PLOS ONE

RESEARCH ARTICLE

Machine learning approach to student

performance prediction of online learning
Jing Wang ID1,2*, Yun Yu2
1 Jiangsu College of Finance and Accounting, Jiangsu, China, 2 Nanjing University of Science and
Technology, Jiangsu, China

* wangjingjscfa@163.com

a1111111111
a1111111111 Abstract
a1111111111
a1111111111 Student performance is crucial for addressing learning process problems and is also an
a1111111111 important factor in measuring learning outcomes. The ability to improve educational sys-
tems using data knowledge has driven the development of the field of educational data min-
ing research. Here, this paper proposes a machine learning method for the prediction of
student performance based on online learning. The critical thought is that eleven learning
OPEN ACCESS behavioral indicators are constructed according to online learning process, following that,
Citation: Wang J, Yu Y (2025) Machine learning
through analyzing the correlation between the eleven learning behavioral indicators and the
approach to student performance prediction of scores obtained by students online learning, we filter out those learning behavioral indica-
online learning. PLoS ONE 20(1): e0299018. tors that are weakly correlated with student scores, meanwhile, retain these learning behav-
https://doi.org/10.1371/journal.pone.0299018
ior indicators being strongly correlated with student scores, which are used as the
Editor: Israr Ullah, Virtual University of Pakistan, eigenvalue indicators. Finally, using the eigenvalue indicators to train the proposed logistic
PAKISTAN
regress model with Taylor expansion. Experimental results show that the proposed logistic
Received: September 16, 2023 regress model defeats against the comparative models in prediction ability. Results also
Accepted: February 3, 2024 indicate that there is a significant dependency between students’ initiative in learning and
Published: January 14, 2025
learning duration, nevertheless, learning duration has a significant effect on the prediction of
student performance.
Copyright: © 2025 Wang, Yu. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
I Introduction
Data Availability Statement: All relevant data are
in the manuscript and Supporting information files. Recently, the research being interested in learning behavioral analysis has become increasingly
strong. Predicting student performance through online learning analysis can visualize student
Funding: This work is supported by the 2023
General Project of Philosophy and Social Sciences
behavior [1–3], to assist teachers understand the trends in student learning behavior, and to
Research in Higher Education Institutions’ improve curriculum design and teaching quality.
Ideological and Political Special Project "Research Higher education institutions consider students’ academic performance as one of the most
on the Development and Optimization Strategy of important issues in providing high-quality education to students [4]. Understanding the
Micro Ideological and Political Work for College important factors that influence student performance is complex. Currently, the academic
Counselors in the Background of Fragmented
community has used various effective tools and approaches to address student performance
Information Age" (No. 2023SJSZ1068). The
funders had no role in study design, data collection
challenges [5–7]. In recent years, along with the continuous progress of technology in predict-
and analysis, decision to publish, or preparation of ing student performance, there is still a gap to be filled to utilize machine learning and data
the manuscript. mining methods to analyze and improve the accuracy of student performance. Many

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 1 / 15

PLOS ONE Student performance prediction of online learning

Competing interests: There are no conflict researchers have identified the factors that affect student performance [8]. However, compared
interests of the authors. The authors agree to to the final student score in the final exam [9], the most common factors depend on learning
publish this paper.
activities [10]. Therefore, we observe that predicting trends in student performance may be
one of the solutions to improve student performance [11].
The Education Data Mining (EDM) method is a solution that may have a potential impact
on supporting higher education managers in making data-driven decisions. EDM aims at uti-
lizing new capabilities in data processing and the maturity of data mining algorithms to
enhance the learning process and transform existing information into knowledge [12]. EDM
analyzes educational data (such as student information, educational records, exam scores, par-
ticipation in online activities and classroom records, etc.) to develop models to improve learn-
ing experiences and institutional effectiveness [13]. Since EDM must discover knowledge from
data stored in different formats and granularity levels from multiple sources (such as enroll-
ment systems, registration systems, learning management systems, etc.), each issue requires
specific handling. Traditional data mining techniques cannot handle these issues, therefore,
the knowledge discovery process requires more advanced data mining methods. EDM applies
data mining, statistical methods, and machine learning (decision trees, neural networks, naive
Bayes, K-nearest neighbors, etc.) to explore large-scale data generated by educational organiza-
tions, in order to better understand the ongoing process.
Researchers have applied EDM methods to curriculum planning and student enrollment
prediction [14], enhancing the understanding of the learning process, and examining the
chances of course success. Various EDM methods have been applied to forecast students’
behavior [15], which provide feedback and suggestions to students [16] and determine stu-
dents’ profile in self-regulated learning [17, 18]. The EDM method helps teachers identify stu-
dents at risk and develop corrective strategies to reduce dropout rates [19, 20] and improve
students’ graduation rates [21]. The goal of all these studies is to improve students’ perfor-
mance. Because of this, most research in this field is focused on modeling students’ perfor-
mance prediction [22, 23].
Discovering hidden patterns and predicting trends in a large scale of the data might be a
potential method to be beneficial for the field of the education [24]. Predictive analysis has
been used to address several educational fields, including student performance, dropout pre-
diction, academic warning systems, and course selection [25]. In addition, predictive analysis
applying in predicting student performance has increased in recent years [26].

Motivation
The prediction of student performance can assist students to improve their grades. Many pre-
vious studies have proposed that machine learning approaches show potential ascendency in
the prediction of student performance. However, for the prediction of student performance, it
is difficult to find relevant work on mechanisms to associate the student performance with
learning behavior [27]. Hence, the research goal in this work is to forecast student perfor-
mance according to online learning behavior, and then to analyze the relations between stu-
dent performance and learning behavior. To finish the goal, here, we firstly constructed eleven
learning behavioral indicators based on the online learning process. Through analyzing the
correlations between these learning behavioral indicators and the scores obtained by students
online learning, we found out learning behavioral indicators that are strong correlated with
the scores, following that, the eigenvalue set consists of these learning behavioral indicators
that are strong correlated with the scores. Finally, the proposed logistic regress model was
trained by the eigenvalue set. Using the trained well logistic regress model to predict student
performance.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 2 / 15

PLOS ONE Student performance prediction of online learning

Contributions
We summarized the main contributions in this work. As follows
(1) We obtain eigenvalue indicators from the constructed learning behavioral indicators
through analyzing the correlation of the both, and then proposed a logistic regress model
with Taylor expansion. Using the eigenvalue indicators to train the logistic regress model
for the prediction of student performance, instead of using the original data to train it.
(2) We find that there is a significant dependency between students’ initiative in learning and
learning duration. However, learning duration has a significant effect on the prediction of
student performance.
This paper is arranged as follows. We summarized the related works in Section II. The
learning behavioral indicators were constructed and the logistic regress model was proposed
in Section III. Experimental settings and results were described in Section IV and Section V,
respectively. Section VI discussed the results, and Section VII drew a conclusion and directs
future work.

II Related work
Some efforts regarding the prediction of student performance have been gain, for instance,
Conijn et al. [28] predict student performance by using multi-level and standard regressions.
Whereas, due to the differences between course data, it is difficult to draw broad conclusions
about the online behavior of students with potential risks. The [29] proposed a convolutional
neural network for predicting student performance, which of the results show that the predic-
tion is successful. This work utilizes traditional and simple features to establish a student per-
formance prediction model for the prediction of student performance. Similarly, the
machined learning method implemented in [30]. Maurya et al. [31] provide a supervised
machine learning classifiers and Asalm et al. [32] apply a deep learning model for student per-
formance prediction. But the deep learning model is solely tested on two datasets. And the
deep learning model implemented by the [7]. Due to the ability to provide accurate and reli-
able results, deep learning approaches have become a popular strategy for predicting student
behavior. Such as, the method proposed in [33–37].
Logistic regress is called cost function, which uses logical functions as representations of
mathematical models. This model performs good contextual analysis on classified data to
understand the relationships between variables [38]. For example, a mixed regress model [39]
is proposed to optimize the accuracy of student performance prediction, which can predict
qualitative values of various factors related to the obtained student grades. However, it is hard
to confirm the reliability of the model. Ahmed et al. [40] proposed an approach using regress
thought to predict student performance. Together, several single classification algorithms are
employed as base classifiers [41], so as to improve the prediction of student performance.
However, the base classifiers need to implement optimization techniques to search parameters
and configuration in the classification algorithms. Similarly, the logistic regress model in [42].
Additionally, machine learning approaches effectively forecast student performance. For
instance, Moreno et al. [43] utilized Waston machine learning approach to derive predictors
for the forecast of student’s final grades based on online university setting. The predictors
focus on investigating student performance. Qunn et al. [44] employed a learning model using
Moodle data to forecast student’s academic performance for a blended education setting, thus
forecasting that whether the student would pass or fail in academic examination. The forecast
accuracy to the learning model reaches 92.2% in forecasting academic grade of students. Due

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 3 / 15

PLOS ONE Student performance prediction of online learning

Fig 1. Graphical illustrations of the method. The process contains four stages.
https://doi.org/10.1371/journal.pone.0299018.g001

to relying on Moodle data, using Moodle logs in the previous ten weeks can forecast those fail-
ing students, accurately, but unfortunately, using those in the first six weeks to suffer frustra-
tion in prediction. Similarly, the [45] utilized Moodle data to the prediction of student
performance. Qiu et al [46] developed the behavior classification-based e-learning perfor-
mance (BCEP) machine learning model to achieve the prediction of student’s learning perfor-
mance. BCEP model shows significant ascendency in forecasting because of combining feature
fusion with behavioral data, however, BCEP model relied on empirical values. Certainly, also
including the decision tree implemented in [47], and support vector machine model in [48]
and Naïve Bayes model in [48], etc.

III Methodology
A. Overall scheme
Fig 1 unveils the overall scheme of the method, which involves four stages. In first stage, i.e.,
data collection. Our goal is the prediction of student performance based on online learning,
therefore, the data is collected from online learning platform. The collected data is original
from diverse type on the online learning platform, such as relational type and non-relational
type, moreover, the data might hide incomplete and anomalous values. Consequently, there
must pretreat the data based on the learning behavior indicators constructed by us. By doing
so, it can create a condition for the performing of the second stage.
Learning behavior analysis, i.e., the second stage, which classifies students according to cer-
tain standards. The purpose of this stage is to compare the learning behaviors for different
types of students, thus analyzing their behavior characteristics. To identify whether behavioral
indicators are related to the outcome, we analyzed the correlation between learning behavioral
indicators and online learning. If the analysis results are no correlation, the behavioral indica-
tors are discarded. Instead, if they are relevant, the behavioral indicators are retained as the
eigenvalues.
The task in the third stage is behavior modeling. We constructed a logistic regress model,
which is trained by the eigenvalues. The fourth stage is the prediction of student performance
using our model.

B. Analysis of learning behavior

Online learning shows multiple diversity, that is, behavior indicators based on online learning
are multiple dimensions. Based on this, we took account into eleven learning behavior indica-
tors, illustrated in Table 1. These learning behavior indicators are described as follows.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 4 / 15

PLOS ONE Student performance prediction of online learning

Table 1. Behavior indicators for online learning.

Learning process Illustrations Indicators ID
Preparation stage Number of views Course introduction CI-N
Number of registers Course register CR-N
Number of attendances Course login CL-N
Major learning behavior Monitor the learning time Resource monitoring time RM-T
Resource utilization Resource utilization efficiency RU-E
Number of repeated views for resource Repeated views of resources RV-N
Number of repeated learning after finishing course Repeated learning of resources RL-N
Degree of duplicate monitoring of resources Resource density utilization RD-U
Secondary learning behavior Number of forum browse Forum browsing FB-N
Number of forum post Forum posting FP-N
Number of forum reply Forum replying FR-N
https://doi.org/10.1371/journal.pone.0299018.t001

The learning process consists of three parts in this work, i.e., preparation stage, major learn-
ing behavior and secondary learning behavior. In preparation stage, we took account into the
number of viewing course introduction, that of course register and that of course login. Fol-
lowing that, major learning behavior and secondary learning behavior are constructed, among
them, major learning behavior, which is treated as a critical monitoring factor, consists of the
five behavior indicators, including the monitoring of learning time for students, denoted as
RM-T, and resource utilization, namely RU-E, which is calculated through the time spent by
students on learning resources divided by the total time spent on learning resources (recom-
mended time). As well as, the number of repeated viewing resource, namely RV-N, and the
number of repeated learning after finishing course, i.e., RL-N. Resource density utilization,
i.e., RD-U, refers to the time of view resource divided by the time difference between last view
resource and first view, reflecting students’ concentration. As for the secondary learning
behavior, it is regarded as a learning interaction behavior, containing the number of browsing
learning forum FB-N, that of posting learning forum FP-N and that of replying learning
forum FR-N.
We analyzed the correlation between the eleven learning behavioral indicators in Table 1
and the course of average score achieved by students in Table 2, through using the SPSS tools.
Following that, we filter out those learning behavioral indicators with weak correlation, instead,
those with higher correlation should be retained. The filtered details are that those learning
behavioral indicators with the correlation coefficient below 0.6 are filtered out, otherwise, they
are retained. Finally, those retained learning behavioral indicators are used as the eigenvalue
indicators affecting online learning. For the convenience of description, in subsequent sections,
we denote those retained learning behavioral indicators as the eigenvalue indicators.

C. Behavioral modeling
We construct a logistic regress model according to the obtained eigenvalue indicators. Having

hðxÞ ¼ gðy1 x1 þ y2 x2 þ . . . þ yi xi þ . . . þ yn xn Þ ¼ gðyT xÞ ð1Þ

Where x1, x2,. . ., xi,. . ., xn is the eigenvalue indicators. θ1, θ2,. . ., θi,. . ., θn is the weight corre-
sponding to the eigenvalue indicators. The value of h(x) denotes the probability of taking 1.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 5 / 15

PLOS ONE Student performance prediction of online learning

Table 2. Dataset descriptions.

# Course name Requirement learning time Learning unit quantity Score for a learning unit Average score Learning flag
C1 Operation System 60 hours 10
C2 Java Programing 60 hours 10
C3 Python Programing 60 hours 10
C4 C++ Programing 60 hours 10
C5 C Programing 60 hours 10
C6 C# Programing 60 hours 10
C7 Probability Theory 30 hours 6
C8 Graph Theory 30 hours 6
C9 Calculus 30 hours 6
C10 Optimal Theory 30 hours 6
C11 Linear Algebra 30 hours 6
C12 Software Engineering 40 hours 8
C13 Computer Network 40 hours 8
C14 Artificial Intelligence 40 hours 8
C15 Business English 48 hours 12
https://doi.org/10.1371/journal.pone.0299018.t002

The joint density function of n samples can be calculated, having

yðiÞ 1 yðiÞ
Lðyjx; yÞ ¼ Pni¼1 ðhðxÞÞ ð1 hðxÞÞ ð2Þ

To predict accurately the results, we introduced penalized log-likelihood. That is, Eq (2) is
converted into Eq (3). As follows
lX m
L∗ ¼ Lðyjx; yÞ b2 ð3Þ
2 j¼1 j

To simplify, taking the logarithm of Eq (3), as follows

X
n X
n
lX m
L∗ ¼ yi log pi þ ð1 yi Þlogð1 pi Þ b2 ð4Þ
i¼1 i¼1
2 j¼1 j

Where λ is the penalty item. The larger the values of λ are, the stronger the effects are. yi is the
i-th eigenvalue indicator. pi is the probability that yi = 1. β1, β2,. . ., βi,. . ., βm are parameters
which could be estimated by maximum likelihood criterion. For the estimate of βi, we referred
the [49]. According to Eq (4), we can obtain the predictor, as follows
pi
Y ¼ log ð5Þ
1 pi

To solve the pi, let y = [y1, y2,.., yn]T, p = [p1, p2,.., pn]T, β = [β1, β2,. . ., βi,. . ., βm]T and let us
take the derivative of L* with respect to βi. Having that
@L∗=@b ¼ 0 ) yT ðy pÞ ¼ lb ð6Þ

Obviously, Eq (6) is non-linear due to the non-linear relations between p and β. To obtain a
linear equation, let us take the first order Taylor expansion at pi.
X
m
@p
pi � p~i þ i
ðbj ~Þ
b ð7Þ
j
j¼1
@bj

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 6 / 15

PLOS ONE Student performance prediction of online learning

~ are an approximate solution. For the p~ , we use Eq (8) to calculate, having that
Where p~i and b j i

X
n
p~i ¼ bj xi ð8Þ
j¼1

~ according to βi.
At the beginning of iterations, we can initialize a starting value about b j

Hence, the value of pi can be calculated.

D. Algorithm implementation
The model algorithm is as shown in Algorithm 1. The input and output are the eleven learning
behavioral indicators LBI(k), the average score of j-th course AS(j). Firstly, parameters are ini-
tialized in Step 1. Following that, the procedure between Step 2 and Step 12 displays the selec-
tion of the eigenvalue indicators. Through calculating correlation coefficient CR(k,j) between
LBI(k) and AS(j), we can obtain eigenvalue indicator LBIðkÞ,^ illustrated in Step 4 to Step 7.
^ in Step 13, we utilize LBIðkÞ
After successfully obtaining eigenvalue indicator LBIðkÞ, ^ con-
^ ^
struct a matric Mð300; k�15Þ with 300 rows and k�15 columns. Here, in the matric
^
Mð300; k�15Þ, the row is the number of students, and the column is constructed by both the
number of eigenvalue indicators and that of courses. In Step 14 and Step 15, the matric
^
Mð300; k�15Þ is randomly divided into training set Train(M) and testing set Test(M). The
procedure between Step 16 and Step 23 illustrates the training of logistic regress model h(x).
Using training set Train(M) to train h(x), once maximum training accuracy is obtained, the
training is terminated. Meanwhile, current trained logistic regress model h(x, Train(M)).
Finally, using testing data Test(M) to verify the trained h(x, Train(M)), prediction accuracy is
outputted, illustrated in Step 24 and Step 25.
Algorithm 1. Model algorithm
input LBI(k), AS(j)
output Prediction accuracy
1 Initialize parameter;
2 For k = 1 to 11:
3 For j = 1 to 15:
4 Calculating the correlation between LBI(k) and AS(j);
5 Obtaining the correlation coefficient CR(k,j);
6 If CR(k, j) > = 0.6 Then:
7 saving eigenvalue indicator LBIðk ^ Þ;
8 End If
9 j++;
10 End For
11 k++;
12 End For
13 Using LBIðk ^ Þ to construct a matric Mð300; k
^ �15Þ with 300 rows and
^
k �15 columns;
14 Obtaining training set Train(M) through randomly selecting 80%
^ �15Þ;
Mð300; k
15 Obtaining testing set TestðMÞ ¼ Mð300; k ^ �15Þ 80%∗Mð300; k^ �15Þ;
16 For i = 1 to i = Imax:
17 Using training set Train(M) to train logistic regress model h(x)
in Eq (1);
18 If current training accuracy = = maximum value True:
19 Saving the model h(x, Train(M));
20 Obtain current training accuracy;
21 Break;

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 7 / 15

PLOS ONE Student performance prediction of online learning

22 End If
23 End For
24 Using testing set Test(M) to verify h(x, Train(M));
25 Obtaining forecasted accuracy;

IV Experimental settings
A. Datasets
Experimental datasets are provided from the MOOC platform (https://www.icourse163.
org/). The datasets include 15 courses. We manually collected information for 15 courses,
and the course is lasted for 10 weeks, among them, 300 students enrolled in the course. We
used the average grade to characterize student’s performance, that is, we set the score to be
excellent or not as the dependent variable. If the score is greater than 90 points, it is treated
as excellent. Otherwise, it is treated as not excellent. The Table 2 lists the details of the 15
courses, where requirement learning time refers to the sum time fulfilling the correspond-
ing course. Learning unit quantity is the number of learning units included in a course.
Score for a learning unit indicates the testing grade obtained by a student after fulfilling a
learning unit. Average score is the average value of the testing grade for all learning units.
Learning flag includes two metrics Ex and NE, if average score is greater than 90 points,
learning flag is marked with Ex, otherwise, it is marked with NE. The dataset is illustrated
in Table 2.

B. Competitors and evaluated indicators

Apart from our logistic regress model, the logistic model [28], the mixed regress model [39], the
logistic model [42], decision tree model [47] and SVM model [48] are used as a comparison.
We chose the five comparative models to have a fair comparison. Please note that our model
and the five competitors, the same eigenvalues are used as the training set and testing set.
In addition, the metrics Accuracy, Precision, Recall and F1-score are used to evaluate the
predicted ability of these methods. We utilized the Python language to achieve the four algo-
rithms corresponding to the four models, unless other stating, the four algorithms were run on
the same experimental setting.

C. Experimental designs
We conducted three groups of experiments to verify the proposed method. As follows,
Experiment (I). Eigenvalue indicators selection. To obtain critical indicators impacting student
grade, we screened the eleven learning behavioral indicators by using correlation analysis
and clustering methods. Then, the results were compared.
Experiment (II). Comparisons of prediction performance. To verify the proposed logistical
regress model, we compared it with the five opponents on the eigenvalue indicators. Then,
the compared results were analyzed.
Experiment (III). Learning efficiency. To analyze the learning efficiency online learning, we
calculated the learning scores of 300 students for viewing courses on 10 weeks. Then, the
calculated results were analyzed.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 8 / 15

PLOS ONE Student performance prediction of online learning

V Results analysis
This section presents the experimental results. We discussed the result analysis from three
aspects, including eigenvalue selection, the prediction for student performance and students’
enthusiasm for active learning. The details are as follow.

A. Selections of eigenvalue indicators

We analyzed the relations between the eleven learning behavior indicators (i.e., the indicators
in Table 1) and student scores by SPSS tool, as shown in Table 3. The five Indicators CI-N,
CR-N, CL-N, RM-T and RL-N are no significant correlated with the learning scores. The three
indicators FB-N, FP-N and FR-N are weakly correlated with the learning scores. While for the
three indicators the number of repeated views for resource RV-N, resource utilization effi-
ciency RU-E and resource density utilization RD-U, they have a strong correlation with the
learning scores. Consequently, together, the three learning behavioral indicators RV-N, RU-E
and RD-U are used as the eigenvalue indicators, which are used to train and test our logistic
regress model and the four opponents.
We clustered the eleven learning behavior indicators by k-means clustering. These values
regarding the eleven learning behavior indicators are obtained via the 300 students learned 15
courses on ten weeks. According to the results analyzed by Table 3, the clustered results are
shown in Fig 2, where c = 3. In Fig 2, the clustering results of the eleven learning behavioral
indicators are significant from the perspective of correlations when c is equal to 3. Hence, it is
reasonable to choose the three learning behavioral indicators RV-N, RU-E, RD-U as the eigen-
value indicators from the view of the correlation.

B. Comparisons of predication performance

Now, we start to verify the predication ability of our logistic regress model. The eigenvalue set
was randomly divided into a training set and a testing set, among them, the 80% of the set was
used as the training set, and the rest 20% is testing set. The comparative results are given in
Table 4, showing that our logistic regress model defeats against the five opponents. Moreover,
our model can predict student performance more accurately.
The results in Table 4 are based on the eigenvalue indicators consisted of the three learning
behavioral indicators RV-N, RU-E and RD-U that are strong correlated with student scores.

Table 3. Correlations of learning behavioral indicators and student scores. Sign ** indicates that there is a signifi-
cant correlation at 0.05 level (two tailed).
Learning behavior indicators Pearson correlation Scores Number of samples
Sig(two tailed)
CI-N 0.036 0.302 300
CR-N 0.099 0.314 300
CL-N 0.036 0.322 300
RM-T 0.014 0.181 300
RL-N 0.069 0.283 300
FB-N 0.212 0.040 300
FP-N 0.247 0.032 300
FR-N 0.256 0.033 300
RV-N 0.781 ** 0.000 300
RU-E 0.701 ** 0.000 300
RD-U 0.708** 0.000 300
https://doi.org/10.1371/journal.pone.0299018.t003

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 9 / 15

PLOS ONE Student performance prediction of online learning

Fig 2. Clustering for the eleven learning behavioral indicators. The three indicators RV-N, RU-E and RD-U are
marked with red circles. The three indicators FB-N, FP-N and FR-N are marked with black circles. The five Indicators
CI-N, CR-N, CL-N, RM-T and RL-N are marked with green circles.
https://doi.org/10.1371/journal.pone.0299018.g002

To objectively assess the model, we supplemented the three learning behavioral indicators
FB-N, FP-N and FR-N that are weakly correlated with student scores. Together, the six learn-
ing behavioral indicators RV-N, RU-E, RD-U, FB-N, FP-N and FR-N are used as the eigen-
value set, which test the six models.
Table 5 unveils the prediction results, showing that our model is still better than the five
competitors in prediction capabilities. Compared these results in Table 5 to those in Table 4,
we find that the prediction results of the six models did not show significant changes. More-
over, the changes in the predicted results of using the three learning behavioral indicators and
using the six learning behavioral indicators, indicating that although the three learning behav-
ioral indicators FB-N, FP-N and FR-N were supplemented (i.e., six learning behavioral indica-
tors in total), there is a weak influence in predicting the results. Hence, these results confirm
each other with the results in Table 3.

Table 4. Predication results on three eigenvalue indicators. These results are averages of 100 times. The best results are highlighted.
Training set Testing set
Accuracy Precision Recall F1-score Accuracy Precision Recall F1-score
Our model 0.947 0.928 0.933 0.979 0.933 0.909 0.922 0.966
Logistic model [28] 0.930 0.922 0.894 0.955 0.912 0.900 0.927 0.873
Mixed regress model [39] 0.942 0.935 0.886 0.951 0.903 0.902 0.918 0.918
Logistic model [42] 0.937 0.912 0.872 0.967 0.909 0.903 0.909 0.903
Decision Tree [47] 0.911 0.922 0.909 0.918 0.882 0.906 0.937 0.946
SVM [48] 0.907 0.918 0.911 0.887 0.898 0.904 0.871 0.951
https://doi.org/10.1371/journal.pone.0299018.t004

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 10 / 15

PLOS ONE Student performance prediction of online learning

Table 5. Predication results on six eigenvalue indicators. These results are averages of 100 times. The best results are highlighted.
Training set Testing set
Accuracy Precision Recall F1-score Accuracy Precision Recall F1-score
Our model 0.959 0.894 0.937 0.916 0.933 0.901 0.917 0.959
Logistic model [28] 0.911 0.838 0.880 0.944 0.884 0.900 0.909 0.944
Mixed regress model [39] 0.912 0.847 0.893 0.902 0.869 0.901 0.906 0.941
Logistic model [42] 0.909 0.811 0.882 0.917 0.913 0.875 0.914 0.950
Decision Tree [47] 0.922 0.825 0.859 0.916 0.863 0.901 0.900 0.952
SVM model [48] 0.847 0.888 0.826 0.833 0.820 0.863 0.913 0.955
https://doi.org/10.1371/journal.pone.0299018.t005

C. Learning efficiency
Fig 3 unveils average viewing time of a week for the fifteen courses, showing that majority stu-
dents (i.e., 63.0% students) consume 60 to 180 minutes on the course viewing per week. And
14.9% students prefer to consume more than 180 minutes for the course viewing per week.
However, 19.3% students choose 30 to 59 minutes for the course viewing per week. Certainly,
rest of 2.8% spend less than 30 minutes for the course viewing per week, or not for the viewing.
These indicate that students exhibit certain proactive learning behaviors.
Fig 4 unveils the number of excellent students. As the viewing time increase, the number of
excellent students shows a slight downward trend. This implies that there is a significant
dependency between students’ initiative in the learning and learning duration. Noting that we
did not consider the difficulty of the course, i.e., the impact of course difficulty on students’
learning motivation. This is to objectively evaluate students’ initiative in the learning.

VI Discussions
A. Advantages
The predictor Θ in Eq (5) have ascendency forecast ability. In the process of deriving the pre-
dictor Θ, we took account into the first order Taylor expansion, illustrated in Eq (7). By doing
so, the predictor Θ can better approximate the original data. In fact, we sufficiently borrowed

Fig 3. Proportion distribution of viewing course time per week.

https://doi.org/10.1371/journal.pone.0299018.g003

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 11 / 15

PLOS ONE Student performance prediction of online learning

Fig 4. Relations between excellent students and viewing time. The score greater than 90 points is regarded as
excellent. Otherwise, not excellent.
https://doi.org/10.1371/journal.pone.0299018.g004

that the advantage Taylor expansion can approximate object functions. This is one advantage
of our method. Additionally, we did not impose any assumptions on the data distribution and
our model, which is another advantage. In summary, that is why our method won.

B. Limitations
Our goal is the student performance prediction for online learning, therefore, we constructed
the eleven learning behavioral indicators. However, many factors have an effect on learning
behavioral of students, such as, subjective factors, including students’ emotions. For those sub-
jective factors, in this work, they were not taken, instead, major objective factors were consid-
ered. From the perspective of data level, the constructed eleven learning behavioral indicators
own limitations. From the view of method level, using the eigenvalues indicators selected from
the eleven learning behavioral indicators to train our method, therefore, the construction of
the eleven learning behavioral indicators directly influences our forecast results. But please
note there are various learning behavioral indicators in real applications, consequently, it is
unrealistic to list all those indicators. That is why we chose major learning behavioral indica-
tors in this work.

C. Insights
Generally, learning behaviors between students exist a difference. Indeed, the characteristics of
learning behaviors to students can objectively reflect student performance (regarding the inter-
pretation, please see Section Introduction), therefore, there must be the correlation between
the characteristics of learning behaviors and student performance. However, this work used
learning behavioral indicators as specific manifestations to the characteristics of learning
behaviors. If the impact indicators are different (our eleven learning behavior indicators can
be considered as eleven different impact indicators), then their impact on the results may be
different. As such, based on this, here, we utilized the strength-weakness of the correlation to
explore the latent relationships between learning behavioral indicators and student perfor-
mance, assisting that we find those relative critical learning behavioral indicators impacting
student performance. Moreover, this also help to understand the underlying educational
dynamics.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 12 / 15

PLOS ONE Student performance prediction of online learning

VII Conclusion
Early prediction of students’ performance is helpful for teachers to determine which students
may perform poorly in final examination. Teachers can pay extra attention to those students,
meanwhile, take intervened measures to assist them. Timely intervened measures employed by
teachers can significantly reduce the number of failed students.
In this work, we constructed eleven learning behavioral indicators aiming at online leaning.
Based on the constructed eleven learning behavioral indicators, here proposed a logistic
regress model to forecast student performance. The critical thought is that the eigenvalue indi-
cators were chose by calculating the correlations of the eleven learning behavioral indicators
and student scores. Following that, the eigenvalue indicators are used to the training set of the
proposed logistic regress model. Finally, experimental results show that our model defeats
against the comparative models in predicted student performance. We indicate that we did
not impose any assumptions on the data distribution and the proposed method, therefore, our
method is suitable for the prediction of student performance in complex education environ-
ments. We also find that there is a significant dependency between students’ initiative in learn-
ing and learning duration. However, learning duration has a significant effect on the
prediction of student performance. We demonstrate that the constructed learning behavioral
indicators are reasonable, which can provide suggestions to promote students’ enthusiasm for
the learning.
Although we accurately predict student performance by the learning behavioral indicators,
many factors have an influence on student performance, such as subjective factors in the learn-
ing process. Therefore, in future work, we will look at exploring students’ subjective factors
impacting student performance.

Supporting information
S1 Dataset.
(TXT)

Author Contributions
Conceptualization: Jing Wang.
Funding acquisition: Jing Wang.
Methodology: Jing Wang.
Project administration: Yun Yu.
Resources: Yun Yu.
Software: Jing Wang, Yun Yu.
Supervision: Jing Wang.
Writing – original draft: Jing Wang.

References
1. Cazarez Rosa Leonor Ulloa, Garcı́a-Dı́az Noel and Equigua Leonel Soriano. Multi-layer Adaptive Fuzzy
Inference System for Predicting Student Performance in Online Higher Education. IEEE Latin America
Transactions. 2021; 19(1):98–106.
2. Alhazmi Essa and Sheneamer Abdullah. Early Predicting of Students Performance in Higher Education.
IEEE Access. 2023; 11:27579–27589.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 13 / 15

PLOS ONE Student performance prediction of online learning

3. Bujang Siti Dianah Abdul, Selamat Ali, Ibrahim Roliana, Krejcar Ondrej, Viedma Enrique Herrera, Fujita
Hamido and Md. Ghani Nor Azura. Multiclass Prediction Model for Student Grade Prediction Using
Machine Learning. IEEE Access. 2021; 9:95608–95621.
4. Francis B. K., Babu S. S. Predicting academic performance of students using a hybrid data mining
approach. J. Med. Syst. 2019; 43(6):1–15. https://doi.org/10.1007/s10916-019-1295-4 PMID:
31037484
5. Yağcı M. Educational data mining: Prediction of students’ academic performance using machine learn-
ing algorithms. Smart Learn. Environ. 2022; 9(1):1–19.
6. T. Le Quy, T. H. Nguyen, G. Friege, E. Ntoutsi. Evaluation of group fairness measures in Student perfor-
mance prediction problems. arXiv. 2022; 2208.10625.
7. X. Liu, L. Niu. A student performance predication approach based on multi-agent system and deep
learning. In: Proc. IEEE Int. Conf. Eng., Technol. Educ. (TALE); 2021. p. 681–688.
8. Alyahyan E., Düştegör D. Predicting academic success in higher education: Literature review and best
practices. Int. J. Educ. Technol. Higher Educ. 2020; 17(1):1–1.
9. Tatar A. E., Düştegör D. Prediction of academic performance at undergraduate graduation: Course
grades or grade point average. Appl. Sci. 2020; 10(14):1–15.
10. Moreno-Marcos P. M., Pong T.-C., Munoz-Merino P. J., Kloos C. D. AnalysisofthefactorsinfluencingLe-
arners’performancepredictionwith learning analytics. IEEE Access. 2020; 8:5264–5282, 2020.
11. Zhang Y., Yun Y., Dai H., Cui J., Shang X. Graphs regularized robust matrix factorization and its appli-
cation on student grade prediction. Appl. Sci. 2020; 10:1755.
12. Chalaris M., Gritzalis S., Maragoudakis M., Sgouropoulou C., Tsolakidis A. Improving quality of educa-
tional processes providing new knowledge using data mining techniques. Procedia-Soc. Behav. Sci.
2014; 147:390–397.
13. Zhang M., Fan J., Sharma A., Kukkar A. Data mining applications in university information management
system development. J. Intell. Syst. 2022; 31:207–220.
14. Vora D.R., Iyer K. EDM-survey of performance factors and algorithms applied. Int. J. Eng. Technol.
2018; 7:93–97.
15. El Mourabit, I., Jai-Andaloussi, S., Abghour, N. Educational Data Mining Techniques for Detecting
Undesirable Students’ Behaviors and Predicting Students’ Performance: A Comparative Study. In:
Advances on Smart and Soft Computing. Advances in Intelligent Systems and Computing; Saeed, F.,
Al-Hadhrami, T., Mohammed, E., Al-Sarem, M., Eds.; 2022; p.1399.
16. Juhaˇ nák L., Zounek J., Rohlı́ková L. Using process mining to analyze students’ quiz-taking behavior
patterns in a learning management system. Comput. Hum. Behav. 2019; 92:496–506.
17. Araka E.; Oboko R., Maina E., Gitonga R. Using Educational Data Mining Techniques to Identify Pro-
files in Self-Regulated Learning: An Empirical Evaluation. Int. Rev. Res. Open Distrib. Learn. 2022;
23:131–162.
18. Araka E., Oboko R., Maina E., Gitonga R.K. A Conceptual Educational Data Mining Model for Support-
ing Self-Regulated Learning in Online Learning Environments. In: Handbook of Research on Equity in
Computer Science in P-16 Education. 2021; p.278–292.
19. Amala Jayanthi, M., Shanthi, I.E. Role of Educational Data Mining in Student Learning Processes with
Sentiment Analysis: A Survey. In: Research Anthology on Interventions in Student Behavior and Mis-
conduct; Management Association. 2022; p.412-427.
20. Padhy, N., Mishra, D., Panigrahi, R. The survey of data mining applications and feature scope. 2022;
arXiv:1211.5723.
21. Trakunphutthirak R., Lee V.C.S. Application of Educational Data Mining Approach for Student Aca-
demic Performance Prediction Using Progressive Temporal Data. J. Educ. Comput. Res. 2022;
60:742–776.
22. Cruz M.E.L.T., Encarnacıon R.E. Analysis and Prediction of Students’ Academic Performance and
Employability Using Data Mining Techniques: A Research Travelogue. Eurasia Proc. Sci. Technol.
Eng. Math. 2021; 16:117–131.
23. Zhang Y., Yun Y., An R., Cui J., Dai H., Shang X. Educational Data Mining Techniques for Student Per-
formance Prediction: Method Review and Comparison Analysis. Front. Psychol. 2021; 12:698490.
https://doi.org/10.3389/fpsyg.2021.698490 PMID: 34950079
24. Aldowah H., Al Samarraie H., Fauzy W. M. Educational data mining and learning analytics for 21st cen-
tury higher education: A review and synthesis. Telematics Informat. 2019; 37:13–49.
25. Ang K. L.-M., Ge F. L., Seng K. P. Big educational data & analytics: Survey, architecture and chal-
lenges. IEEE Access. 2020; 8:116392–116414.

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 14 / 15

PLOS ONE Student performance prediction of online learning

26. A. Hellas, P. Ihantola, A. Petersen, V. V. Ajanovski, M. Gutica, T. Hynninen, et al. Predicting academic
performance: A systematic literature review. In: Proc. 23rd Annu. Conf. Innov. Technol. Comput. Sci.
Educ. 2018; p.175-199.
27. Abu Zohair L. M. Prediction of student’s performance by modelling small dataset size. Int. J. Educ.
Technol. Higher Educ. 2019; 16(1):1–8.
28. Conijn R., Snijders C., Kleingeld A., Matzat U. Predicting student performance from LMS data: A com-
parison of 17 blended courses using Moodle LMS. IEEE Trans. Learn. Technol. 2017; 10(1):17–29.
29. Dien T. T., Luu S. H., Thanh-Hai N., Thai-Nghe N. Deep learning with data transformation and factor
analysis for student performance prediction. Int. J. Adv. Comput. Sci. Appl. 2020; 11(8):1–11.
30. Ha D. T., Loan P. T. T., Giap C. N., Huong N. T. L. An empirical study for student academic performance
prediction using machine learning techniques. Int. J. Comput. Sci. Inf. Secur. (IJCSIS). 2020; 18(3):1–
8.
31. Maurya L. S., Hussain M. S., Singh S. Developing classifiers through machine learning algorithms for
student placement prediction based on academic performance. Appl. Artif. Intell. 2021; 35(6):403–420.
32. Aslam N., Khan I., Alamri L., Almuslim R. An improved early student’s academic performance prediction
using deep learning. Int. J. Emerg. Technol. Learn. 2021; 16(12):108–122.
33. Hussain S. Survey on current trends and techniques of data mining research. London Journal of
Research in Computer Science and Technology. 2017; 17(1):11.
34. Hussain, M., Hussain, S., Zhang, W., Zhu, W., Theodorou, P., & Abidi, S. M. R. Mining moodle data to
detect the inactive and low-performance students during the moodle course. In: Proceedings of the 2nd
International Conference on Big Data Research. 2018; p.133-140.
35. Hussain S. Educational data mining using R programming and R studio. Journal of applied and funda-
mental sciences. 2015; 1(1):45.
36. Gaftandzhieva S., Talukder A., Gohain N., Hussain S., Theodorou P., Salal Y. K., et al. Exploring online
activities to predict the final grade of student. Mathematics. 2022; 10(20):3758.
37. Hussain S., & Hazarika G. C. Educational data mining model using rattle. International Journal of
Advanced Computer Science and Applications. 2014; 5(6).
38. Hussain M., Zhu W., Zhang W., Abidi S. M. R., Ali S. Using machine learning to predict student difficul-
ties from learning session data. Artif. Intell. Rev. 2019; 52(1):381–407.
39. Alshanqiti A., Namoun A. Predicting student performance and its influential factors using hybrid regres-
sion and multi-label classification. IEEE Access. 2020; 8:203827–203844.
40. D. M. Ahmed, A. M. Abdulazeez, D. Q. Zeebaree, F. Y. H. Ahmed. Predicting university’s students per-
formance based on machine learning techniques. In: Proc. IEEE Int. Conf. Autom. Control Intell.
Syst.2021; p.276-281.
41. Evangelista E. D. L., Sy B. D. An approach for improved students’ performance prediction using homo-
geneous and heterogeneous ensemble methods. Int. J. Electr. Comput. Eng. 2022; 12(5):5226.
42. Yun Wu. Multi-task achievement prediction model based on grade change trend. 2020 IEEE Interna-
tional Conference on Information Technology, Big Data and Artificial Intelligence. 2020; p.1-8.
43. Moreno Ger P., Burgos D. Machine Learning and Student Activity to Predict Academic Grades in Online
Settings in Latam. In: Radical Solutions for Digital Transformation in Latin American Universities;
Springer: Singapore. 2021; p.243–257.
44. Quinn R.J., Gray G. Prediction of student academic performance using Moodle data from a Further
Education setting. Ir. J. Technol. Enhanc. Learn. 2020; 5:1.
45. Hussain, M.; Hussain, S., Zhang, W.; Zhu, W., Theodorou, P., Abidi, S.M.R. Mining moodle data to
detect the inactive and low-performance students during the moodle course. In: Proceedings of the 2nd
International Conference on Big Data Research. 2018; p.133-140.
46. Qiu F., Zhang G., Sheng X., Jiang L., Zhu L., Xiang Q., et al. Predicting students’ performance in e-
learning using learning process and behaviour data. Sci. Rep. 2020; 12:453.
47. Alhassan A., Zafar B., Mueen A. Predict students’ academic performance based on their assessment
grades and online activity data. Int. J. Adv. Comput. Sci. Appl. 2020; 11:4.
48. Ya gcı M. Educational data mining: Prediction of students’ academic performance using machine learn-
ing algorithms. SmartLearn. Environ. 2022; 9:11.
49. Shen Li, Tan Eng Chong. Dimension reduction-based penalized logistic regression for cancer classifica-
tion using microarray data. IEEE/ACM Transactions on Computational Biology and Bioinformatics.
2005; 2(2):166–175. https://doi.org/10.1109/TCBB.2005.22 PMID: 17044181

PLOS ONE | https://doi.org/10.1371/journal.pone.0299018 January 14, 2025 15 / 15