, precision,

 Table III example, the data associated with students that are
P ERFORMANCE E VALUATION OF T RADITIONAL ML AND DL predicted to fail the course can be further studied
M ODELS F OR DATASET 3
and analyzed to determine the root cause of their
Algorithm
Optimized SVM
Accuracy
0.88
Precision
0.91
Recall
0.85
F-score
0.88
poor performance. This can be done using techniques
Optimized K-NN 0.86 0.87 0.86 0.86 such as association rules to determine which learning
Optimized RF 0.82 0.82 0.82 0.82
Optimized NB 0.87 0.86 0.87 0.86 outcomes are these students struggling with. Subse-
CNN 0.91 0.92 0.91 0.91
RNN-LSTM 0.82 0.83 0.82 0.82
quently, additional resources can be provided to these
students (e.g. additional videos, additional readings,
practice exercises, etc.) to ensure that they are meet-
Accordingly, based on the aforementioned results, ing the expected student learning outcomes, making
it can be concluded that deep learning models have their learning experience feel better customized and
significant promise in accurately predicting student personalized.
performance at the midpoint stage of the online
course. Therefore, they can be relied on by university R EFERENCES
instructors to identify students who are at risk of [1] M. Injadat, A. Moubayed, A. B. Nassif, and A. Shami,
failing at an earlier stage. In turn, this would result in “Multi-split optimized bagging ensemble model selection for
better student performance as these students can then multi-class educational data mining,” Applied Intelligence,
vol. 50, pp. 4506–4528, 2020.
be guided to offer them further help and ensure they [2] ——, “Systematic ensemble model selection approach for
successfully complete and pass the course at hand. educational data mining,” Knowledge-Based Systems, vol.
200, p. 105992, 2020.
VI. C ONCLUSION [3] A. M. Musleh Al-Sartawi, “E-learning improves accounting
education: case of the higher education sector of bahrain,”
Analyzing and evaluating students’ progress in any in Information Systems: 17th European, Mediterranean, and
learning environment is stressful and time consuming Middle Eastern Conference, EMCIS 2020, Dubai, United
if done using traditional analysis methods. With the Arab Emirates, November 25–26, 2020, Proceedings 17.
Springer, 2020, pp. 301–315.
shift towards integrating the Internet technologies [4] T. Assaraira, Z. Al-Anber, S. Bani-Mohammad, Z. Al-
in education and the focus of academic institutions bashaireh, K. Aldheisat, Z. Alreqeb, O. Saafneh, and N. Al-
on moving toward e-Learning, blended, or online hindawi, “A study on the implementation of integrating
online learning (fully and blended forms) and its quality as-
learning models, this issue is further exasperated. surance in jordanian universities,” in 2021 22nd International
Consequently, the topic of student performance pre- Arab Conference on Information Technology (ACIT), 2021,
diction has become a vibrant research area in recent pp. 1–10.
[5] T. Assaraira, N. Alhindawi, S. Bani-Mohammad, Z. Al-
years. To address this, machine learning and data Anber, and Z. Albashaireh, “The jordanian universities ex-
mining techniques have emerged as a viable solution. perience in integrating online learning and its quality assur-
To that end, this work proposed the use of deep learn- ance,” International Arab Journal of Information Technology,
vol. 19, no. 3, pp. 544 – 565, 2022.
ing techniques, namely convolutional neural networks [6] S. Kausar, X. Huahu, I. Hussain, Z. Wenhao, and M. Zahid,
(CNN) and recurrent neural networks with long short “Integration of data mining clustering approach in the person-
term memory (RNN-LSTM), to predict the students’ alized e-learning system,” IEEE access, vol. 6, pp. 72 724–
72 734, 2018.
performance at the midpoint stage of the online [7] M. Murtaza, Y. Ahmed, J. A. Shamsi, F. Sherwani, and
course delivery using three distinct datasets collected M. Usman, “Ai-based personalized e-learning systems: Is-
from three different regions of the world. Exper- sues, challenges, and solutions,” IEEE Access, 2022.
[8] S. V. Kolekar, R. M. Pai, and M. P. MM, “Rule based adaptive
imental results showed that deep learning models user interface for adaptive e-learning system,” Education and
had promising performance. More specifically, they Information Technologies, vol. 24, pp. 613–641, 2019.
outperform other optimized traditional ML models in [9] Z. Zhang, T. Cao, J. Shu, and H. Liu, “Identifying key factors
affecting college students’ adoption of the e-learning system
two of the three considered datasets while also having in mandatory blended learning environments,” Interactive
comparable performance for the third dataset. Learning Environments, vol. 30, no. 8, pp. 1388–1401, 2022.
In terms of model performance, the main take- [10] M. T. Zavalsız, S. Alhajj, K. Sailunaz, T. Özyer, and R. Al-
away is that no single deep learning model can be hajj, “A comparative study of different pre-trained deeplearn-
ing models and custom cnn for pancreatic tumor detection,”
generalized across multiple datasets. Therefore, it is International Arab Journal of Information Technology, 2023.
recommended that multiple models are combined as [11] K. L. Narayanan, R. S. Krishnan, and Y. H. Robinson, “A
ensembles using different techniques (e.g. majority hybrid deep learning based assist system for detection and
classification of breast cancer from mammogram images,” In-
voting, bagging, boosting, etc.) to ensure that the ternational Arab Journal of Information Technology, vol. 19,
student performance prediction is performed more no. 6, 2022.
accurately and in a more generic manner. [12] A. Ahmed and I. S. Elaraby, “Data mining: A prediction
for student’s performance using classification method,” World
From the application point of view, these promising Journal of Computer Application and Technology, vol. 2,
results open the door to offering more customized no. 2, pp. 43–47, 2014.
and personalized learning experiences. More specif- [13] B. Khan, M. S. H. Khiyal, and M. D. Khattak, “Final grade
prediction of secondary school student using decision tree,”
ically, the results achieved using the aforementioned International Journal of Computer Applications, vol. 115,
framework open the door to further analysis. For no. 21, 2015.
[14] R. Saxena, “Educational data mining: performance evalua- Education, vol. 34, no. 2, pp. 137–156, 2020.
tion of decision tree and clustering techniques using weka [20] ——, “Relationship between student engagement and per-
platform,” International Journal of Computer Science and formance in e-learning environment using association rules,”
Business Informatics, vol. 15, no. 2, pp. 26–37, 2015. in 2018 IEEE World Engineering Education Conference
[15] A. Jain and S. Solanki, “An efficient approach for multiclass (EDUNINE), 2018, pp. 1–6.
student performance prediction based upon machine learn- [21] Y. Nakata, Y. Kozaki, K. Bannaka, M. KONDO, Y. Mi-
ing,” in 2019 International Conference on Communication zokoshi, K. Mitsunari, and K. Takamatsu, “Sustainability of
and Electronics Systems (ICCES). IEEE, 2019, pp. 1457– equal evaluations among teachers of first-year students in
1462. higher education,” in 2019 International Congress on Applied
[16] S. Rana and R. Garg, “Evaluation of student’s performance Information Technology (AIT), 2019, pp. 1–6.
of an institute using clustering algorithms,” International [22] M. Vahdat, L. Oneto, D. Anguita, M. Funk, and M. Rauter-
Journal of Applied Engineering Research, vol. 11, no. 5, pp. berg, “A learning analytics approach to correlate the aca-
3605–3609, 2016. demic achievements of students with interaction data from an
[17] X. Wang, Y. Zhang, S. Yu, X. Liu, Y. Yuan, and F.-Y. educational simulator,” in Design for Teaching and Learning
Wang, “E-learning recommendation framework based on in a Networked World: 10th European Conference on Tech-
deep learning,” in 2017 IEEE international conference on nology Enhanced Learning, EC-TEL 2015, Toledo, Spain,
systems, man, and cybernetics (SMC). IEEE, 2017, pp. 455– September 15-18, 2015, Proceedings 10. Springer, 2015,
460. pp. 352–366.
[18] W. W. Fok, Y. He, H. A. Yeung, K. Law, K. Cheung, Y. Ai, [23] A. Sharma and R. Rani, “Classification of cancerous profiles
and P. Ho, “Prediction model for students’ future develop- using machine learning,” in International Conference on
ment by deep learning and tensorflow artificial intelligence Machine Learning and Data Science (MLDS’17), Dec. 2017,
engine,” in 2018 4th international conference on information pp. 31–36.
management (ICIM). IEEE, 2018, pp. 103–106. [24] M. Injadat, A. Moubayed, A. B. Nassif, and A. Shami,
[19] A. Moubayed, M. Injadat, A. Shami, and H. Lutfiyya, “Multi-stage optimized machine learning framework for net-
“Student engagement level in an e-learning environment: work intrusion detection,” IEEE Transactions on Network
Clustering using k-means,” American Journal of Distance and Service Management, pp. 1–1, Aug. 2020.