International Journal of Information and Education Technology, Vol. 14, No.

2, 2024

A Comparative Analysis for GPA Prediction of

Undergraduate Students Using Machine and Deep Learning
Ibrahim Alnomay, Abdullah Alfadhly*, and Aali Alqarni
King Abdulaziz City for Science and Technology KACST, Riyadh, Saudi Arabia
Email: [email protected] (I.A.); [email protected] (A.A.); [email protected] (A.A.)
*Corresponding author

Manuscript received July 21, 2023; revised August 8, 2023; accepted August 29, 2023; published February 18, 2024

Abstract—Recently the research field of machine learning has the fact that many courses were transferred from different
experienced a huge rise in popularity and growth. Machine universities, and courses were changed during the student
Learning (ML) is a way of improving computational prediction academic life cycle, and courses were taken more than once,
models by allowing the computer to generate its own algorithm
to predict outcomes, based on an existing dataset. In this paper,
all of that have been taken in consideration in the cleansing
we demonstrate the application of Machine Learning to enhance and data processing stage.
the educational processes. We implemented regression and The contribution of our work is in the ability to deal with
supervised learning techniques on data from King Saud large academic dataset that span more than 20 years, serving
University, Riyadh, Saudi Arabia, to construct a predictive multiple academic plans. Most of the research that predict the
model for student performance. This allows for timely student performance, uses a specific plan to develop their
interventions in students' academic paths. We utilized extensive
and diverse course records, encompassing several academic
machine learning algorithms. This may not be the optimal
years and programs, to conduct a comparative analysis of solution in situations where big data are available to
various Machine and Deep Learning methodologies, assessing university administration and they need a way to deal with
their efficacy through performance metrics. The developed such dataset. In our work, we devised multiple steps to
ML/DL algorithms use Grade Point Averages (GPAs) of courses convert the currently available data to tables with unified
and semesters as explanatory features to predict the student’s features that can be fed to the machine learning algorithms.
final GPA, which is the target value of the models. Based on the
results, the linear and bagging regression models have the best
Moreover, since we did not depend on a specific academic
Mean Absolute Error (MAE) performance metric. To ensure plan, we derived a way from the course code in order to know
there will be enough time for academic intervention, data of when these courses are likely taken by students during their
early courses and semesters are used. academic study. This paper is a part of a larger project to
achieve the goal of new generation of Smart universities, by
Keywords—deep learning, Grade Point Average (GPA) which Artificial Intelligence (AI) will play a greater rule in
prediction, machine learning, student performance
the education system.
In the following sections of the paper, we highlight works
I. INTRODUCTION
done to utilize AI in the education sector in Section II. In
The application of machine learning (ML) to the prediction Section III, we described our research methodology,
of student grades has been studied intensively in recent years. consisting of preprocessing data, features selection, and
Many researchers [1–6] have proposed machine learning machine learning development. Our results are discussed in
models to predict Grade Point Averages, with their findings detail in Section IV.
showing promising outcomes. To predict student grades in
diverse outcomes requires collecting information from a II. RELATED WORK
variety of sources. However, a significant challenge persists
Jishan et al. [7] employed SMOTE approaches to enhance
in predictive modelling for unbalanced datasets. Addressing
the accuracy of learners’ absolute grade prediction. A number
this challenge has become a focal point, necessitating further
of classification methods including Decision Tree (DT),
exploration and study. This is because unbalanced datasets
Naive Bayes (NB), and Neural Network (NN), were used to
are skewed towards one class, making it difficult for the
divide learners’ grades into five groups: A, B, C, D, and F.
model to learn to predict the minority class. Jishan et al. [7]
They found that using SMOTE, NN and NB exceeded
attempted to address this issue by employing oversampling
different approaches with the same elevated accuracy of 75%.
techniques. They specifically utilized the Synthetic Minority
However, as compared to NN, NB performed better since the
OverSampling Technique (SMOTE) to handle unbalanced
best period to apply the prediction instances is immediate.
datasets and enhance the experimental outcomes of their
Nowadays, educational settings can create a large amount
study.
and diversity of data, such as those linked to students’ school
In this paper, we investigate the power of machine learning
records, evaluation files, curriculum reports, and records of
in predicting students’ expected Grade Point Average (GPA).
e-mail contacts between students and professors. Educational
In order to perform this task we firstly collected a dataset of
environments create a lot of data, which is useful for decision-
students graduated in a range of disciplines from the largest
making and improving the learning process. This can be
university in Saudi Arabia, King Saud University. The total
achieved through analysis of student data, and their behaviour,
number of students exceeded 12,499 student, number of
contentment, and performance [8]. Data Mining (DM)
graded courses is 766,278 grades. The dataset include both
approaches may be used to extract information from this data
male and female. Finally, practical considerations had to be
and, as a result, improve the quality of education [1].
taken into account during the data processing stage, such as

doi: 10.18178/ijiet.2024.14.2.2050 287

 International Journal of Information and Education Technology, Vol. 14, No. 2, 2024

Polyzou and Karypis [2] established a strategy for changes in the structure and format of the historical grade
forecasting future course grades acquired from the University information.
of Minnesota. The findings showed that Matrix Factorization Several works can be found in the literature which identify
(MF) and Linear Regression (LinReg) achieved better than various educational issues. Depending on the eventual user’s
current conventional approaches based on the proposed perspective, these works have varied goals (students,
methodologies. The author also discovered that using a instructors, administrators, or other stakeholders). Some of
course-specific dataset can help forecast future course grades these intended works will be presented in the sequel. The
with greater accuracy. scope of this study does not allow for a comprehensive
Another study used MF, Collaborative Filtering (CF), and examination of these works; nevertheless, further information
Restricted Boltzmann Machines (RBM) approaches to may be found in [11, 12].
predict student grades in various courses using 225 actual Iam-On and Boongoen [13] have addressed a different
data from undergraduate students [3]. They conclude that educational issue, concentrating on a critical issue in higher
when compared to MF, utilizing CF did not have great education: student dropout. Early discovery of susceptible
predictive power, specifically when the dataset is sparse. students, as described in this paper, can lead to the success of
However, their observations showed that the suggested RBM any engagement approach. Academic and administrative help
delivers effective instruction and greater prediction accuracy would be offered to at-risk students to boost their chances of
than CF and MF. finishing the course. Based on students’ pre-university traits,
Another research [4] created a predictive model for admission data, and first academic performance at university,
predicting students’ exam results in the curriculum at the the proposed work seeks to reveal intriguing patterns that
beginning of the semester. They used Waikato Environment might help forecast students’ performance and dropout. The
for Knowledge Analysis (WEKA) to compare eleven authors presented a new development in the social approach
machine learning algorithms in five categories. They used to increase the accuracy of traditional classifiers and, as a
attribute selection for details preparation to decrease high result, optimize students’ attention while also proposing
dimensionality and imbalanced data. The authors used courses depending on their progress. With the advancement
SMOTE to equalize the instance’s distribution in three and widely usage of on-line education, machine learning is
distinct classes. playing a crucial role in improving the quality of teaching.
Al-Barrak et al. [5] discovered classification methods to The education sector, in general, could benefit in different
predict the Grade Point Average (GPA) of students based on aspects [14] among them, which are: 1—students expected
their grades in preceding studies using the DT. They used 236 performance, 2—students best suitable field of study, 3—
students who completed from King Saud University’s student future career, and in many cases 4—the teachers
Computer Science College in 2012. They discovered the DT performance. The recent advancement of E-learning and the
classification algorithm that may detect early indicators and widely acceptance of Learning Management Systems (LMS)
extract valuable information for students based on their GPA, give deep insight into the date collected for the learning
which can help them to improve their performance. activities, providing more detailed information than simply
Abana [6] used several DT algorithms to predict a student’s the grades and gender of the students. A deep analysis of the
grade performance. Cross-checking was utilized for performance of e-learning systems is now with the help of
accessing the prediction model’s interpretation. According to machine learning is more accredited and accessible.
the statistics, Random Tree (RT) had a maximum accuracy of In [15], the authors utilized a plethora of studies in
75.188 %, which was superior to other algorithms. Educational Data Mining (EDM) for predicting a student’s
The prediction model’s accuracy may be enhanced by academic success at graduation time, investigating which of
increasing the experiment number and characteristics of the the individual course grades or grade averages is more
dataset. relevant for predicting student graduation academic
In Ahmad et al.’s [9] research, they established a performance. Although both types of data are
framework for forecasting student academic achievement. interchangeably used in the literature, there is no study
The study employed 399 records of students only from the comparing the performance of EDM models using grade
departmental registry throughout the courses of eight years of averages vs. individual course grades. It is unknown when
admissions, which included student demographics, past and how to use these two college performance representations
academic histories, and family background information. In to attain best predictive power. To elucidate this matter, a
comparison to DT and NB, the Rule-Based (PART) model comprehensive set of experiments were conducted on the
was shown to be the most accurate, with 71.3 % accuracy. recent student data compiled from the second author’s college.
However, due to the incomplete and incorrect values in the
dataset, adopting a limited sample size affected the accuracy III. RESEARCH METHODOLOGY
of this study. Generally, in machine learning research, two phases are
From 2006 through 2015, Anderson [10] conducted conducted. First, preparing and pre-processing the data.
experimental research on 683 students at California State Second, applying machine learning algorithms in order to
University’s Craig School of Business, using a machine figure out the one which has the best performance among
learning approach system to predict academic achievement. them.
The best classifier, according to the study, is Support Vector
Machine (SVM). It regularly beats a simple average strategy A. Data Description
that optimizes each data class with the lowest error rate. The The data under investigation are the official records of
outcome for a large dataset might be different due to huge more than 12,000 students graduated from King Saud

288
 International Journal of Information and Education Technology, Vol. 14, No. 2, 2024

University in two colleges (College of Engineering and course code as shown in Fig. 2. Specifically, we selected
College of Computer) and 35 departments as shown in Table courses with a level 1 or 2 as these courses are typically taken
1. For each graduate student the dataset consist of grades of in the early stages of the program and are more likely to be
all completed courses based upon an official curriculum. predictive of overall academic performance.
Graduated GPA is also available for each student. Moreover,
transferred courses are also highlighted and the student Converting Data to one Table
gender is also identified. Table 2 shows all features contained
within the dataset.
Finding and Selecting common
Table 1. Dataset description courses among all students
Variable Value
Total students 12,499
Studied semesters 73 Computing average grades for
Years 20 courses taken more than once
No of Departments 35
No of Graded Courses 766,278
Finding and Selecting the
Table 2. Dataset features courses based on their levels
Feature Type Example
Student ID Numerical 222197475
Gender Binary Female, Male
A table with selected courses
College Nominal Engineering fed to ML development step
Department Nominal Electrical Engineering
Specialization Nominal Power Fig. 1. Data preprocessing steps.
The university
Nominal KFUPM
transferred from
Cumulative GPA Numerical 3.5

B. Analysis
We have analyzed in two categories. First, we used GPAs
obtained by students in courses as predictors of the final GPA.
We utilized courses of level one and two given the fact that it Fig. 2. Criteria of the level of the course
is these courses are usually taken by students early on in their
academic plan. We will show later how we determine the E. Machine Learning Model Development
course level. Second, we studied the effect of an accumulated As we identified in the previous section, changes of
semester’s GPA in predicting the final GPA. We started with courses names over the years, adding new courses, and
one semester and added up one more at a time until we removing others cause some inconsistency in data. Therefore,
reached the fourth semester. during the preprocessing step, and before applying machine
C. Data Preprocessing and Cleansing learning algorithms, we selected departments with the fewest
The original dataset is composed of two large files. The changes so that more courses are used as features in
first file has all the metadata of all students, and the second predicting the final GPA.
file has all graded courses. We construct one table for each To predict the final GPA with a high accuracy, we
academic department. Each row in that table represent one evaluated different machine learning regression algorithms
student, and each column represent a graded course. including ensemble techniques and deep learning methods.
These algorithms are:
D. Feature Selection  Linear Regression: The linear regression models a
Due to the many changes in the curriculum in almost all relationship between single or multiple explanatory
departments, different courses have been taken by the variables (features) and a target variable. The equation of
students. This complicates the selection for our ML models. linear regression with multiple explanatory variables is
In order to solve this, we selected only the common courses defined as follows:
taken by all the students and ignore other courses.
y = w0x0+ w1x1+ ... + wmxm (1)
In this study, we have selected the electrical engineering
program to be the dataset used in this analysis. This is mainly Here w0 is the y axis intercept with x0 = 1. The objective of
due to the consistency of the curriculum over the studied the model solution is to find the best line fit that relate
period. We found 34 common courses that have been taken dependent variables to the output (target) value.
by all students over all of the examined semesters. If a student  Random Forest Regression: The random forest algorithm
took a course more than once, we computed the grade average. is a combination of multiple decision trees. Due to the
Fig. 1 shows the steps taken during the data preprocessing randomness that helps reduce model variance, a random
stage. To be able to predict in an early stage, we further forest typically achieves superior generalization
limited the selected features by choosing a lower course level. performance than an individual decision tree.
We developed a criteria based on the course code to Additionally, random forests are less sensitive to
determine the course level. We chose the course level based anomalies in the dataset and do not require extensive
on the most significant digit in the numerical part of the parameter tuning compared to other machine learning

289
 International Journal of Information and Education Technology, Vol. 14, No. 2, 2024

models . Typically, the only parameter that requires exploratory phase, we computed the correlation coefficients
optimization? in random forests is the number of trees in between individual courses. Achieved GPAs courses A and B
the ensemble. in Fig. 3 show linearity with the final GPA in contrast with
 Bagging Regression: The bagging algorithm is an GPAs of courses D and C where there is no linearity. This is
ensemble technique similar to random forest. However, an interesting finding showing the importance of examining
rather than using the same training set to model the the skill sets of each student before choosing their field of
individual classifiers in the ensemble, bootstrap samples study.
(random samples with replacement) are drawn from the
initial training set, which is why bagging is also known as
bootstrap aggregating.
 Adaptive Boosting Regression: In boosting, the ensemble
is comprised of extremely simplistic base models, also
referred to as weak learners, that have a marginal
performance advantage over random guessing. A typical
illustration of a poor learner is a decision tree trunk. The
central idea of boosting is to focus on training samples
that are difficult to predict, i.e., to let weak learners learn
from misclassified training samples to enhance the
ensemble’s performance. In contrast to bagging, the initial
formulation of boosting, the algorithm utilizes random
subsets of training samples drawn without replacement
from the training dataset.
 Gradient Boosting Regression: Gradient Boosting Fig. 3. Achieved GPA versus final GPA of students.
Regression is similar to Adaptive Boosting Regression in
terms of using weak learners. However, In Adaptive A. Prediction by Selected Courses
Boosting Regression, shift is done by up-weighting Table 3 shows the MAE results of the Linear Regression
observations that were mispredicted before where Model for different selected features (courses) based on the
Gradient Boost identifies difficult observations by large level of the course. As we mentioned earlier, it is better to
residuals computed in the previous iterations. choose features based on a lower level to predict the final
 Deep Learning: Deep Learning is based on multilayer results of the students in their early semesters.
artificial neural networks which consists of three layers:
input layer, hidden layers, and output layer. The number Table 3. Linear regression MAE results based on the selected courses level
of hidden layers between the input and output layer is No. of Selected
Level MAE
Courses(Features)
configurable and they are created as a hyper-parameter to
First 8 0.37
the problem that is to be solved. In this paper, we Second 15 0.21
constructed an architecture for our deep neural networks Third 28 0.12
which consist of one input layer, multiple hidden layers, Fourth 34 0.12
and one output layer. The total size of the electrical
engineering dataset used in the above algorithms is 231 As shown in the Table 3 the MAE decreases if meaningful
entries. We allocated 70% for the training dataset, and the features are added before applying the machine learning
remaining for testing. mode. However, after Level 3, the MAE does not decrease
To evaluate the performance of the algorithms, two metrics since the added features do not add any improvement to the
are computed: final prediction. Level 2 has a MAE of 0.21 and has a good
 Mean Absolute Error (MAE): MAE is defined as follows: warning time ahead before later semesters. Therefore, the
remaining results in this section are based on this level.
∑𝑛
𝑖=1 |𝑦𝑖 −𝑥𝑖 |
𝑀𝐴𝐸 = (2) The predicted and actual final GPA values for the first 15
𝑛
students are shown in Fig. 4. The predicted values are
where yi is the predicted value and xi is the actual value. The remarkably close to the actual value, as indicated in the graph
number of samples in the dataset is n. MAE is commonly with a MAE of 0.21 as we discussed later.
used to measure the performance of regression models.
 Accuracy: Accuracy is defined as follows:

𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (3)
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑆𝑎𝑚𝑝𝑙𝑒𝑠

This metric is often used to measure the performance of

supervised models.

IV. RESULTS AND DISCUSSIONS

We conducted our prediction analysis in two categories:
the first by using selected courses (features), and the second
category by using semester GPAs. During the data Fig. 4. Predicted values vs Actual values.

290
 International Journal of Information and Education Technology, Vol. 14, No. 2, 2024

To reach a low MAE, we evaluated several regression As shown in the table, we restrict the selected features until
algorithms. Table 4 shows the evaluated algorithms and MAE the fourth semester so that the prediction time is suitable to
values obtained for each algorithm. warn students in an early stage of their academic journey.
However, the MAE of the first method, using selected courses
Table 4. Evaluated ML regression algorithms
is better for predicting the final GPAs.
Algorithm MAE
Linear Regression 0.21
Random Forest Regression 0.27 V. CONCLUSION
Bagging Regression 0.21 In this work, we applied machine learning and deep
Gradient Boosting Regression 0.24
Adaptive Boosting Regression 0.22
learning algorithms to a dataset of undergraduate student
records collected from king Saud university, Riyadh. The
Per the above evaluation, Linear regression and Bagging dataset covers twenty years and seventy three semesters, and
regression have the best performance with a MAE value of therefore spans multiple academic plans. We first chose a
0.21. This is due to the high linearity between the target value department with fewer changes to its academic plan and then
and the attributes (courses) in the dataset. Tree based restricted selected features to the courses that are taken by all
regressors perform better than linear regressors if a non-linear students. To reach the best performance, linear regression,
relationship exists. ensemble techniques, and deep neural networks were
To complete our analysis, we applied Deep Learning evaluated and compared. The linear regression and bagging
technique to our dataset. We created a deep neural network of algorithms have the best performance with MAE of 0.2.
input layer, two internal layer of size 64, and output layer. Finally, we tested whether the performance was improved if
The Deep Learning model has the worst performance the dataset was analyzed using supervised learning with
analysis with MAE 0.49 as shown in Fig. 5. definite categories. This second approach did not exceed the
linear regression, resulting in an accuracy of 84%. We finally
complemented our analysis by predicting by the GPA per
semester. We started at the first semester and add the next
semesters to the selected features till the fourth semester.

CONFLICT OF INTEREST
The authors declare no conflict of interest.

AUTHOR CONTRIBUTIONS
Ibrahim Alnomay conducted the research, did the literature
review, and developed the used methodologies. Abdullah
Alfadhly and Aali alqarni wrote the analysis code, developed
Fig. 5. Loss values of DL algorithm. and tested the machine learning and deep learning algorithms.
All authors wrote the paper text, reviewed it, and approved
Finally, we converted the problem from regression to the final version.
supervised Machine Learning by predicting if the final
student grade fall into specific categories. The categories are REFERENCES
Fair, Good, Very Good, Excellent. We used Decision Tree [1] M. Ezz and A. Elshenawy, “Adaptive recommendation system using
Classifier algorithm, and get an accuracy of 84%. The chosen machine learning algorithms for predicting student’s best academic
max depth of the tree is three. If we did not specify a max program,” Education and Information Technologies, vol. 25, no. 4, pp.
2733–2746, Jul. 2020.
depth, the algorithm will reach a perfect fit on the training set, [2] A. Polyzou and G. Karypis, “Grade prediction with course and student
but the performance is worse on the testing set, reaching an specific models,” Advances in Knowledge Discovery and Data Mining,
accuracy of 79%. Our result is comparable with the result in ser. Lecture Notes in Computer Science, Springer International
Publishing, 2016, pp. 89–101.
[16] where they used four categories as we did and achieved [3] Z. Iqbal, J. Qadir, A. N. Mian, and F. Kamiran, “Machine learning
an accuracy between 48% and 86% for tree-based classifier. based student grade prediction: A case study,” arXiv:1708.08744 [cs],
However, they used a larger dataset of 530 rows and 64 Aug. 2017, arXiv: 1708.08744.
[4] I. Khan, A. Al Sadiri, A. R. Ahmad, and N. Jabeur, “Tracking student
attributes. performance in introductory programming by means of machine
learning,” in Proc. 2019 4th MEC International Conference on Big
B. Prediction Using Semester GPAs
Data and Smart City (ICBDSC), Jan. 2019, pp. 1–6.
In this analysis category, we predict the student final GPA [5] M. A. Al-Barrak, M. Al-Razgan, and King Saud University, Saudi
by using the GPA of semesters. At first, we predict by only Arabia, “Predicting students final GPA using decision trees: A case
study,” International Journal of Information and Education
the first semester and see the results, then we add the GPA of Technology, vol. 6, no. 7, pp. 528–533, 2016.
the second semester to the selected features, and so on. Table [6] E. Abana, “A decision tree approach for predicting student grades in
5 shows the results of the evaluated performance metrics. research project using weka,” International Journal of Advanced
Computer Science and Applications, vol. 10, Jul. 2019.
Table 5. Prediction performance results using semester GPAs [7] S. T. Jishan, R. I. Rashu, N. Haque, and R. M. Rahman, “Improving
accuracy of students’ final grade prediction model using optimal equal
Semester MAE
width binning and synthetic minority oversampling technique,”
First 0.47 Decision Analytics, vol. 2, no. 1, p. 1, Mar. 2015.
Second 0.41 [8] M. W. Rodrigues, S. Isotani, and L. E. Zárate, “Educational Data
Third 0.36 Mining: A review of evaluation process in the e-learning,” Telematics
Fourth 0.31 and Informatics, vol. 35, no. 6, pp. 1701–1717, Sep. 2018.

291
 International Journal of Information and Education Technology, Vol. 14, No. 2, 2024

[9] F. Ahmad, N. H. Ismail, and A. A. Aziz, “The prediction of students’ International Journal of Machine Learning and Cybernetics, vol. 8, no.
academic performance using classification data mining techniques,” 2, pp. 497–510, Apr. 2017.
Applied Mathematical Sciences, vol. 9, pp. 6415–6426, 2015. [14] H. Luan and C.-C. Tsai, “A review of using machine learning
[10] T. Anderson and R. Anderson, “Applications of ma-chine learning to approaches for precision education,” Educational Technology &
student grade prediction in quantitative business courses,” Global Society, vol. 24, no. 1, pp. 250–266, 2021, International Forum of
Journal of Business Pedagogy, vol. 1, no. 3, pp. 13–23, Dec. 2017, Educational Technology & Society.
publisher: Institute for Global Business Research. [15] S. Liu, X. Wang, M. Liu, and J. Zhu, “Towards better analysis of
[11] B. Bakhshinategh, O. R. Zaiane, S. ElAtia, and D. Ipperciel, machine learning models: A visual analytics perspective,” Visual
“Educational data mining applications and tasks: A survey of the last Informatics, vol. 1, no. 1, pp. 48–56, Mar. 2017.
10 years,” Education and Information Technologies, vol. 23, no. 1, pp. [16] N. Alangari and R. Alturki, Predicting Students Final GPA Using 15
537–553, Jan. 2018. Classification Algorithms, p. 12.
[12] C. Romero and S. Ventura, “Educational data mining and learning
analytics: An updated survey,” WIREs Data Mining and Knowledge Copyright © 2024 by the authors. This is an open access article distributed
Discovery, vol. 10, no. 3, May 2020. under the Creative Commons Attribution License which permits unrestricted
[13] N. Iam-On and T. Boongoen, “Improved student dropout prediction in use, distribution, and reproduction in any medium, provided the original
Thai University using ensemble of mixed-type data clusterings,” work is properly cited (CC BY 4.0).

292