Volume 20, 2021

USING EDUCATIONAL DATA MINING TO PREDICT

STUDENTS’ ACADEMIC PERFORMANCE FOR APPLYING
EARLY INTERVENTIONS
Sarah Alturki* Data and Web Science Group, Faculty of alturki@[Link]-
Business Informatics and Mathematics, [Link]
University of Mannheim, Mannheim,
Germany
Nazik Alturki Information Systems Department, Col- namalturki@[Link]
lege of Computer and Information Sci-
ences, Princess Nourah bint Abdulrahman
University, Riyadh, Saudi Arabia
Heiner Stuckenschmidt Data and Web Science Group, Faculty of heiner@[Link]-
Business Informatics and Mathematics, [Link]
University of Mannheim, Mannheim,
Germany
* Corresponding author

ABSTRACT
Aim/Purpose One of the main objectives of higher education institutions is to provide a
high-quality education to their students and reduce dropout rates. This can be
achieved by predicting students’ academic achievement early using Educational
Data Mining (EDM). This study aims to predict students’ final grades and iden-
tify honorary students at an early stage.
Background EDM research has emerged as an exciting research area, which can unfold valu-
able knowledge from educational databases for many purposes, such as identi-
fying the dropouts and students who need special attention and discovering
honorary students for allocating scholarships.
Methodology In this work, we have collected 300 undergraduate students’ records from three
departments of a Computer and Information Science College at a university lo-
cated in Saudi Arabia. We compared the performance of six data mining meth-
ods in predicting academic achievement. Those methods are C4.5, Simple
CART, LADTree, Naïve Bayes, Bayes Net with ADTree, and Random Forest.

Accepting Editor Dennis Kira│ Received: May 9, 2021│ Revised: June 24, July 11, July 13, July 16, 2021 │
Accepted: July 17, 2021.
Cite as: Alturki, S., Alturki, N., & Stuckenschmidt, H. (2021). Using educational data mining to predict students’
academic performance for applying early interventions. Journal of Information Technology Education: Innovations in
Practice, 20, 121-137. [Link]
(CC BY-NC 4.0) This article is licensed to you under a Creative Commons Attribution-NonCommercial 4.0 International
License. When you copy and redistribute this paper in full or in part, you need to provide proper attribution to it to ensure
that others can later locate this work (and to ensure that others do not accuse you of plagiarism). You may (and we encour-
age you to) adapt, remix, transform, and build upon the material for any non-commercial purposes. This license does not
permit you to use this material for commercial purposes.
Predicting Students’ Academic Performance

Contribution We tested the significance of correlation attribute predictors using four differ-
ent methods. We found 9 out of 18 proposed features with a significant corre-
lation for predicting students’ academic achievement after their 4th semester.
Those features are student GPA during the first four semesters, the number of
failed courses during the first four semesters, and the grades of three core
courses, i.e., database fundamentals, programming language (1), and computer
network fundamentals.
Findings The empirical results show the following: (i) the main features that can predict
students’ academic achievement are the student GPA during the first four se-
mesters, the number of failed courses during the first four semesters, and the
grades of three core courses; (ii) Naïve Bayes classifier performed better than
Tree-based Models in predicting students’ academic achievement in general,
however, Random Forest outperformed Naïve Bayes in predicting honorary
students; (iii) English language skills do not play an essential role in students’
success at the college of Computer and Information Sciences; and (iv) studying
an orientation year does not contribute to students’ success.
Recommendations We would recommend instructors to consider using EDM in predicting stu-
for Practitioners dents’ academic achievement and benefit from that in customizing students’
learning experience based on their different needs.
Recommendations We would highly endorse that researchers apply more EDM studies across vari-
for Researchers ous universities and compare between them. For example, future research
could investigate the effects of offering tutoring sessions for students who fail
core courses in their first semesters, examine the role of language skills in so-
cial science programs, and examine the role of the orientation year in other
programs.
Impact on Society The prediction of academic performance can help both teachers and students
in many ways. It also enables the early discovery of honorary students. Thus,
well-deserved opportunities can be offered; for example, scholarships, intern-
ships, and workshops. It can also help identify students who require special at-
tention to take an appropriate intervention at the earliest stage possible. Moreo-
ver, instructors can be aware of each student’s capability and customize the
teaching tasks based on students’ needs.
Future Research For future work, the experiment can be repeated with a larger dataset. It could
also be extended with more distinctive attributes to reach more accurate results
that are useful for improving the students’ learning outcomes. Moreover, exper-
iments could be done using other data mining algorithms to get a broader ap-
proach and more valuable and accurate outputs.
Keywords Educational Data Mining (EDM), prediction of academic achievement, higher
education

INTRODUCTION
Since one of the benchmarks of a high-quality university is based on the students’ excellent record
of academic performance, students’ failure or low achievement is a big concern for higher education
institutions. Moreover, low academic achievement is a problem that negatively affects the individual
and the community, e.g., it may lead to unemployment. Therefore, it is essential to address such issues
by investigating the factors associated with students’ success and finding ways for early intervention
to help low-performing students (Jayaprakash & Jaiganesh, 2019).

122
 Alturki, Alturki, & Stuckenschmidt

Data Mining (DM) is widely applied in the Education field (Romero & Ventura, 2010) and is one of
the most popular techniques to analyze students’ performance. DM refers to extracting or “mining”
knowledge from large amounts of data to discover hidden patterns and relationships that are helpful
in decision making. Currently, there is an increase of research interest in using DM in education due
to its high potential in improving educational institutes (Baradwaj & Pal, 2011). The practice of DM
methods applied to educational data is known as Educational Data Mining (EDM) (Baker & Yacef,
2009). It concerns developing methods that discover knowledge from educational environment data
(Han et al., 2011) that are drawn from a variety of domains, including DM and machine learning,
psychometrics and other areas of statistics, information visualization, and computational modeling
(Romero & Ventura, 2007). EDM provides educators and students with useful insights into the edu-
cation process, resulting in suitable actions and decisions that improve academic success (Kotsiantis,
2009).
Predicting students’ academic performance is one of EDM’s main focuses (Fan et al., 2019). Such
prediction brings many advantages, for instance, providing valuable feedback, offering recommenda-
tions, supporting personalized learning, and strategically planning educational programs (Hellas et al.,
2018). There are three types of predictions in higher education: (i) predicting students’ academic per-
formance or GPA at a degree level; (ii) predicting students’ failure or drop out of a degree; and (iii)
predicting students’ results on particular courses (Alturki et al., 2020). In this study, we address the
first type. We provide a conceptual framework that may be used to create a recommender system that
uses classification algorithms to predict students’ academic achievement at an early stage. Such a sys-
tem will enable higher education institutions to create quality graduates and reduce student attrition.
Furthermore, it will assist educational institutions in early intervention to improve students’ perfor-
mance.
The rest of the paper is organized as follows. The next section presents the study’s overview and the
research questions, the background of the study, and the related work. Following that, the research
methodology is explained. Then we provide details of the experimental results and discussion. Fi-
nally, we conclude with a summary of the study’s primary outcomes, outline our research limitations,
and suggest future lines of research.

STUDY OVERVIEW
In this section, the research questions are outlined, the empirical study background is presented, and
related previous work is discussed.

R ESEARCH QUESTIONS
The four research questions proposed for this study are as follows:
RQ.1: What are the significant attributes for predicting student academic achievement in the
College of Computer and Information Sciences?
RQ.2: Is it possible to discover honorary students at an early stage of their studies?
RQ.3: Do academic language skills play a role in students’ success?
RQ.4: Does studying an Orientation year (or Preparatory year) affect students’ success?

T H E C OLLEGE OF C OMPUTER AND INFORMATION S CIENCES

At the studied College of Computer and Information Sciences, a 5.00-grade point average (GPA)
system is used for evaluating students’ achievement each semester. The final GPA is calculated by di-
viding the total quality points earned by the total number of credit hours for which grades were as-
signed. The college offers three majors: Computer Science (CS), Information Technology (IT), and
Information Systems (IS). These majors are within the same realm of study. However, each major

123
Predicting Students’ Academic Performance

focuses on specific aspects of the field of computer and information sciences. The CS major focuses
on the theory of computational applications, i.e., understanding the “why” behind computer pro-
grams. The major is based on algorithms and mathematics – the language of computers. CS students
also learn the fundamentals of programming languages, linear and discrete mathematics, and soft-
ware design and development. Students study the machine itself and understand how and why vari-
ous computer processes operate the way they do.
The IT major focuses more on network models and their protocols and the types of traffic gener-
ated, and their quality of service requirements. Students learn how to deal with performance issues in
networks and competence in the use of techniques to analyze and optimize performance. IT students
also focus on internet design principles, internet routing design, internet application protocols, and
cryptography and security.
Students enrolled in the IS major focus on meeting the needs of users in an organizational context
through the selection, creation, practical application, integration, and administration of computing
technologies. Students also learn how to use and apply current technical concepts and practices and
how to analyze, identify and define the requirements that must be satisfied to address IT problems or
opportunities faced by organizations or individuals. Moreover, students learn the fundamentals of
effectively designing IT-based solutions and integrating them into the user environment, along with
identifying and evaluating current and emerging technologies and discuss their applicability to solve
the users’ needs.
Previously, all three majors could be completed within five years, i.e., ten semesters (two semesters as
part of the orientation [Preparatory] year, eight semesters as part of the major). An orientation year
is a program designed to prepare students for their higher education. During the orientation year,
students are given courses of the English language and foundational undergraduate-level courses
such as basics in physics and mathematics. The aim of this year is to allow students to adjust to the
new academic environment and teaching system. However, the orientation year has been recently
eliminated from the college programs, making it possible to graduate within four academic years.
There are twelve core college courses shared between the three majors taught in the English lan-
guage. Programming Language (1), Programming Language (2), Database Fundamentals, and Com-
puter Networks Fundamentals, are among the introductory courses taught during the first two years
of all three majors.

R ELATED WORK ON P REDICTIONS IN H IGH ER E DUCATION

Substantial research on the effectiveness of teaching methods indicates that the quality of teaching is
often reflected by the achievements of learners (Ganyaupfu, 2013). Educators should apply appropri-
ate teaching methods that best suit specific objectives (Ganyaupfu, 2013). Adjusting the teaching ac-
cording to specific students’ needs can be achieved by using methods of EDM. Various EDM appli-
cations can be applied in educational institutes (Romero & Ventura, 2013). Predicting student aca-
demic performance is one of the leading applications. This section of the paper provides a literature
review on related previous studies and draws particular attention to the attributes that researchers
have widely been using to predict academic achievement.
Based on Alturki et al. (2020), the features that are used for predicting academic achievement can be
classified into three categories. They are: (i) demographics, (ii) pre-enrollment features, and (iii) post-
enrollment features. Although the demographical features are heavily used for predicting academic
achievement, the extent to which they are useful is unclear (Alturki et al., 2020). One of the top used
features in this category is gender (Aulck et al., 2016; Daud et al., 2017; Garg, 2018; Kovačić, 2010;
Osmanbegović & Suljic, 2012; Shakeel & Butt, 2015). However, some researchers found that gender
does not significantly impact the overall prediction (Kovačić, 2010; Osmanbegović & Suljic, 2012).
Age is also one of the common features used to predict academic achievement (Aulck et al., 2016;
Kemper, 2020; Kovačić, 2010; Yehuala, 2015). However, Kovačić (2010) reported that age does not

124
 Alturki, Alturki, & Stuckenschmidt

significantly impact predicting academic success. Country of origin is also used widely for predicting
academic achievement (Abu Saa, 2016; Aulck et al., 2016; Kemper, 2020).
Using pre-enrollment features to predict students’ academic achievement is significant, especially if
the prediction is to be performed at an early stage (Alturki et al., 2020). Previous GPA is one of the
most popular used features for predicting academic success (Abu Saa, 2016; Aluko et al., 2018; Garg,
2018; Huang & Fang, 2013; Kabakchieva, 2013; Kovačić, 2010; Osmanbegović & Suljic, 2012; Pal &
Pal, 2013; Thai-Nghe et al., 2007). Academic language skills have also been widely used for predicting
academic success (Abu Saa, 2016; Asif et al., 2017; Badr et al., 2016; Bani-Salameh, 2018; Thai-Nghe
et al., 2007). Although some researchers (e.g. Arsad et al., 2014) claim that academic language skills
do not affect students’ success in “non-linguistic courses,” others (e.g. Wait & Gressel, 2009) found a
significant relationship. Bani-Salameh (2018) performed a study on medical students’ performance
and its relation to the teaching language. Students were given the same test in two languages, namely
Arabic (native language), and English (teaching language), and compared their performance. The re-
sults indicate a weak correlation between students’ performance and the teaching language.
Using post-enrollment features for predicting students’ academic achievement can maximize the pre-
diction accuracy as such features represent students’ current situation in the program (Alturki et al.,
2020). The achievement of the previous semester, which is part of the total GPA, has been used in
many studies to predict students’ success (e.g. Abu Saa, 2016; Al luhaybi et al., 2018; Asif et al., 2017;
Kabakchieva, 2013; Thai-Nghe et al., 2007; Yehuala, 2015); that is, since students’ success is highly
dependent on previously acquired knowledge. Asif et al. (2017) found that the results of a four-year
program’s first and second-year courses play a significant role in predicting graduation performance.
Using students’ grades that are earned in quizzes and examinations have also been widely used in var-
ious studies for predicting students’ academic success (e.g. Al luhaybi et al., 2018; Aulck et al., 2016;
Badr et al., 2016; Huang & Fang, 2013; Kemper et al., 2020; Pradeep & Thomas, 2015; Shakeel &
Butt, 2015; Villwock et al., 2015; Yadav et al., 2011; Yassein et al., 2017). Failure in examinations has
also been used as a predictor; for instance, Kabakchieva (2013) used a dataset of 10,330 students to
predict their performance using five classes (bad, average, good, very good, and excellent) and found
that the number of failures at the first-year exams is among the most influencing features in the clas-
sification. It is relatively reasonable to use such a feature as a predictor as a “high rate of academic
failure is often observed in the first year of studies” (Gilar-Corbi et al., 2020, p. 1). Academic load,
measured in terms of credit hours and course difficulty per a single semester (Szafran & Austin,
2002), is also used to predict students’ success (e.g. Abu Saa, 2016; Yehuala, 2015). In fact, Yehuala
(2015) found that the number of credit hours is one of the main significant attributes for predicting
academic achievement.
In terms of the performance of different DM methods, there is no single method that works best in
all settings. Therefore, researchers often explore two or more DM methods to reveal which gener-
ates the best accuracy in their particular case (Alturki et al., 2020). Yohannes and Ahmed (2018) de-
signed an application to assist higher education institutions in predicting their students’ academic per-
formance upon graduation (eighth semester). They used students’ scores for core and non-core
courses from the first to the sixth semester and found that Support Vector Regression and Linear Re-
gression performed better than Neural Networks. A study by Khasanah and Harwati (2017) con-
ducted Feature Selection to find high influence attributes with students’ performance. They used stu-
dents’ demographics, pre-enrolment information, and post-enrolment information. They found that
Bayesian Network outperforms Decision Trees and that students’ attendance and students’ GPA in
the first semester are the two most important features for predicting academic achievement. Shakeel
and Butt (2015) used demographics, pre-enrollment features, and post-enrollment features to predict
students who are likely to drop out. They found that Naïve Bayes performed best, followed by Ran-
dom Forest, then J48, then Logistic Regression.
One identified research gap is that none of the reviewed studies explored the effect of the orienta-
tion year on predicting students’ academic achievement. McMullen (2014) emphasizes the

125
Predicting Students’ Academic Performance

importance of the orientation year on students’ success as they provided statistical proof that stu-
dents in Saudi Arabia believe that they leave secondary school without the skills necessary to enter
their academic majors. Since the orientation year has been shown to improve students’ attrition
(Davig & Spain, 2003), it is, therefore, essential to examine its impact on predicting students’ achieve-
ment.

RESEARCH METHODOLOGY
DATA C OLLECTION
Participants
The data set of 300 female students aged between 20 and 22 used in this study was obtained ran-
domly from a College of Computer and Information Sciences at a Saudi university. One hundred of
the collected records are from the IS major, one hundred from the IT major, and one hundred rec-
ords from the CS major. While 117 of the collected students have studied an orientation year, the re-
maining 183 have not. It is important to note that the GPA of the orientation year is not accounted
for those who have studied it, i.e., it has no effect on the final GPA. In the collected data, the major-
ity of the students had an “Accepted” GPA, and the minority had a “Poor” GPA. Figure 1 shows the
distribution of the students’ final academic achievement.

100
94
90
Students Final Academic Grade Distribution

82
80

70

60

50 49
44
40
31
30

20

10

0
Excellent Very Good Good Accepted Poor

Figure 1. Distribution of the final academic grade of the collected data

There are three types of features for predicting students’ academic achievement: demographics such
as gender and age, pre-enrollment features such as previous GPA, and post-enrolment features such
as grades achieved in each course (Alturki et al., 2020). In our study, we applied one pre-enrollment
feature, which is the ‘secondary school graduation percentage’, and the rest are post-enrollment fea-
tures, which are: two ‘English course grades’, ‘the GPA of the first four semesters’, ‘academic load in
first four semesters’, ‘number of failed courses in the first four semesters, and grades of the college’s
core courses. Although there are multiple shared courses between all three majors, we only consid-
ered the courses taken in the first four semesters as we aim to perform the prediction at an early
stage, i.e., before the third academic year. Table 1 below shows the dataset characteristics.

126
 Alturki, Alturki, & Stuckenschmidt

Table 1. Main characteristics of the collected dataset

Feature Description Type Value
Excellent, Very good,
GradGPA Graduation grade Nominal Good, Accepted, and
Poor
SecPer Secondary school achievement Numeric 0 – 100%
Major The major of the student Nominal CS, IS, and IT
Whether the student studied an orientation
PY Boolean Yes or no
year or not
GPA1 Student’s GPA in 1st academic semester Numeric 0-5
GPA2 Student’s GPA in 2nd academic semester Numeric 0-5
GPA3 Student’s GPA in 3rd academic semester Numeric 0-5
GPA4 Student’s GPA in 4th academic semester Numeric 0-5
Hrs/sem1 Student’s academic load per 1st semester Numeric 12 - 24 hours
Hrs/sem2 Student’s academic load per 2nd semester Numeric 12 - 24 hours
Hrs/sem3 Student’s academic load per 3rd semester Numeric 12 - 24 hours
Hrs/sem4 Student’s academic load per 4th semester Numeric 12 - 24 hours
Number of failed courses in the 1st aca-
F/year1 Numeric ≥1
demic year

Feature Description Type Value

Number of failed courses in the 2nd aca-
F/year2 Numeric ≥1
demic year
A+, A, B+, B, C+, C,
Prog1 Student’s grade in programming (1) Nominal
D+, D, and F
A+, A, B+, B, C+, C,
Prog2 Student’s grade in programming (2) Nominal
D+, D, and F
A+, A, B+, B, C+, C,
DB Student’s grade in Database’s fundamentals Nominal
D+, D, and F
Student’s grade in Computer Networks fun- A+, A, B+, B, C+, C,
NW Nominal
damentals D+, D, and F

Data Preprocessing
In the raw dataset, the final GPA is within the range of 0–5.0, where 5.0 is the best possible GPA
score. However, since the final GPA is in the form of an integer, and the predicted class should be
categorical (nominal) values, we transformed the GPA into five categories according to the grading
system. Table 2 shows the grading classification that is used in this study.

127
Predicting Students’ Academic Performance

Table 2. Classification of academic grading

GPA Grade Symbol
4.5 - 5 Excellent A
4.00 – 4.5 Very good B
3.25 – 4.00 Good C
2.5 – 3.25 Accepted D
Less than 2.5 Poor E

DATA M INING M ETH ODS

Data Mining (DM) methods can be classified into two categories: supervised and unsupervised meth-
ods. The unsupervised methods uncover hidden patterns in unlabeled data to find patterns in a da-
taset, and there are no output variables to predict. On the other hand, supervised methods (also
known as predictive or directive) predict the value of the output variables based on the inputs. In this
study, we address supervised DM methods to predict the value of the output variables based on the
inputs. To achieve this, a model is developed from training data where the values of inputs and out-
puts are previously labeled. The model generalizes the relationship between the inputs and outputs
and uses it to predict other datasets where only inputs are known (Witten et al., 2017).
Several classification algorithms can be used to predict the students’ graduation performance at the
end of the degree. However, the literature review suggests that, in general, there is no single classifier
that works best in all contexts to provide a good prediction. Following are the six DM methods ap-
plied in this study:
C4.5: The C4.5 (J48 in Weka) is an extension of the ID3 algorithm (Interactive Dichotomize 3). This
algorithm is used to generate a tree-shaped structure that represents sets of decisions. At each node
of the tree, C4.5 chooses the attribute of the data that most effectively splits its set of samples into
subsets enriched in one class or the other. The splitting measure is the normalized information gain.
The attribute with the highest normalized information gain is selected to make the decision.
Simple CART: The Classification and Regression Tree (CART) analysis uses a learning sample
which is a set of historical data with preassigned classes for all observations for building decision
trees. It is a learning technique that gives the results as either classification or regression trees, de-
pending on the categorical or numeric data set. It uses cross-validation or a large independent test
sample to select the best tree from the sequence of trees considered in the pruning process. During
the implementation phase of CART, the dataset is split into the two subgroups that are the most dif-
ferent concerning the outcome. This procedure is continued on each subgroup until some minimum
subgroup size is reached (Kalmegh, 2015).
LADTree: The Logical Analysis of Data (LAD) is a classification method built based on learning a
logical expression. Since LAD is a binary classifier, it can differentiate between positive and negative
samples (Amudha et al., 2011). For a dataset processed by LAD, a large set of patterns are produced,
and a subset of them is selected to satisfy the above assumption such that each pattern in the model
satisfies specific requirements in terms of prevalence and homogeneity (Buhmann, 2003).
Naïve Bayes: The Naive Bayes classifier, which is based on the work of Thomas Bayes, simplifies
learning by assuming that features are independent given class. The Naïve Bayes follows the hypothe-
sis that the data belongs to a particular class, then the probability for the hypothesis is calculated to
be true. Thus, only one scan of the data is required. In cases of training data, each training example
can incrementally increase/decrease the probability that a hypothesis is correct. Thus, Naïve Bayes
perfectly fits domains containing uncertainty (Nielsen & Jensen, 2007).

128
 Alturki, Alturki, & Stuckenschmidt

Bayes Net with ADTree: A Bayesian network is a probabilistic graphical model representing a set
of variables and their conditional dependencies via a directed acyclic graph. They are ideal for taking
an event that occurred and predicting the possibility that any of several possible known causes was
the contributing factor. An Alternating Decision Tree (ADTree) is a machine learning method for
classification. It generalizes decision trees and has connections to boosting. It consists of an alterna-
tion of decision nodes that specify a predicate condition, and prediction nodes that contain a single
number. An ADTree classifies an instance by following all paths for which all decision nodes are true
and summing any prediction nodes that are traversed. Bayes Net associated with an ADTree com-
bines both methods to improve the classification accuracy.
Random Forest: As the name implies, a Random Forest is a tree-based classifier that functions as
an ensemble depending on a collection of random variables (Cutler et al., 2012). Random Forest is a
mixture of tree predictors such that each tree depends on the values of a random vector sampled in-
dependently and with the same distribution for all trees in the forest. In the training stage, random
forests apply the general technique known as “bagging” to individual trees in the ensemble (Caie et
al., 2021). Bagging repetitively chooses a random sample with replacement from the training set and
fits trees to these samples where each tree is grown without any pruning.

DATA M INING T OOL

For this study, we used the WEKA software package developed at the University of Waikato in New
Zealand. The name stands for Waikato Environment for Knowledge Analysis. This package has been
implemented in the Java software language and stands out as one of the most competent and com-
prehensive packages with machine learning algorithms. WEKA supports several standard data mining
tasks, specifically data preprocessing, clustering, classification, regression, visualization, and feature
selection (Kalmegh, 2015).

EXPERIMENTAL RESULTS
Different options are available for evaluating performance measures. A few of them are cross-valida-
tion using a different number of folds and percentage split, which is used to split the dataset into two
partitions – one is used for training the dataset, and the other is used for testing. In this study, a per-
centage split is applied, i.e., 70% of the dataset has been used for training the model, and the remain-
ing 30% is used for testing purposes. Data sets were randomly partitioned into training and testing
datasets via 10-fold CV. In 10-fold CV, the data set is divided into 10-subsets, and the holdout
method is repeated ten times. Each time, one of the ten subsets is used as the test set, and the other
nine subsets are put together to form the training set.

I NDIVIDUAL F EATURE ANALYSIS

As selecting the right features improves classifiers’ performance, we have explored different feature
selection techniques to understand the influence of different features on the target “final grade”.
They are Search-Based, Correlation Based, Information Gain Based, and Wrapper with Naïve Bayes.
The search method is either BestFirst, which searches the space of attribute subsets by greedy hill-
climbing augmented with a backtracking facility, or Ranker, which ranks attributes by their evalua-
tions. Table 3 describes the selected methods and their results.

129
Predicting Students’ Academic Performance

Table 3. Comparison between different attribute selection methods

Search Selected
Attribute Evaluator Description
Method Attributes
Evaluates the effect of a sub- GPA1 (100%)
set of features by considering GPA2 (100%)
the predictive ability of each
Search Based one along with the degree of BestFirst GPA3 (100%)
redundancy between them. GPA4 (100%)
Subsets of features that are DB (100%)
(CfsSubsetEval) highly correlated with the F/year1 (80%)
class while having low inter-
correlation are preferred NW (30%)

1. GPA3
2. GPA4
Evaluates the influence of an 3. F/year2
Correlation Based
attribute by measuring the Ranker 4. GPA2
correlation (by Pearson for- 5. GPA1
mula) between attribute and
(CorrelationAttributeEval) 6. F/year1
the target attribute.
7. SecGPA
8. DB
9. Hrs/sem4
1. GPA3
2. GPA4
3. DB
Information Gain Based Evaluates the influence of an
4. GPA2
attribute by measuring the in-
Ranker 5. GPA1
formation gain concerning the
(InfoGainAttributeEval) target attribute. 6. prog1
7. prog2
8. F/year2
9. F/year1
GPA3 (100 %)
Evaluates attribute sets by us- GPA4 (100 %)
Wrapper with Naïve Bayes
ing a learning scheme. Cross GPA1 (90%)
validation is used to estimate BestFirst NW (90%)
the accuracy of the learning
(WrapperSubsetEval) F/year1 (70%)
scheme for a set of attributes.
DB (60%)
Prog1 (40%)

Based on the results provided above, we conclude that the attributes that have the most influence on
academic achievement are the student’s GPA for each semester, the number of failed courses in 1st
and 2nd year, their grade in the ‘Database fundamentals’, and ‘Programming 1’ core courses. On the
other hand, ‘English skills’, ‘secondary school GPA’, ‘academic load’, and ‘programming 2’ are found
surprisingly to have no effect on the prediction of students’ academic achievement.

130
 Alturki, Alturki, & Stuckenschmidt

P ERFORMANCE E VALUATION OF DIFFERENT DM M ETH ODS

In this section, two case studies are provided for predicting students’ academic achievement. In the
first case, the prediction is performed after the 3rd academic semester whereas, in the second, it is
performed after the 4th semester. In both cases, we compare the performance of six classifiers in
terms of the following: (1) classifier accuracy, which is the total number of correct predictions di-
vided by the total number of predictions made for a dataset; (2) Receiver Operating Characteristic
(ROC) to examine the performance of a binary classifier by creating a graph of the True Positives vs
False Positives for every classification threshold (the higher the ROC, the better the results); (3) F
Measure, which provides a way to combine both precisions and recall into a single measure that cap-
tures both properties (a poor F-Measure score is 0.0 and a best F-Measure score is 1.0); and finally
(4) the accuracy of allocating excellent students.

Predicting final academic grade after the 3rd semester

In this academic prediction case, six attributes are selected: the GPA from the 1st, 2nd, and 3rd semes-
ter; the number of failed courses during the 1st year; and the grades of the two core courses that are
taken during the first three semesters, i.e., Programming Language (1), and Database Fundamentals.
Table 4 below compares the different classifiers’ results in predicting students’ academic achievement
after their 3rd semester. It can be observed that Naïve Bayes and Random Forest perform the best
with an accuracy of 63.33% and 63%, respectively. Those two classifiers also perform the best in pre-
dicting honorary students with an equal accuracy of 77.6%.
Table 4. Accuracy of predicting final academic grade after the 3rd semester
Honorary
Classifier Accuracy ROC F Measure students
(Accuracy)

J48 55.67% .783 .549 71.4%

SimpleCart 59.67% .824 .576 75.5%

LADTree 58.33% .850 .585 75.5%

Naïve Bayes 63.33% .892 .637 77.6%

Bayes Net with ADTree 60.33% .877 .603 77.6 %

Random Forest 63% .872 .615 85.8%

Predicting final academic grade after the 4th semester

In this academic prediction case, nine attributes are selected to predict students’ graduation grades:
the GPA from semesters 1, 2, 3, and 4, number of failed courses during the 1st year, number of
failed courses during the 2nd year, Programming Language (1) grade, Database Fundamentals grade,
and Computer Networks Fundamentals grade. Table 5 compares the different classifiers’ results in
predicting students’ academic achievement after their 4th semester. We can notice that Random Forest
performs the best compared to the rest of the Tree-based Models with 67.6% accuracy. However,
Naïve Bayes outperformed all five tree-based classifiers with an accuracy of 69.67%. In terms of
which classifier performs the best in predicting honorary students, Random Forest significantly out-
performed the rest of the classifiers with 92.6% accuracy.

131
Predicting Students’ Academic Performance

Table 5. Accuracy of predicting final academic grade after the 4th semester
Honorary
Classifier Accuracy ROC F Measure students
(Accuracy)
J48 61.67% .815 .611 77.6%
SimpleCart 63.3% .833 .633 79.6%
LADTree 63% .875 .630 75.5%
Naïve Bayes 69.67% .917 .697 77.6%
Bayes Net with ADTree 66.33% .911 .666 79.6%
Random Forest 67.6% .909 .660 92.6%

DISCUSSION OF TH E R ESULTS
In this section, we answer our research questions and discuss the results obtained from the study.
As a general observation, it is clear that, with the increase of attributes, the model’s accuracy in-
creases as all six classifiers performed better in the second case than they did in the first; that is, they
acquired a higher accuracy, ROC, and F measure. In terms of which classifier performed better in
predicting academic achievement, the results are similar to the results of Shakeel and Butt (2015). We
found that Random Forest performs the best compared to the rest of the Tree-based Models with
63% and 67.6% accuracy in the first and second case respectively. Random Forests generally yield
better results as they are much more robust than a single decision tree; that is, they aggregate many
decision trees to limit error and overfitting due to bias. While single decision trees search for the
most important feature when splitting a node, Random Forests search for the best feature among a
random subset of features. Such wide diversity usually results in a better model. The second-best
Tree-based classifier was Bayes Net with ADTree with an accuracy of 60.33% and 66.33% in the first
and second case respectively. In binary classification trees, such as J48, CART, an instance follows
only one path through the tree, ADTree follows all paths for which all decision nodes are true and
sums any prediction nodes that are traversed. However, Naïve Bayes outperformed all five tree-based
classifiers with an accuracy of 63.33% and 69.67% in the first case and second case respectively. The
reason behind the good performance of Naïve Bayes is described by Domingos & Pazzani (1996) as
follows: “Naïve Bayes is commonly thought to be optimal, in the sense of achieving the best possible
accuracy, only when the independence assumption holds, and perhaps close to optimal when the at-
tributes are only slightly dependent. However, this very restrictive condition seems to be inconsistent
with the Naïve Bayes’ surprisingly good performance in a wide variety of domains, including many
where there are clear dependencies between the attributes.”
Referring to our first research question (“what are the significant attributes for predicting student ac-
ademic achievement in the College of Computer and Information Sciences?”), we found that stu-
dents’ earned GPA in all four semesters plays a significant role in predicting students’ academic
achievement. This supports the findings of Asif et al. (2017). However, the students’ earned GPAs in
the 3rd and 4th semesters are found to have more influence on the prediction than the 1st and 2nd se-
mesters. That is relatively reasonable as the courses become more challenging as students escalate in
semesters and the variation in students’ skills start to show more. We have also found that the num-
ber of failed courses in the first two years of the program plays a significant role in predicting stu-
dents’ academic achievement. This is in line with the results of Kabakchieva (2013). In terms of
which fundamental course has the largest effect on students’ success, we found that the Database
Fundamentals course had the highest impact. On the other hand, we have unexpectedly found that
the academic load does not play a significant role in predicting students’ success. This contradicts the

132
 Alturki, Alturki, & Stuckenschmidt

findings of Yehuala (2015). Although previous GPA from secondary school is one of the most pop-
ular used features for predicting academic success (Abu Saa, 2016; Aluko et al., 2018; Garg, 2018;
Huang & Fang, 2013; Kabakchieva, 2013; Kovačić, 2010; Osmanbegović & Suljic, 2012; Pal & Pal,
2013; Thai-Nghe et al., 2007), we have found it to have no significant effect on the overall prediction.
When it comes to our second research question regarding the ability to predict honorary students, we
can conclude that it is possible to identify honorary students at an early stage of their bachelor’s stud-
ies at Computer and Information Sciences colleges. In terms of which classifier performed best in
such prediction, Random Forest outperformed the other Tree-based classifiers and outperformed
Naïve Bayes with an accuracy of 85.8% in the first case and 92.6% in the second.
To answer our third research question regarding whether academic language skills play a role in stu-
dents’ success, we found that English language skills did not play a role in students’ success. This is in
accordance with the findings of Arsad et al. (2014) and Bani-Salameh (2018) and might be due to the
fact that the nature of courses under the College of Computer and Information Science is not lin-
guistics and rather more scientific.
The fourth research question in this paper was related to investigating whether studying an Orienta-
tion year affects students’ success. Although the orientation year has shown to improve students’ aca-
demic achievement (Davig & Spain, 2003; McMullen, 2014), we found that it has no effect on stu-
dents’ success in the College of Computer and Information Science and can therefore conclude that
removing the orientation year from the study programs was a reasonable action that did not carry
drawbacks on the students.

CONCLUSION
The capabilities of the DM techniques can provide useful insights for predicting the final academic
performance of students. This type of prediction can help both teachers and students in many ways.
It can also enable the early discovery of honorary students. Thus, well-deserved opportunities can be
offered, e.g., scholarships, internships, and workshops. It can also help identify students who require
special attention to take an appropriate intervention at the earliest stage possible. Moreover, instruc-
tors can be aware of each student’s capability and thus can customize the teaching tasks based on stu-
dents’ needs, e.g., offering extracurricular learning material to students facing difficulties, using differ-
ent teaching strategies, and providing online tutoring videos for those who need.
The main objectives of this study were to predict students’ final grades at a degree level and identify
honorary students at an early stage of their studying journey. We tested the significance of correla-
tion attributes predictors using four different methods. We have found 9 out of 18 proposed features
with a significant correlation for predicting students’ academic achievement after their 4th semester.
Those features are student GPA during the first four semesters, the number of failed courses during
the first four semesters, and the grades of three core courses, i.e., Database Fundamentals, Program-
ming Language (1), and Computer Network Fundamentals. We performed the prediction of stu-
dents’ academic achievement by investigated six supervised DM algorithms. As the classifiers’ perfor-
mances are evaluated based on their predictive accuracy, we found that Naïve Bayes performed better
than the five selected Tree-based Models for predicting students’ final graduation grades. However,
Random Forest performed significantly better than Naïve Bayes in predicting honorary students with
an accuracy of 85.8%, 92.6% after the third and fourth semesters, respectively.
Evaluating bachelor’s programs should be an ongoing cycle. Therefore, it was essential to make sure
that eliminating the orientation year from the College of Computer and Information Systems was an
ideal change. The study results reveal existing evidence that the orientation year does not have an im-
pact on students’ success.
One of the limitations of this present study is that we did not investigate the influence of gender
and age on the prediction of academic achievement, since all of the participants are females and

133
Predicting Students’ Academic Performance

belong to the same age group. Also, due to time limitations, this research has been carried out includ-
ing only students from one college and one university.
For future work, the results of this study can be used to design a recommender system that enables
timely interventions at Computer and Information Systems Colleges. The experiment can be re-
peated with a larger dataset. It could also be extended with more distinctive attributes to reach more
accurate results helpful in improving the students’ learning outcomes, e.g., gender, students’ attend-
ance, students’ e-learning activity information, and students’ satisfaction. Future research could also
investigate the effects of offering tutoring sessions for students who fail core courses in their first
semesters, examine the role of language skills in social science programs, and examine the role of
orientation year in other programs.
We would recommend instructors to consider using EDM in predicting students’ academic achieve-
ment and benefit from that in customizing students’ learning experience based on their different
needs. We would also highly endorse that researchers apply more EDM studies across different uni-
versities and compare between them.

ACKNOWLEDGMENT
This research was funded by the Deanship of Scientific Research at Princess Nourah bint Abdulrah-
man University through the Fast-track Research Funding Program.

REFERENCES
Abu Saa, A. (2016). Educational data mining & students’ performance prediction. International Journal of
Advanced Computer Science and Applications, 7(5): 212-200. [Link]
Al luhaybi, M., Tucker, A., & Yousefi, L. (2018). The prediction of student failure using classification methods:
A case study. Computer Science & Information Technology, 79-90. [Link]
Alturki, S., Hulpus, I., & Stuckenschmidt, H. (2020). Predicting academic outcomes: A survey from 2007 till
2018. Technology, Knowledge and Learning, 1-33. [Link]
Aluko, R. O., Daniel, E. I., Shamsideen Oshodi, O.., Aigbavboa, C. O., & Abisuga, A. O. (2018). Towards
reliable prediction of academic performance of architecture students using data mining techniques. Journal
of Engineering, Design and Technology, 16(3), 385-397. [Link]
Amudha, J., Soman, K., & Kiran, Y. (2011). Feature selection in top-down visual attention model using WEKA.
International Journal of Computer Applications, 24(4). [Link]
Arsad, P. M., Buniyamin, N., & Manan, J. A. (2014). Students’ English language proficiency and its impact on
the overall student’s academic performance: An analysis and prediction using Neural Network Model.
WSEAS Transactions on Advances in Engineering Education, 11, 44-53.
[Link]
Asif, R., Merceron, A., Ali, S. A., & Haider, N. (2017). Analyzing undergraduate students’ performance using
educational data mining. Computers & Education, 113, 177-194.
[Link]
Aulck, L., Velagapudi, N., Blumenstock, J., & West, J. (2016). Predicting student dropout in higher education.
Proceedings of the ICML Workshop on #Data4Good: Machine Learning in Social Good Applications, New York.
[Link]
Badr, G., Algobail, A., Almutairi, H., & Almutery, M. (2016). Predicting students’ performance in university
courses: A case study and tool in KSU Mathematics Department. Procedia Computer Science, 82, 80-89.
[Link]
Baker, R., & Yacef, K. (2009). The state of educational data mining in 2009: A review and future visions. Journal
of Educational Data Mining, 1(1), 3-17. [Link]
Bani-Salameh, H. N. (2018). Teaching language effects on students’ performance. Health Professions Education,
4(1), 27-30. [Link]

134
 Alturki, Alturki, & Stuckenschmidt

Baradwaj, B. K., & Pal, S. (2011). Mining educational data to analyze students’ performance. International Journal
of Advanced Computer Science and Applications, 2(6), 63-69. [Link]
Buhmann, M. D. (2003). Radial basis functions: Theory and implementations (1st ed.). Cambridge University Press.
[Link]
Caie, P. D., Dimitriou, N., & Arandjelović, O. (2021). Precision medicine in digital pathology via image analysis
and machine learning. In S. Cohen (Ed.), Artificial intelligence and deep learning in pathology (pp. 149-173).
Elsevier. [Link]
Cutler, A., Cutler, D. R., & Stevens, J. R. (2012). Random forests. In C. Zhang & Y. Ma (Eds.), Ensemble machine
learning (pp. 157-175). Springer. [Link]
Daud, A., Aljohani, N. R., Abbasi, R., Lytras, M. D., Abbas, F., & Alowibdi, J. S. (2017). Predicting student
performance using advanced learning analytics. Proceedings of the 26th International Conference on World Wide
Web Companion, 415-421. [Link]
Davig, W. B., & Spain, J. W. (2003). Impact on freshmen retention of orientation course content: Proposed
persistence model. Journal of College Student Retention: Research, Theory & Practice, 5(3), 305-323.
[Link]
Domingos, P. M. & Pazzani, M. (1996). Beyond independence: Conditions for the optimality of the simple
bayesian classifier Proceedings of the 13th ICML Conference on on Machine Learning, 105–112.
[Link]
Fan, Y., Liu, Y., Chen, H., & Ma, J. (2019). Data mining-based design and implementation of college physical
education performance management and analysis system. International Journal of Emerging Technologies in
Learning, 14(6), 87-97. [Link]
Ganyaupfu, E. M. (2013). Teaching methods and students’ academic performance. International Journal of
Humanities and Social Science Invention, 2(9), 29-35.
Garg, R. (2018). Predicting student performance of different regions of Punjab using classification techniques.
International Journal of Advanced Research inComputer Science, 9(1), 236-241.
[Link]
Gilar-Corbi, R., Pozo-Rico, T., Castejón, J. L., Sánchez, T., Sandoval-Palis, I., & Vidal, J. (2020). Academic
achievement and failure in university studies: Motivational and emotional factors. Sustainability, 12(23), 1-14.
[Link]
Han, J., Kamber, M., & Pei, J. (2011). Data mining: Concepts and techniques (3rd ed.). Elsevier.
Hellas, A., Ihantola, P., Petersen, A., Ajanovski, V., utica, M., Hynninen, T., Knutas, A., Leinonen, J., Messom,
C., & Liao, S. (2018). Predicting academic performance: A systematic literature review. In B. Scharlau & G.
Rossling (Eds.), Proceedings of the 23rd Annual ACM Conference on Innovation and Technology in Computer Science
Education. Association for Computing Machinery. [Link]
Huang, S., & Fang, N. (2013). Predicting student academic performance in an engineering dynamics course: A
comparison of four types of predictive mathematical models. Computers and Education, 61, 133-145
[Link]
Jayaprakash, S., & Jaiganesh, V. (2019 ) A conceptual framework to predict academic performance of students
using classification algorithm. International Research Journal of Engineering and Technology, 6(9), 721-733.
Kabakchieva, D. (2013). Predicting student performance by using data mining methods for classification.
Cybernetics and Information Technologies, 13(1), 61-72. [Link]
Kalmegh, S. (2015). Analysis of WEKA data mining algorithm REPTree, simple cart and randomtree for
classification of Indian news. International Journal of Innovative Science, Engineering & Technology, 2(2), 438-446.
[Link]
Kemper, L., Vorhoff, G., & Wigger, B. U. (2020). Predicting student dropout: A machine learning approach.
European Journal of Higher Education, 10(1), 28-47. [Link]

135
Predicting Students’ Academic Performance

Khasanah, A. U., & Harwati, H. (2017). A comparative study to predict student’s performance using
educational data mining techniques. IOP Conference Series: Materials Science and Engineering, 215(1).
[Link]
Kotsiantis, S. (2009). Educational data mining: A case study for predicting dropout-prone students. International
Journal of Knowledge Engineering and Soft Data Paradigms, 1(2), 101.
[Link]
Kovačić, Z. J. (2010, June). Early prediction of student success: Mining students enrolment data. Proceedings of
the 2010 Informing Science and Information Technology Education Conference, Cassino, Italy, 647-665.
[Link]
McMullen, M. (2014). The value and attributes of an effective preparatory English program: Perceptions of
Saudi University students. English Language Teaching, 7(7). [Link]
Nielsen, T. D., & Jensen, F. V. (2007). Bayesian networks and decision graphs (2nd ed.). Springer.
Osmanbegović, E., & Suljic, M. (2012). Data mining approach for predicting student performance. Journal of
Economics and Business, 10(1), 3-12.
Pal, A. K., & Pal, S. (2013). Analysis and mining of educational data for predicting the performance of
students. International Journal of Electronics Communication and Computer Engineering, 4(5), 1560-1565.
Pradeep, A., & Thomas, J. (2015). Predicting college students dropout using EDM techniques. International
Journal of Computer Applications, 123(5), 26-34. [Link]
Romero, C., & Ventura, S. (2007). Educational data mining: A survey from 1995 to 2005. ScienceDirect, 33(1),
135-146. [Link]
Romero, C., & Ventura, S. (2010). Educational data mining: A review of the state of the art. IEEE Transactions
on Systems, Man and Cybernetics, Part C: Applications and Reviews, 40(6), 601-618.
[Link]
Romero, C., & Ventura, S. (2013). Data mining in education. WIREs Data Mining and Knowledge Discovery, 3(1),
12-27. [Link]
Shakeel, K., & Butt, N. (2015, May). Educational data mining to reduce student dropout rate by using
classification. Proceedings of the 253rd OMICS International Conference on Big Data Analysis & Data Mining,
Lexington, Kentucky.
Szafran, R. F., & Austin, S. F. (2002). The effect of academic load on success for new college students: Is lighter
better? NACADA Journal, 22(2), 26-38. [Link]
Thai-Nghe, N., Janecek, P., & Haddawy, P. (2007). A comparative analysis of techniques for predicting
academic performance. 37th Annual Frontiers in Education Conference - Global Engineering: Knowledge without
Borders, Opportunities without Passports, T2G-7-T2G-12. [Link]
Villwock, R., Appio, A., & Andreta, A. A. (2015). Educational data mining with focus on dropout rates.
International Journal of Computer Science and Network Security, 15(3), 17-23.
Wait, I. W., & Gressel, J. W. (2009). Relationship between TOEFL score and academic success for international
engineering students. Journal of Engineering Education, 98(4), 389-398. [Link]
9830.2009.tb01035.x
Witten, I. H., Frank, E., Hall, M. A., & Pal., C. J. (2017). Part II: More advanced machine learning schemes,
Data mining: Practical machine learning tools and techniques (4th ed, pp. 205-208). Morgan Kaufmann.
[Link]
Yadav, S. K., Bharadwaj, B., & Pal, S. (2011). Data mining applications: A comparative study for predicting
student’s performance. International Journal of Innovative Technology and Creative Engineering, 1(12), 13-19.
Yassein, N. A., Gaffer, R., Helali, M., & Mohomad, S. B. (2017). Predicting student academic performance in
KSA using data mining techniques. Journal of Information Technology & Software Engineering, 7(5).
[Link]

136
 Alturki, Alturki, & Stuckenschmidt

Yehuala, M. A. (2015). Application of data mining techniques for student success and failure prediction (the
case of Debre_Markos University). International Journal of Scientific & Technology Research, 4(4), 91-94.
Yohannes, E., & Ahmed, S. (2018). Prediction of student academic performance using neural network, linear
regression and support vector regression: A case study. International Journal of Computer Applications, 180(40),
39-47. [Link]

AUTHORS
Sarah Alturki is currently a Ph.D. student working with the Data and
Web Science group at the School of Business Informatics and Mathemat-
ics at the University of Mannheim, Germany. She earned her Master’s de-
gree in Applied Computer Science from SRH University of Applied Sci-
ences in Heidelberg, Germany. She is currently focusing on projects and
research related to empowering technology in higher education.

Nazik Alturki is an Assistant Professor at the Information Systems De-

partment, College of Computer and Information Sciences, Princess Nou-
rah bint Abdulrahman University, Saudi Arabia. She earned her Ph.D. de-
gree in Information Systems from the University of Melbourne, Aus-
tralia. Her research interests include Health Informatics, Big Data, Data
Analytics, and Data Mining.

Heiner Stuckenschmidt is a Full Professor of the Artificial Intelligence

group at the University of Mannheim. He is carrying out fundamental
and applied research in Knowledge Representation and Reasoning as well
as Machine Learning applications and Social and Economic Data Science.
He is the author of more than 250 publications in international journals
and conferences. More information about Heiner and his work can be
found here:
[Link]
enschmidt/

137