Received 10 October 2023, accepted 20 November 2023, date of publication 23 November 2023,

date of current version 29 November 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3335985

Student Performance Prediction Approach Based

on Educational Data Mining
ZILING CHEN , GANG CEN , YING WEI, AND ZIFEI LI
School of Information and Electronic Engineering, Zhejiang University of Science and Technology, Hangzhou 310000, China
Corresponding author: Gang Cen (gcen@[Link])
This work was supported by Zhejiang University of Science and Technology Teaching Research and Reform Key Projects
under Grant 2022-jg15.

ABSTRACT Predicting student performance is crucial for improving students’ future academic
achievements. Within student groups, common characteristics can reveal trends in overall student
learning. Most studies tend to focus on the common characteristics of students but ignore their
individual characteristics. However, individual characteristics are important in promoting student academic
performance because they allow us to understand the unique learning performances of each student.
To address this issue, this paper proposes a student performance prediction approach. First, addressing
the problem of difficulty in effectively dividing the student samples under multi-dimensional discrete data,
we propose a method that combines the relationship matrix-based bipartite network approach (RMBN) with
Louvain clustering. Second, the hybrid neural network model based on a relationship matrix (RMHNN) is
proposed to address the problem that discrete types of features are difficult to fit by algorithms. The results
show that the implementation of the model on real student data can effectively predict student performance
with an accuracy of 93.1% and an F1-score of 90.45%. With the model’s predicted student performance,
educators can provide individualized support and assistance to each student.

INDEX TERMS Performance prediction, data mining, common characteristics, individual characteristics,
relation network.

I. INTRODUCTION Numerous studies related to student development have

Academic performance is the most important measure of been published within the field of education. Numerous
the quality of college students, serving as the fundamental studies cover various aspects, including student dropout
criterion for universities to monitor teaching quality and rates [2], student profiling [3], [4], [5], monitoring student
evaluate students’ academic abilities [1]. In school period, academic performance [6], [7], [8], [9], and predicting at-risk
student’s performance is reflected in both academic and prac- students [10]. In research concerning student dropout rates,
tical performance. Academic performance refers to students’ students are typically classified as either ‘‘good’’ or ‘‘poor’’
performance in various courses, while practical performance [11], [12], [13], with targeted interventions aimed at high-risk
includes students’ performance in research competitions and student groups. Student profiling studies [14] involve the
on-campus scholarships, among other aspects. Students in the creation of student profiles by collecting behavioral and
same college or university have similar test scores at the time academic data to enable early detection of students exhibiting
of enrollment. However, a variety of factors can lead to a abnormal behavior. Additionally, there have been endeavors
gradual widening of performance between students during to predict groups of students facing learning difficulties based
their studies. Numerous researchers have conducted exten- on their assignment submissions [15], as well as efforts to
sive studies to address the issues of incomplete assessment identify students’ learning styles through the analysis of their
of student performance and the lack of tools for assessing academic data characteristics [16].
students’ performance in most universities. Most of the research on predicting student performance
focuses on extracting common characteristics among student
The associate editor coordinating the review of this manuscript and groups while neglecting the consideration of individual
approving it for publication was Sandra Baldassarri. characteristics of students. There are three problems with
2023 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
131260 For more information, see [Link] VOLUME 11, 2023
Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

these studies. Firstly, in terms of the choice of data objectives traditional machine learning algorithms and neural
[17], [18], most studies choose to use students’ overall networks. The experimental results show that our model
grades to assess students’ academic performance. However, is valid.
this approach ignores the performance of each student in The remaining sections of this paper are organized as
different subjects, making it difficult to extract individual follows. Section II presents the related work. Section III
characteristics. Secondly, in terms of the choice of research provides an overview of data collection, data integration, and
method, most of the studies employ a supervised learning data cleaning processes for the dataset. Section IV offers
approach [19]. Since assessing student learning is a complex a detailed explanation of the solution design for predicting
concept, it is difficult for us to accurately measure the students’ future academic performance. Section V evaluates
problem with a single target variable. Using an unsupervised the performance of the adopted methods and presents the
learning approach can more fully utilize all the data features results analysis. Finally, section VI concludes the paper.
to mine the individual differences between students. Thirdly,
in terms of student group division [11], [15], [20], most II. RELATED WORK
studies tend to categorize students as ‘‘good’’ or ‘‘poor’’. With the rapid growth of student data and the continuous
While this categorization helps provide early warnings for development of algorithm models, we are now able to
students classified as ‘‘poor’’, it is overly broad for students construct more effective student prediction models. One
classified as ‘‘good’’, potentially overlooking individual common approach is to use a grading mechanism for cate-
differences within this category. gorization. Nghe et al. [17] divided students into categories
This paper focuses on both individual and common based on grade point average (GPA) and used students’
characteristics of students, aiming to predict students’ multidimensional data features and yearly academic GPA
academic performance by exploring these characteristics. to predict their GPA category for the next academic year.
Firstly, to address the problem of k-means struggling to effec- Additionally, depending on the research needs and student
tively cluster student samples in multi-dimensional discrete characteristics, different methods can be used for student cat-
data, we propose a method that combines the relationship egorization and prediction. Akram et al. [15] aimed to predict
matrix-based bipartite network approach (RMBN) with Lou- students with learning difficulties based on their assignment
vain. Secondly, to tackle the issue that existing studies using submission behavior. They used the k-means clustering
a single objective variable may not accurately assess student algorithm to label students as procrastinators and non-
performance, we use a multi-objective assessment method. procrastinators, followed by applying different classification
We propose RMBN approach. Thirdly, to address the problem methods to categorize the students. By integrating behavioral
that discrete features are unsuitable for learning within a categorization models with behavioral data, they obtained
feature space based on Euclidean distance, we propose the category-specific feature values for each behavior, enabling
hybrid neural network model based on a relationship matrix the construction of machine learning-based performance
(RMHNN). Fourthly, to resolve the concern that existing prediction. Asif et al. [20] conducted research based on
research tends to focus on students’ common characteristics college students’ performance and divided the student group
but neglects their individual characteristics, we propose the into low-achieving and high-achieving for nurturing and
student performance prediction approach. intervention purposes.
The main contributions of the paper are summarized as In the application of student performance prediction, most
follows: studies have employed machine learning algorithms. Yağcı
• We address the difficulty of k-means in effectively [21] proposed a new model based on machine learning
classifying student samples under multi-dimensional algorithms to predict undergraduates’ final exam scores using
discrete data by combining the proposed RMBN midterm exam scores as the source data. The performance
approach with Louvain. of the proposed model was compared with other machine
• We use a multi-objective assessment method and learning algorithms, and the results showed a classification
propose RMBN. By using the RMBN data processing accuracy of 70%-75%. In terms of student dropout rates,
approach, we can comprehensively consider multi- Niyogisubizo et al. [22] used an ensemble model to predict
ple performance indicators, addressing the difficulty students’ dropout status in university courses. When it comes
of capturing associations between sample features to students’ choices of elective courses, Huang et al. [23]
when analyzing multidimensional data, which helps developed a score-based elective course recommendation
us achieve the goal of accurately assessing student system that accurately predicted the course scores for each
performance. student by leveraging the course grade distributions of similar
• We propose the RMHNN model to solve the problem senior students. Tomasevic et al. [11], through increasing
of discrete types of features that are difficult to fit. the available data sources gradually and employing state-of-
Experimental results show that the model can accurately the-art supervised machine learning techniques to perform
predict student performance. a comprehensive analysis and comparison. They discovered
• We apply the proposed a student performance prediction that increasing the available data improves the detection
approach to real student data and compare it with performance of educational data mining. To identify the

VOLUME 11, 2023 131261

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

importance of factors influencing student performance,

Alshanqiti and Namoun [24] proposed a hybrid regression
model and multi-label classifier to explore the various factors
influencing student grades. In addition to machine learning,
some studies have also utilized deep learning methods.
In contrast to the majority of existing research that utilized
machine learning to predict student performance, Akour et al. FIGURE 1. Structure of the analyzed data (student-related features).
[25] addressed and validated the prediction of student grades
using a convolutional neural network-based deep learning
algorithm. research awards (No research competition, With research
Although there have been many studies using data mining competition).
techniques to predict student performance, most of them
have focused on mining the common characteristics of the IV. METHODOLOGY
student groups and lacked a deeper dive into the individual A. APPROACHES TO PREDICTING STUDENT
characteristics of each student. In the student performance PERFORMANCE
prediction task, this paper considers both the common We propose the student performance prediction approach.
characteristics of the student body as a whole as well as the As shown in Figure 2. To comprehensively assess student
individual characteristics of students in different subgroups. performance, we collect student performance characteristics
at both the academic and practical performance levels to
enable the construction of a dataset that can accurately
III. DATA reflect student performance. The integrated and cleaned
A. STUDENT DATASET data includes both discrete-type features and continuous-type
The availability of data and the feasibility of the research features. In the case of discrete-type features, the distribution
are fundamental aspects of conducting scientific studies. of eigenvalues is too sparse. During data preprocessing,
As shown in Table 1, we utilizes student data from all we transform continuous features into discrete ones using
majors within the information school at a university for data binning and convert character-type features into discrete
2017 and 2018. features through one-hot coding. After the data preprocessing
is complete, we obtain a grade matrix S with too large
a feature space and too sparse a feature representation
B. DATA PREPROCESSING vector. Because the eigenvalues of the distribution in the
We extract the students’ names and course GPAs from five grade matrix S exhibit high dimensionality and discrete
data tables, integrated them using students’ names as the properties, we have difficulty capturing the associations
key, and remove duplicate data. Additionally, we standardize between students from this data. First, we propose the
synonymous attribute fields in the dataset to a uniform format RMBN data processing approach to address the problem of
and retain only one attribute feature. Sensitive information in weak correlation between students in a high-dimensional and
the data tables is anonymized. discrete grade matrix S. The RMBN data processing approach
Since the cleaned dataset contains both discrete and projects the data into a bipartite network G after encoding
continuous features, we apply a binning strategy to the the entire multidimensional data in one-hot. With the data
continuous course grade data [26], [27]. This process structure of this graph network, the associations between
transforms continuous data into discrete categories. Figure 1 students and variables can be made explicit, which helps
illustrates the mapping relationship for discretizing course subsequent algorithms capture similarities and differences
grades, where continuous GPAs are segmented into five between features. Second, the Louvain clustering algorithm
distinct categories. Moreover, we use one-hot encoding for can fully take into account the associations between students
categorical data types, representing each category with a in complex graph networks. We use Louvain clustering
binary vector. Additionally, character data is converted into as a graph-based feature extraction method for bipartite
integer data types. network G to form grouping results with strong similarities
Finally, the grade matrix S is generated, which com- between students within groups and strong differences
prises 20 feature attributes, including grade level (Grade), between students outside groups. Third, after obtaining
major (Major), ideological and political education (Pol.), metrics to accurately assess student performance, we research
foreign language (Fore.), physical education (Phys.), basic the prediction of student performance. In response to the
sciences (Math.), KAB entrepreneurship foundation (KAB), fact that the feature values in the grade matrix S are
college chinese (Chi.), health guidance (Health), disciplinary in the form of high-dimensional and discrete data, and
majors (Core), core majors (Prof.), extended majors (Exp.), the discrete features are not suitable for learning within
composite majors (Com.), special design (Spec.), gradu- a feature space based on Euclidean distance, we pro-
ation design (Grad. Proj), gender, post-graduation plans, pose the RMHNN model to predict students’ academic
political affiliation, school awards (No-JSA, With-JSA), and performance.

131262 VOLUME 11, 2023

Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

TABLE 1. Data sources and dataset variables.

FIGURE 2. Student performance prediction approach based on educational data mining.

B. THE RMBN DATA PROCESSING APPROACH relation matrix as a bipartite network [28], [29], [30], [31],
The feature space of the grade matrix S obtained after data and this treatment helps to clarify the structure of relations in
preprocessing is too large and the feature representation the data.
vector is too sparse. When processing this data using
traditional machine learning algorithms and neural networks, C. THE LOUVAIN CLUSTERING ALGORITHM
it is difficult to capture the association of eigenvalues In complex graph networks, the Louvain clustering algorithm
between samples. We propose the RMBN data processing can comprehensively take into account the associations
approach. The RMBN data processing approach undergoes between students [32], [33], [34], which in turn results
two processing steps. First, we encode the entire preprocessed in groupings of high similarity between students within
dataset with one-hot coding to generate a relation matrix with groups and significant differences between students between
only 0 and 1 eigenvalues. Doing so helps to highlight patterns groups. Therefore, this paper adopts Louvain clustering as
and structures in the data and clarify the links between a graph-based feature extraction method to capture both
samples and variables in the grade matrix S. In addition, common characteristics of student groups and individual
numerical differences between categories in the grade matrix characteristics of each student within high-dimensional and
S can be eliminated by this treatment, so that subsequent discrete data.
algorithms do not incorrectly assign greater weights to certain
categories because of the magnitude of the values, realizing D. THE RMHNN MODEL
the goal of coding each category with the same weight. As shown in Figure 3. The RMHNN model includes an input
Second, to solve the problem that the data feature space layer, a graph convolution layer, and a cross-attention layer.
becomes too large and the feature representation vector First, the grade matrix S and the student relation network A
becomes too sparse after one-hot coding, we project the are input to the input layer. Second, the graph convolutional

VOLUME 11, 2023 131263

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

network layer embeds the relation network A into the grade student’s academic performance into n categories, student
matrix S. By setting the threshold range of relation edges group 1, student group 2, etc., which are denoted as target =
and other methods, it helps the model maximize the learning {1, . . . , n}.
of the hidden feature association information in the graph In the RMHNN model, the graph convolution layer
network, thus forming a new matrix representation. Third, utilizes graph convolutional network (GCN) [35], [36] to
the cross-attention layer helps to assign weights to factors embed the student relation network into the student data,
that affect student performance. Finally, the linear layer is enriching the representation information of the student data
utilized to fuse the newly formed data features to complete and better capturing the mutual influence and network social
the prediction of student performance. structure among students. In the graph convolutional layer,
To capture correlations in student academic performance, we optimize the adjacency matrix in the GCN by changing
we design a method based on the concept of social network the threshold of relationship edges, using a linear scaling
and according to our defined rule that ‘‘if students have function to resize the adjacency matrix, and other methods to
similar course grades under the same course, there is a improve model performance. Furthermore, we also conduct
dependency relationship between these two students, and the normalization and adjust the feature dimensions.
weight of the undirected edges connecting these two students’ In the RMHNN model, the cross-attention layer utilizes
nodes is 1’’. Under the rule, we design the single graph a cross-attention network [37], [38] to handle the corre-
extraction algorithm based on node connectivity relationships lation between different features, better understanding the
(BNSN) that can extract the student nodes and the undirected importance and influence among different features. In the
edges between them from the bipartite network G generated cross-attention layer, we enhance the model’s predictive
based on the RMBN data processing approach, and calculate performance by integrating various data sources to explore
the weights of these undirected edges. The pseudocode for the best dataset combinations.
the BNSN(G) algorithm1 is as follows:
E. EVALUATION METRICS
Algorithm 1 BNSN(G) Algorithm Modularity(Q) is an important metric used to measure the
input : Bipartite network G density of community structure distribution [39]. The formula
output: Relation network A for modularity(Q) [40] is as follows:
Convert the bipartite network G into a triplet 1 XX
N N
di dj
T = (U , V , E); Q= (Ai,j − )δ(ci , cj ) (1)
2m 2m
for i = 0 to T .size[0] do i=1 j=1
Grouping of students according to their node
number; In equation (1), the undirected graph formula, N is the
end network size, Ai,j represents the adjacency matrix between
C=[]; node i and node j. The variable 2m represents the sum of all
for i = 0 to T .size[0]-1 do edges in the adjacency matrix A. di represents the degree of
Q=[]; node i, dj represents the degree of node j. ci represents the
initialization count; community cluster to which node i belongs, and cj represents
Extract all variables v owned by the student with the community cluster to which node j belongs, and the δ
node i ; function takes if node i and node j are in the same community,
Vi = { vi1 , vi2 , vi3 ,. . . } ; δ(ci , cj ) is equal to 1; otherwise, it is equal to 0.
for j = i+1 to T .size[0] do Accuracy (yaccuracy ) is one of the commonly used evalua-
count=0; tion metrics in classification models. Precision (yprecision ) is
Extract all variables v owned by the student one of the evaluation metrics for classification models. Recall
with node j ; (yrecall ) is one of the evaluation metrics for classification
Vj = { vj1 , vj2 , vj3 ,. . . } ; models. F1-score (yF1 ) is a comprehensive evaluation metric.
if X ∈ Vi ∩ Vj then count++; Their formulas are as follows:
else count=0; yTP + yTN
yaccuracy = (2)
[Link](i,j,count); (yTP + yFP + yTN + yFN )
end yTP
yprecision = (3)
[Link](Q); (yTP + yFP )
end yTP
yrecall = (4)
(yTP + yFN )
yprecision × yrecall
In the input layer of the RMHNN model, we use the student yF1 = 2 × ( ) (5)
(yprecision + yrecall )
relation network A with the grade matrix S as the student
characteristics for the student performance prediction task. In equation (2-5), yTP represents the number of samples
Based on the Louvain grouping scheme, we categorize each where both the actual and predicted labels are positive,

131264 VOLUME 11, 2023

Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

FIGURE 3. The hybrid neural network model based on a relationship matrix (RMHNN).

indicating consistent predictions with the actual labels. yTN logistic regression, plain bayes, and k-nearest neighbor
represents the number of samples where both the actual (KNN) from machine learning algorithms as classifiers.
and predicted labels are negative, indicating consistent Literature [21] uses a convolutional neural network (CNN)
predictions with the actual labels. yFP represents the number as a classifier. Literature [43] uses an attention-based graph
of samples where the actual label is negative and the predicted convolutional network as a classifier. To achieve accurate
label is positive, indicating inconsistent predictions. yFN student performance prediction, we introduce the RMHNN
represents the number of samples where the actual label is model. To validate the effectiveness of our proposed RMHNN
positive but the classifier model predicts it as negative. model, we apply the classification algorithms from previous
studies to our data with substitutions. The experimental
V. RESULTS AND DISCUSSION results demonstrate that our proposed model surpasses other
A. ANALYSIS OF THE STUDENT PERFORMANCE classification algorithms in accuracy and can effectively
PREDICTION MODEL predict student performance.
First, in the task of evaluating student performance, to assess Finally, for student performance assessment and student
student performance, literature [5], [8], [16], [20] uses performance prediction, we propose a combination of the
clustering algorithms to group similar or related objects RMBN data processing approach and Louvain clustering for
together. The k-means clustering algorithm is often used to assessed student performance, and the RMHNN model for
deal with continuous data The k-means clustering algorithm predicted student performance, respectively. The experimen-
is often used for continuous data because it clusters based tal results show that our proposed a student performance
on the distances between samples, and distance measures prediction approach is effective.
are typically meaningful for continuous data [41]. The
dataset used in this paper contains both discrete-type features B. LOUVAIN ALGORITHM CLUSTERING RESULTS
and continuous-type features, and after data preprocessing, In the Louvain clustering algorithm, resolution is a parameter
the resulting grade matrix S is all discrete-type features. used to adjust the quantity of community nodes. In Table 2,
Therefore, the k-means algorithm may be less suitable for when the resolution is set to 1, Newman*, Dugue*, and Potts*
the dataset used in this paper. To address the challenge of use data processed based on the RMBN data processing
clustering multiple discrete features accurately, we combine approach. The purpose of this approach is to convert the
the proposed RMBN data processing approach with Louvain preprocessed student dataset into a bipartite network, where
clustering. The experimental results demonstrate that this the values of the relationship edge connecting the student
combination reduces the dimensionality of the encoded data, nodes to the variable nodes in the formed graph network
clarifies the connections between students and variables, and differ due to the inconsistency of the eigenvalues between the
achieves highly modular grouping results. different features. This may lead to subsequent algorithms
Second, in the task of predicting student performance, lit- confusing the correlations between students. Therefore,
erature [16] uses decision trees (DT) as classifiers. Literature to balance the weights of the relationship edges in the graph
[42] uses random forest, support vector machines (SVM), network, this paper encodes the dataset once again with

VOLUME 11, 2023 131265

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

one-hot coding to make the connection between the student

nodes and the variable nodes in the graph network clearer.
Newman, Dugue, and Potts used standardized processed data.
Based on the modularity values (Q), it can be observed
that the clustering results in the left three columns have
higher Q, namely 73.37%, 71.37%, and 19.30%, respectively.
On the other hand, the clustering results in the right three
columns have lower Q, namely 4.39%, 48.99%, and 6.28%,
respectively. The clustering results in the left three columns
yield 6, 7, and 72 communities, while the results in the
right three columns yield 10, 23, and 50 communities,
respectively. According to the modularity evaluation metric,
the clustering results in the left three columns exhibit higher
Q, indicating a higher density of connections within com-
FIGURE 4. Visualization of student academic performance in different
munities compared to connections between communities, groups (student academic performance and student practical
and thus a tighter community structure. The experimental performance).
results indicate that after the data processed by the RMBN
data processing approach, the Louvain clustering algorithm
can effectively identify the community structure within
the network and assign student nodes to their respective
communities. Based on the experimental results, it can be
observed that after the data is processed by the RMBN
data processing approach, and then applying the Louvain
clustering method for grouping yields better experimental
results compared to directly using Louvain clustering for
grouping the preprocessed dataset after standardization. This
is due to the fact that one-hot coding is taken to discretize
the features throughout the RMBN data processing approach, FIGURE 5. Heatmaps of different student groups’ performance in
where each feature is transformed into a binary feature. extracurricular practices.

Because each feature is turned into an explicit node, this

explicit feature representation helps Louvain clustering to is set, the GCN achieves remarkably low accuracy and F1-
capture similarities and differences between features more score, with values of only 7.07% and 2.64%, respectively,
easily. while the graph attention network (GAT) performes slightly
In this paper, we use the Louvain (Newman*) clustering better with an accuracy of 34.48% and an F1-score of 17.68%.
algorithm to group the data processed through the RMBN This indicates that both models struggle to classify the student
data processing approach and then conduct a visual analysis data without a threshold. As the threshold for the connection
of the generated six student groups. In the visual results, these strength increases, the accuracy and F1-score of both the
six student groups exhibit distinct trends in their academic GCN and GAT gradually improve. When the threshold value
performance. As shown in Figure 4 and Figure 5. Students of the relationship edge is set to be greater than 8, the
in student group 2 excel academically and in practice. The accuracy and F1-score of the GCN significantly improve to
student in student group 5 needs to be strengthened in terms of 63.64% and 48.65%, respectively. Similarly, the accuracy
academics and in terms of practice. Students in student group and F1-score of the GAT also see significant improvements,
5 need to enhance their theoretical knowledge in academics reaching 77.01% and 77.51%, respectively. As the relational
and participate in research competitions to improve their edge thresholds are adjusted, the prediction accuracies and
practical skills. Students in student group 1, student group 3, F1-score of both the GCN and GAT stabilize. Specifically,
student group 4, and student group 5 need to strengthen their when the threshold value of the relationship edge is set to
academic performance in courses related to mathematical be greater than 14, the GCN achieves its highest prediction
fundamentals and subject-specific foundational courses. accuracy and F1-score at 84.85% and 63.64%, respectively.
By examining the academic and practical performance of For the GAT, the peak prediction accuracy and F1-score are
students in different programs, we can see that there is room reached when the threshold value of the relationship edge is
for improvement for all student groups. set to be greater than 15, at 81.61% and 81.02%, respectively.
These experiments demonstrate that adjusting the adjacency
C. RESULTS OF THE RMHNN MODEL matrix by setting a threshold range for relationship edges
In the experiment, we adjust the relationship edge threshold effectively enhances the model’s predictive performance.
range of the adjacency matrix to observe its impact on the When the threshold value of the relationship edge is set low,
model’s predictions. As shown in Figure 6, when no threshold there is difficulty in filtering out weaker connections between

131266 VOLUME 11, 2023

Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

TABLE 2. Comparative analysis of experimental results on module degree value Q and the number of community partition count.

FIGURE 6. The predictive effectiveness of the model is tested by setting a FIGURE 7. Adjusting the adjacency matrix by setting a scaling factor using
threshold range to adjust the adjacency matrix(%). a linear scaling function to evaluate the model’s predictions(%).

nodes, leading to the misinterpretation of unrelated nodes as

being connected. Conversely, when the threshold value is set thresholds to observe the model performance. Figure 8
high, it’s easy to filter out stronger connections, resulting in illustrates the effect of normalized and un-normalized on
the misinterpretation of related nodes as not being connected. the model performance with the relational edge threshold
The method of adjusting the adjacency matrix by setting a set. Under the premise of using the method of adjusting the
threshold for the relationship edges suffers from the problem threshold range of relationship edges, comparing the experi-
of not being able to capture subtle changes between the mental results in Figure 6 with the experiments in Figure 8
weights, which may lead to loss of information and affect where normalization is not applied, we can observe the
the performance of the model. The use of a linear scaling impact of introducing the cross-attention layer on the model’s
function as the weight adjustment function can better preserve prediction accuracy and F1-score. In Figure 6, the F1-score of
the information between the weights [44]. As shown in the model reaches its highest value when the threshold of the
Figure 7, in the GCN, the accuracy and F1-score of the model relationship edge is set to be greater than 12, with a value
peak at scale=0.0695 with values of 85.86% and 64.72%, of 71.62%. The highest accuracy of the model is achieved
respectively. In the GAT, the accuracy and F1-score of the when the threshold value of the relationship edge is set to
model peaked at scale=0.075 with values of 83.991% and be greater than 14, with a value of 84.85%. In Figure 8,
83.79%, respectively. The experimental results show that the F1-score of the model reaches its highest value with
using the linear scaling function as a weight function to a value of 81.06% when the threshold of the relationship
scale the data matrix helps improve the performance and edge is set to be greater than 11 without normalization.
generalization of the model, enhancing its ability to express The highest accuracy of the model, 80.98%, is achieved
complex relationships. when the threshold of the relationship edge is set to greater
In the GCN, the student relation network presents a certain than 12.
focus, while the attention mechanism in cross-attention Figure 7 shows how the data matrix is scaled in the GCN
networks can assign weights according to the degree of by using a linear scaling function and observing how it affects
importance of the features, calculate the weights of each the model performance. Figure 9 illustrates the effect of
student’s performance features, and highlight the key factors normalized and un-normalized on the model performance
affecting the student’s performance, which can solve the with the application of a linear scaling function. Comparing
specific weighting problem. Therefore, a cross-attention the experimental results in Figure 7 with those without
layer is introduced to balance the weights of the features normalization in Figure 9 under the premise of using a
affecting student performance. Figure 6 illustrates the model linear scaling function to scale the data matrix approach,
performance in the GCN by adjusting the relational edge we can observe the effect of introducing the cross-attention

VOLUME 11, 2023 131267

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

FIGURE 8. Validation of the effect of normalization on model

performance in the context of the method of setting a threshold range to FIGURE 9. Validating the effect of normalization on model performance
adjust the adjacency matrix(%). in a method that uses a linear scaling function to set the scaling factor to
adjust the adjacency matrix(%).

layer on the model prediction accuracy and F1-score.

In Figure 7, the highest accuracy of the model is achieved student performance is evaluated by selecting different data
when scale=0.0695 and its value is 85.85%. The F1-score dimension outputs, constructing student and course relation
of the model is highest when scale=0.08 and its value is networks using the grade matrix S, and employing methods
68.84%. In Figure 9, the accuracy and F1-score of the model such as integrating the grade matrix, student adjacency
peaks at scale=0.075 with values of 86.20% and 87.94%, matrix, and course adjacency matrix in the cross-attention
respectively, under the normalization condition. These results layer. Under the condition ‘‘RMHNN (Edge > 14) (900)’’, the
indicate that the introduction of the cross-attention layer can accuracy of the model is 90.80 and the F1-score is 91.71%.
indeed improve the F1-score of the model by considering Under the condition ‘‘RMHNN (scale=0.069) (900)’’, the
differences in how various characteristics affect different accuracy of the model is 91.95% and the F1-score is 89.67%.
students. The cross-attention network helps address these Under the condition ‘‘RMHNN (Edge > 14) (402)’’, the
variations in feature importance. accuracy of the model is 93.10% and the F1-score is 90.46%.
Normalization is a data preprocessing technique that aims Under the condition ‘‘RMHNN (scale=0.069) (402)’’, the
to standardize the scale of various features by mapping accuracy of the model is 93.10% and the F1-score is 90.45%.
the data to a specific range [45]. As shown in Figure 8, This shows that the choice of using different data dimensions
the impact of normalization and non-normalization on output in the graph convolutional layer can improve the
model performance is demonstrated under the condition F1-score of the model. ‘‘Using the student matrix alone
of setting the relationship edge threshold. The accuracy (Edge > 14) (900)’’ represents the condition that the threshold
and F1-score of the RMHNN model are highest when the of the relation edge is set to greater than 14 in the graph
relationship edge threshold is set to greater than 14, with convolution layer and the output dimension of the data is
values of 90.80% and 91.71%, respectively. The experimental increased to 900, and we use the student performance matrix
results show that when optimizing the model by setting S alone in the cross-attention layer for our experiments. ‘‘Add
the relationship edge threshold range, performing data Course Relationship Matrix (Edge > 14) (900)’’ means that
normalization can significantly improve the prediction of in the graph convolutional layer, with the threshold value
the model. In the student performance prediction task in for the relationship edges set to greater than 14 and the
this paper, performing data normalization can better improve data output dimension increased to 900, we conduct data
model performance and help identify correlations between fusion and experimental verification in the cross-attention
students. layer using three tables: the student grade matrix S, the
As depicted in Figure 9, the influence of normalization student adjacency matrix obtained from the student relation
and non-normalization on model performance is illustrated network, and the course adjacency matrix obtained from the
under the premise of applying a linear scaling function. Under course relation network. ‘‘Remove Student Relation Network
the condition of normalization, when scale=0.069, the model (Edge > 14) (900)’’ means that in the graph convolutional
achieves its highest accuracy and F1-score, which are 91.95% layer, with the threshold value for the relationship edges set
and 89.67%, respectively. In the absence of normalization, to greater than 14 and the data output dimension increases
when scale=0.075, the model achieves its highest accuracy to 900, we conduct data fusion and experimental verification
and F1-score, which are 86.20% and 87.94%, respectively. in the cross-attention layer using two tables: the student
The experimental results show that the effectiveness of grade matrix S and the course adjacency matrix obtained
model prediction can be improved by using the normalization from the course relation network. The experimental results
method. show that using the student adjacency matrix obtained based
As shown in Figure 10. In the graph convolutional on the student relation network in the cross-attention layer
layer, the performance of these methods in predicting can improve the model prediction. In the cross-attention

131268 VOLUME 11, 2023

Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

TABLE 3. Comparison of experimental results(%).

FIGURE 10. By using different data dimensions and different relational

networks to assess their impact on model performance(%).

layer, fusing the new graph representation results obtained

in the graph convolution layer with the student adjacency
matrix, helps the model to take into account the performance
characteristics of each student more accurately and improves
the accuracy and F1-score of the model. Theoretically, the
course relation network can improve the accuracy of the
model prediction, but since the course relation network in
this model is constructed based on the student performance
matrix S, its use does not improve the model’s accuracy and FIGURE 11. Embedding the student relation network in the graph
F1-score. convolutional layer and combining the grade matrix S with the student
adjacency matrix in the cross-attention layer to visualize the model’s
The results of the model comparison analysis experiments prediction performance through heatmaps.
are shown in Table 3 where the RMHNN model outperforms
the other models overall. The prediction accuracies of KNN,
SVM, and DT were low, only 61.68%, 72%, and 66.6%. This results show that the RMHNN model can effectively deal
is because traditional machine learning algorithms and neural with the problem that the feature space is too large and the
network models are not suitable for fitting data tables where feature representation vectors are too sparse, which makes
the feature space is too large and the feature representation it difficult to capture the correlation of eigenvalues among
vectors are too sparse. The SVM although predicts well in the students.
high dimensional data space, requires more support vectors As shown in Figure 11 and Figure 12. Figure 11
to maintain the boundaries leading to a prediction accuracy illustrates the embedding of the student relation network
of 72% and an F1-score of only 57%. DT are able to handle into the graph convolutional layer to obtain a new graph
mixed data types, but in high dimensional spaces, DT need representation. Subsequently, in the cross-attention layer, this
to increase their depth to capture all features, but doing so novel graph representation is fused with the student adjacency
raises the issue of overfitting. The prediction accuracy of the matrix, followed by experimental predictions. This process
decision tree was 66.6% with an F1-score of 51%. Multilayer is presented through visualization to showcase its impact on
perceptron (MLP), backpropagation network (BPNN), and model predictions. Figure 12 demonstrates the embedding of
CNN are deep learning algorithms, but their prediction the course relation network into the graph convolutional layer
accuracies are only 68.97%, 67.2%, and 68.58%, due to the to obtain a new graph representation. Subsequently, in the
problem that deep neural networks are prone to overfitting in cross-attention layer, this novel graph representation is fused
high-dimensional sparse data. The prediction results of GCN with the course adjacency matrix, followed by experimental
and GAT are significantly better than other models, which is predictions. This process is presented through visualization
because they solve the problem of weak correlation of feature to showcase its impact on model predictions. A comparison
values between samples under high-dimensional sparse data of the visualization effects reveals that the student relation
by means of graph network representation. They adjust network constructed based on the grade matrix S helps
the association between the samples through the embedded the model to capture the association of eigenvalues among
adjacency matrix, which allows the model to learn the depen- student samples. By comparing the visualization results,
dency of the eigenvalues between the samples to improve it can be found that constructing the student relation
the performance of the model prediction. The experimental network based on the grade matrix S helps the model to

VOLUME 11, 2023 131269

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

the model is best when scale=0.075. Third, a potential issue

with GCN is that they may suffer from feature-specificity
problems. This paper conducts experiments to compare
whether the introduction of the cross-attention layer has
an impact on model performance. The experimental results
show that the introduction of the cross-attention layer can
effectively recognize the complex connection of eigenvalues
between students and significantly improve the F1-score of
the model. Fourth, normalization increases the stability of the
model. The RMHNN model is used to evaluate the effect of
normalized and unnormalized on the model performance with
FIGURE 12. Embedding the course relation network in the graph a relational edge threshold set. The experimental results indi-
convolution layer and combining the grade matrix S with the course
adjacency matrix in the cross-attention layer, the model’s predictive cate that, under the premise of setting the relationship edge
performance is visualized through heatmaps. threshold, data normalization is conducive to the model’s
learning and generalization. Fifth, under the premise of
applying a linear scaling function, the RMHNN model is used
capture the correlation of eigenvalues among the student to assess the impact of normalization and non-normalization
samples. on model performance. The experimental results indicate that
applying normalization to the RMHNN model, which uses
D. ANALYSIS OF EXPERIMENTAL RESULTS a linear scaling function as a weight adjustment function,
In the task of assessing student performance, to address can improve the accuracy and F1-score of the model in pre-
the difficulty of k-means clustering in dealing with dicting student performance. Six, we conduct a comparative
high-dimensional and sparse discrete data matrices, we use analysis of the proposed RMHNN model with traditional
a combination of the RMBN data processing approach machine learning algorithms and neural network models,
and Louvain clustering to achieve an accurate assessment experimentally verifying the effectiveness of the RMHNN
of student performance. The results of the experiment model. Seventh, the RMHNN model is subjected to ablation
shows that the RMBN data processing approach enhances experiments. The results showed that changing the dimension
the connection between students and variables. Using a of the features in the graph convolution layer within the
combination of the RMBN data processing approach and RMHNN model has a moderate impact. Conversely, replac-
Louvain clustering, helps the algorithm to clarify the ing the student-relational network embedded in the GCN
correlation between the students in the data matrix and with the course-relational network reduces the experimental
to form groupings of students with a higher degree of effect.
modularity.
In the RMHNN model, we use GCN to embed the
relation network into the grade matrix S, addressing the VI. CONCLUSION
problem that using Euclidean distances to measure distances In the developmental process of student learning, predicting
between discrete features is not practically meaningful. The student performance can help educators gain a greater
experiments in this paper are carried out in the following understanding of students’ learning needs. To delve into
seven areas. First, we apply the datasets to the GCN and students’ academic achievements across various subjects and
GAT. By adjusting the adjacency matrices embedded into their practical performance in diverse scenarios, we conduct
the models, we observe that altering the thresholds of the a comprehensive analysis of student data with multiple
relationship edges could impact the accuracy of the models dimensions. This paper’s approach not only centers on
in predicting student performance. The experimental results understanding the overall learning patterns exhibited by the
also validate the effectiveness of the proposed rules for majority of students but also considers the performance
relation networks. In the GCN, the model achieves the best of each student across distinct subjects and their practical
accuracy when the threshold of the relationship edge is set to aptitude in varying situations. Therefore, we propose a
be greater than 14. In the GAT, the accuracy of the model is student performance prediction approach. We apply the
best when the threshold of the relationship edge is set to be proposed model of student performance to real-world
greater than 15. Second, to achieve finer weight adjustment student data. The results demonstrate that our model
and capture similar changes in eigenvalues between samples, is effectively predicting students’ performance. By iden-
this paper uses a linear function as the weight adjustment tifying students’ performance, teachers can adjust their
function to scale the data matrix. Experiments show that teaching strategies to cater to students’ learning needs.
the use of linear scaling functions captures the correlation By mastering the distribution of students’ performance and
between samples and improves the model’s ability to express future performance trends, educational administrators can
complex relationships. In the GCN, the model has the best deploy resources rationally to student groups with different
accuracy when scale=0.0695. In the GAT, the accuracy of performances.

131270 VOLUME 11, 2023

Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

REFERENCES [21] M. Yağcı, ‘‘Educational data mining: Prediction of students’ academic

performance using machine learning algorithms,’’ Smart Learn. Environ.,
[1] E. T. Lau, L. Sun, and Q. Yang, ‘‘Modelling, prediction and classification
vol. 9, no. 1, p. 11, 2022.
of student academic performance using artificial neural networks,’’ Social
Netw. Appl. Sci., vol. 1, no. 9, pp. 1–9, Sep. 2019. [22] J. Niyogisubizo, L. Liao, E. Nziyumva, E. Murwanashyaka, and
P. C. Nshimyumukiza, ‘‘Predicting student’s dropout in university
[2] W. Wang, Y. Zhao, Y. J. Wu, and M. Goh, ‘‘Factors of dropout from
classes using two-layer ensemble machine learning approach: A novel
MOOCs: A bibliometric review,’’ Library Hi Tech, vol. 41, no. 2,
stacked generalization,’’ Comput. Educ., Artif. Intell., vol. 3, Jan. 2022,
pp. 432–453, Jun. 2023.
Art. no. 100066.
[3] X. Li and S. He, ‘‘Research and analysis of student portrait based on [23] L. Huang, C.-D. Wang, H.-Y. Chao, J.-H. Lai, and P. S. Yu, ‘‘A score
campus big data,’’ in Proc. IEEE 6th Int. Conf. Big Data Analytics prediction approach for optional course recommendation via cross-user-
(ICBDA), Mar. 2021, pp. 23–27. domain collaborative filtering,’’ IEEE Access, vol. 7, pp. 19550–19563,
[4] F. Yang and F. W. B. Li, ‘‘Study on student performance 2019.
estimation, student progress analysis, and student potential prediction [24] A. Alshanqiti and A. Namoun, ‘‘Predicting student performance and its
based on data mining,’’ Comput. Educ., vol. 123, pp. 97–108, influential factors using hybrid regression and multi-label classification,’’
Aug. 2018. IEEE Access, vol. 8, pp. 203827–203844, 2020.
[5] Z. Mingyu, W. Sutong, W. Yanzhang, and W. Dujuan, ‘‘An [25] M. Akour, H. A. Sghaier, and O. Al Qasem, ‘‘The effectiveness of
interpretable prediction method for university student academic using deep learning algorithms in predicting students achievements,’’
crisis warning,’’ Complex Intell. Syst., vol. 8, no. 1, pp. 323–336, Indonesian J. Electr. Eng. Comput. Sci., vol. 19, no. 1, p. 388,
Feb. 2022. Jul. 2020.
[6] L. E. Adamova and O. O. Varlamov, ‘‘Logic artificial intelligence [26] V. L. Miguéis, A. Freitas, P. J. V. Garcia, and A. Silva, ‘‘Early
application for the students individual trajectories introduction,’’ in Proc. segmentation of students according to their academic performance: A
ITM Web Conf., vol. 35, 2020, p. 2001. predictive modelling approach,’’ Decis. Support Syst., vol. 115, pp. 36–51,
[7] A. Muhammad, Q. Zhou, G. Beydoun, D. Xu, and J. Shen, ‘‘Learning Nov. 2018.
path adaptation in online learning systems,’’ in Proc. IEEE 20th Int. [27] W. Chango, R. Cerezo, M. Sanchez-Santillan, R. Azevedo, and C. Romero,
Conf. Comput. Supported Cooperat. Work Design (CSCWD), May 2016, ‘‘Improving prediction of students’ performance in intelligent tutoring
pp. 421–426. systems using attribute selection and ensembles of different multimodal
[8] H. Lim, S. Kim, K.-M. Chung, K. Lee, T. Kim, and J. Heo, ‘‘Is college data sources,’’ J. Comput. Higher Educ., vol. 33, no. 3, pp. 614–634,
students’ trajectory associated with academic performance?’’ Comput. Dec. 2021.
Educ., vol. 178, Mar. 2022, Art. no. 104397. [28] N. Gitinabard, F. Khoshnevisan, C. F. Lynch, and E. Yuan Wang,
[9] L. Meng, W. Zhang, Y. Chu, and M. Zhang, ‘‘LD–LP generation ‘‘Your actions or your associates? Predicting certification and dropout
of personalized learning path based on learning diagnosis,’’ in MOOCs with behavioral and social features,’’ 2018, arXiv:1809.
IEEE Trans. Learn. Technol., vol. 14, no. 1, pp. 122–128, 00052.
Feb. 2021. [29] D. J. Watts and S. H. Strogatz, ‘‘Collective dynamics of ’small-
[10] A. A. Mubarak, H. Cao, and W. Zhang, ‘‘Prediction of students’ world’networks,’’ Nature, vol. 393, nos. 66–84, pp. 440–442, 1998.
early dropout based on their interaction logs in online learning envi- [30] R. Van Noorden, ‘‘Online collaboration: Scientists and the social network,’’
ronment,’’ Interact. Learn. Environ., vol. 30, no. 8, pp. 1414–1433, Nature, vol. 512, no. 7513, pp. 126–129, Aug. 2014.
Jul. 2022. [31] G. A. Pavlopoulos, P. I. Kontou, A. Pavlopoulou, C. Bouyioukos,
[11] N. Tomasevic, N. Gvozdenovic, and S. Vranes, ‘‘An overview E. Markou, and P. G. Bagos, ‘‘Bipartite graphs in systems biology and
and comparison of supervised data mining techniques for student medicine: A survey of methods and applications,’’ GigaScience, vol. 7,
exam performance prediction,’’ Comput. Educ., vol. 143, Jan. 2020, no. 4, Apr. 2018, Art. no. giy014.
Art. no. 103676. [32] P. De Meo, E. Ferrara, G. Fiumara, and A. Provetti, ‘‘General-
[12] B. K. Bhardwaj and S. Pal, ‘‘Data mining: A prediction for performance ized Louvain method for community detection in large networks,’’
improvement using classification,’’ 2012, arXiv:1201.3418. in Proc. 11th Int. Conf. Intell. Syst. Design Appl., Nov. 2011,
[13] Y. Zhao, Q. Xu, M. Chen, and G. M. Weiss, ‘‘Predicting student pp. 88–93.
performance in a master of data science program using admis- [33] M. J. Barber, ‘‘Modularity and community detection in bipartite net-
sions data,’’ in Proc. Int. Educ. Data Mining Soc., Jul. 2020, works,’’ Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat. Interdiscip. Top.,
pp. 1–9. vol. 76, no. 6, Dec. 2007, Art. no. 066102.
[14] G. Su-Hui, B. Cheng-Jie, and W. Quan, ‘‘Hadoop-based college student [34] M. E. J. Newman and M. Girvan, ‘‘Finding and evaluating community
behavior warning decision system,’’ in Proc. IEEE 3rd Int. Conf. Big Data structure in networks,’’ Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat.
Anal. (ICBDA), Mar. 2018, pp. 217–221. Interdiscip. Top., vol. 69, no. 2, Feb. 2004, Art. no. 026113.
[15] A. Akram, C. Fu, Y. Li, M. Y. Javed, R. Lin, Y. Jiang, and [35] J. Zhou, ‘‘Graph neural networks: A review of methods and applications,’’
Y. Tang, ‘‘Predicting students’ academic procrastination in blended AI Open, vol. 1, pp. 57–81, Jan. 2020.
learning course using homework submission data,’’ IEEE Access, vol. 7, [36] T. N. Kipf and M. Welling, ‘‘Semi-supervised classification with graph
pp. 102487–102498, 2019. convolutional networks,’’ 2016, arXiv:1609.02907.
[16] N. Hidayat, R. Wardoyo, A. Sn, and H. Dwi, ‘‘Enhanced performance [37] Z. Niu, G. Zhong, and H. Yu, ‘‘A review on the attention mechanism of
of the automatic learning style detection model using a combination of deep learning,’’ Neurocomputing, vol. 452, pp. 48–62, Sep. 2021.
modified K-means algorithm and naive Bayesian,’’ Int. J. Adv. Comput. [38] R. Hou, H. Chang, B. Ma, S. Shan, and X. Chen, ‘‘Cross attention network
Sci. Appl., vol. 11, no. 3, pp. 638–648, 2020. for few-shot classification,’’ in Proc. Adv. Neural Inf. Process. Syst.,
[17] N. T. Nghe, P. Janecek, and P. Haddawy, ‘‘A comparative analysis of vol. 32, 2019, pp. 1–12.
techniques for predicting academic performance,’’ in Proc. 37th Annu. [39] M. E. J. Newman, ‘‘Modularity and community structure in networks,’’
Frontiers Educ. Conf. Global Eng., Knowl. Without Borders, Opportunities Proc. Nat. Acad. Sci. USA, vol. 103, no. 23, pp. 8577–8582, Jun. 2006.
Without Passports, Oct. 2007, p. 7. [40] W. Zhan, Z. Zhang, J. Guan, and S. Zhou, ‘‘Evolutionary method for
[18] D. Kabakchieva, ‘‘Predicting student performance by using data mining finding communities in bipartite networks,’’ Phys. Rev. E, Stat. Phys.
methods for classification,’’ Cybern. Inf. Technol., vol. 13, no. 1, pp. 61–72, Plasmas Fluids Relat. Interdiscip. Top., vol. 83, no. 6, Jun. 2011,
Mar. 2013. Art. no. 066120.
[19] J. L. Rastrollo-Guerrero, J. A. Gómez-Pulido, and A. Durán-Domínguez, [41] A. M. Ikotun, A. E. Ezugwu, L. Abualigah, B. Abuhaija, and J.
‘‘Analyzing and predicting students’ performance by means of Heming, ‘‘K-means clustering algorithms: A comprehensive review,
machine learning: A review,’’ Appl. Sci., vol. 10, no. 3, p. 1042, variants analysis, and advances in the era of big data,’’ Inf. Sci., vol. 622,
Feb. 2020. pp. 178–210, Apr. 2023.
[20] R. Asif, A. Merceron, S. A. Ali, and N. G. Haider, ‘‘Analyzing [42] Z. Li, R.-S. Wang, S. Zhang, and X.-S. Zhang, ‘‘Quantitative function
undergraduate students’ performance using educational data mining,’’ and algorithm for community detection in bipartite networks,’’ Inf. Sci.,
Comput. Educ., vol. 113, pp. 177–194, Oct. 2017. vols. 367–368, pp. 874–889, Nov. 2016.

VOLUME 11, 2023 131271

 Z. Chen et al.: Student Performance Prediction Approach Based on Educational Data Mining

[43] Q. Hu and H. Rangwala, ‘‘Academic performance estimation with YING WEI received the M.E. degree in computer
attention-based graph convolutional networks,’’ 2019, arXiv:2001.00632. application from Zhejiang University, China.
[44] Y.-T. Xu, Y. Zhang, and X. Huang, ‘‘Single-machine ready He was the Head of the Information Center,
times scheduling with group technology and proportional linear where he was responsible for the top-level design
deterioration,’’ Appl. Math. Model., vol. 38, no. 1, pp. 384–391, of the informatization construction. He is cur-
Jan. 2014. rently an Associate Professor with the Zhejiang
[45] D. Singh and B. Singh, ‘‘Investigating the impact of data normalization University of Science and Technology, China.
on classification performance,’’ Appl. Soft Comput., vol. 97, Dec. 2020,
He has led the completion of projects, such as the
Art. no. 105524.
school’s digital campus and data center virtual-
ization. His research interests include educational
informatization and education data mining.

ZILING CHEN received the B.E. degree in

computer science and technology from Huzhou
University, China, in 2021. She is currently pursu-
ing the M.E. degree in applied statistics with the
Zhejiang University of Science and Technology,
China.
Her main research interest includes educational
data mining.

GANG CEN received the B.E. degree from the

Department of Physics, Hangzhou University,
China, in 1982. ZIFEI LI is currently pursuing the B.E. degree
He is currently a Professor with the Zhejiang in computer science and technology with the
University of Science and Technology, China. His Zhejiang University of Science and Technology,
research interests include educational information China.
science and technology and education data mining. His research interests include system design and
He is the Director of the National Computer data analysis.
Education Research Association for Higher Edu-
cation Institutions and the Director of the National
Association for Computer Fundamental Education Research in Higher
Education Institutions, among others.

131272 VOLUME 11, 2023