Predicting Student Enrolments and Attrition Patterns in Higher Educational Institutions using Machine Learning

Full text

562 The International Arab Journal of Information Technology, Vol. 18, No. 4, July 2021
Predicting Student Enrolments and Attrition
Patterns in Higher Educational Institutions using
Machine Learning
Samar Shilbayeh and Abdullah Abonamah
Business Analytics Department, Abu Dhabi School of Management, UAE
Abstract: In higher educational institutions, student enrollment management and increasing student retention are fundamental
performance metrics to academic and financial sustainability. In many educational institutions, high student attrition rates are
due to a variety of circumstances, including demographic and personal factors such as age, gender, academic background,
financial abilities, and academic degree of choice. In this study, we will make use of machine learning approaches to develop
prediction models that can predict student enrollment behavior and the students who have a high risk of dropping out. This can
help higher education institutions develop proper intervention plans to reduce attrition rates and increase the probability of
student academic success. In this study, real data is taken from Abu Dhabi School of Management (ADSM) in the UAE. This
data is used in developing the student enrollment model and identifying the student’s characteristics who are willing to enroll in
a specific program, in addition to that, this research managed to find out the characteristics of the students who are under the
risk of dropout.
Keywords: Machine learning, predictive model, apriori algorithm, student retention, enrolment behaviour, association rule
mining, boosting, ensemble method.
Received March 10, 2020; accepted July 19, 2020
https://doi.org/10.34028/18/4/8
1. Introduction
Machine learning can play a critical role in building up
a strategic enrolment plan for academic institutions.
This includes detecting the enrolment trends over time,
which gives the institution a better ability to manage
student enrolments and align institutional resources
more efficiently and effectively.
Higher education institutions are increasingly
interested in identifying their enrolment pattern to
understand the factors that maximize their enrolment
numbers. These factors allow the student recruitment
department to identify, attract, engage and enrol the
right students. The direct result of this is to help the
institutions to improve their financial and academic
performance. In academic institutions, student attrition
continues to be a significant and costly challenge. It is
considered to be the main concern that negatively
impacts the institution's reputation, ranking, and
success. The attrition rate in the academic institutions is
inspired by churn analysis that is usually performed in
marketing analysis. It is found out that attracting a new
customer in an organization is much more expensive
than retaining existing ones. In the same way, for any
academic institution, maintaining the current students is
one of their higher concern. This can be accomplished
by understanding the student attrition pattern by
detecting the students who are under the risk of attrition,
then building up a general
model that can predict this incident in advance. In this
way, they can develop proper strategies to retain those
students and give them more support if needed.
This paper closely investigates the enrolment pattern
that helps in attracting the type of students who are most
likely to enrol in one of the Abu Dhabi School of
Management (ADSM) academic programs and also
detects the student attrition patterns.
2. Literature Review
There are many enrolment and attrition rate prediction
studies that have utilized machine learning approaches
to identify the student enrolment and attrition pattern.
Petkovski et al. [7] conducted a study focused on the
student's performance as an indicator for the student
dropout rate in their freshman year. This study uses
naïve Bayes, decision tree, and rule-based induction
machine learning algorithm to build the best model that
predicts the attrition rate with 80% accuracy. Yukselturk
et al. [12] examined on their study the dropouts rate in
an online program for 189 students who registered in the
online information technology certificate program. This
study uses four machine learning algorithms K Nearest
Neighbour KNN, Decision Tree (DT), Naive Bays (NB)
and Neural Network (NN), resulting in 87% accuracy
for NN, and 80% accuracy for DT and finding out that
the student demographics information plays a very
critical role in their dropout rate. Rai and Jain [8]

Predicting Student Enrolments and Attrition Patterns in Higher Educational... 563
performed a study for the aim of understanding the
student drop rate factors using machine learning
algorithms utilizes ID3 and J48 decision trees using
weka on 220 undergraduate students in Information
Technology courses. In this research; it is found out that
personal factors is the most important factor that effects
the student attrition and weights for 28% of dropout rate.
On the same study, the institutional factors such as the
university environment, and the course cost weight for
17% of dropout rate, also it is noticed that few students
are most likely to drop out due to homesickness and
adjustment problems [8]. In the same prospect; kemper
et al. [4] managed to predict the student dropout rate and
success factor using logistic regression and decision
trees on a resample examination data at the Karlsruhe
Institute of Technology (KIT) with 95% accuracy after
three semesters of the student enrolments.
On the other hand, multiple authors and studies use
machine learning predictive methods and data science
techniques to predict student behaviour and
performance in educational settings. The researchers
Mulugeta and Berhanu [5] performed three different
machine learning algorithms: mainly J48, naïve bayes,
and neural network to build three prediction models that
predict the student's enrolment at the department level
for higher education institutions in both private and
public universities. The developed model in [5] shows
that the highest model accuracy is reached by applying
neural network machine learning algorithm over J48
and naïve Bayes. In the same prospect; Nakhkob and
Khademi [6] conducted a study that utilized neural
network, bagging, boosting, and naïve bayes to predict
the rate of student enrolment in higher education in
different universities in Iran. In the same study, a
comparison between those four models is performed
using model accuracy, and ROC curve. The result shows
that bagging outperforms neural network, boosting and
naïve Bayes.
Another study conducted by Waters and
Miikkulainen [11] utilized machine learning algorithms
to predict how likely 588 Ph.D. applicants will be
admitted to their colleges based on their information
provided in their applications, the result is used to
reduce the admission process time.
Two machine learning approaches: logistic
regressions, and decision trees are performed to predict
student dropout at Karlsruhe Institute of Technology
(KIT), both used models yield high prediction up to 95%
after three semesters, this study is done by Kemper et al.
[4]. For the same aim; a study is done by Al-Shabandar
et al. [2]. This study utilized two machine learning
models to find out the students who under the risk of
failure and withdraw at the early stages of online
courses. Random forest on decision tree and gradient
boosting machine are used for this aim and result in 89
% and 95% accuracy respectively.
3. Research Methodology
3.1. Data Description and Understanding
Data understanding and description is an important
stage in data analysis, which includes gathering,
understanding, cleaning, describing and verifying the
data. More accurate, comprehensive and higher quality
data will result in more efficient output. The data used
for this research is taken from the Student Information
System (SIS) database of Abu Dhabi School of
Management (ADSM). This particular dataset is used to
capture information about the students who enrolled in
ADSM along with their demographics and academic
background information for 1600 students enrolled
from (2013-2018). Student dataset variables are
described in Table 1.
Data Cleansing process includes filling up the
missing values, deleting the outliers or incorrectly input
data to make sure that the models will be built on clean,
proper data that can be converted into interesting results.
Table 1. Student dataset variables.
Variable name
Description
Program
This includes four master programs available
at ADSM
Workplace
The student workplace
UGSChool
Student undergraduate school
UGMajor
Student undergraduate major
Status
Student enrolment status (active, inactive,
under process, admitted)
UGCountry
Student undergraduate country
Experience Year
Number work experience years
Age
Student age
WorkPlaceSector
Student workplace sector divided into
(private, public)
Position
Working student current position
UGGPA
Undergraduate GPA
Emirates
The emirate region (state)
Nationality
Student Nationality
ELR Score
English Test Score
Sponsor
The name of the organization sponsoring the
student (if any)
Start Year
The student starting year in the academic
program
3.2. Predictive Analysis
For the aim of developing the proposed models, this
research is divided into three stages, the following is the
description for each stage with its results:
 Stage 1: For the aim of predicting the number of
enrolment for the coming four years (2019-2022), a
boosted regression tree algorithm is applied to six
years of data on 1600 students. The boosted tree is
compared with the single regression tree. The
boosting method initially is defined by Schapire and
Freund [10] as a classification supervised ensemble
method. Boosting is one type of ensemble
approaches and defined as the method of applying the
same machine learning algorithm (regression
decision tree in our case) in the whole dataset several
times. On each learning stage; the weight of correctly
classified instances is decreased, and the weight of

564 The International Arab Journal of Information Technology, Vol. 18, No. 4, July 2021
incorrectly classified instances is increased to give
those that are incorrectly classified in the previous
learning stage more attention in the next learning
process [3]. Figure 1 shows the stage 1 framework.
Figure 1. Stage 1 framework.
The final result of the boosted regression model
prediction is the average of those different created
models. In our work, 500 regression trees are created.
The final result of applying boosted regression model is
shown in Figure 3. The boosted regression model is
tested using 10 folds cross validation and gives 89%
accuracy which outperforms the single regression
decision tree that gives only 76% accuracy tested using
10 folds cross-validation as well.
 Stage 2: The association rule Apriori algorithm is
applied to the given dataset for the aim of revealing
interesting patterns about the characteristics of the
students who are most likely to enrol in one of the
school academic programs. Association rules mining
is one of the machine learning techniques that is used
to reveal interesting and meaningful hidden patterns
between different dataset variables. The developed
rules are interestingly unexpected for business
stakeholders. Apriori algorithm is defined by
Agrawal et al. [1]. Apriori algorithm is used to
uncover interesting relations between different items
in the market basket analysis. In Apriori algorithm, a
predefined interestingness measurements called
support and confidence are identified to limit the
number of rules generated [9]. The generated rule
could be written as: If {X} Then {Y}.
The If part of the rule (the {X} above) is known as the
antecedent and the THEN part of the rule is known as
the consequent (the {Y} above). The antecedent is the
condition and the consequent is the result. The support
is defined as the number of transactions that include
items in the {X} and {Y} parts of the rule as a percentage
of the total number of transactions N. It is a measure of
how frequently the collection of items occur together as
a percentage of all transactions. Support formula is
shown in Equation (1)
  󰇛󰇜

Confidence is defined as is the ratio of the number of
transactions that include all items in {Y} as well as the
number of transactions that include all items in {X} to
the number of transactions that include all items in {X}.
Confidence formula is shown in Equation (2)
  󰇛󰇜
󰇛󰇜
The two measurements support and confidence reveal
the degree of interestingness for the generated rules. The
user should define a minimum threshold that meets
his/her requirements for every set of generated rules.
The generated rules should assure minimum support and
minimum confidence that is greater than the user
predefined thresholds. Figure 2 shows stage 2
framework.
Figure 2. Stage 2 framework.
Figure 3. Enrollments number prediction using boosted regression trees (2019 -2022).
ADSM
dataset
Data
Description &
Cleaning
Apply Algorithm
Boosted
Regression
Model
Boosted Trees
Output
Perform
Evaluation
10
folds
CV
Apriori
Algorith
m
ADSM
dataset
Data Description
& Cleaning
Apply Algorithm
Generate Rules
Output
(1)
(2)

Predicting Student Enrolments and Attrition Patterns in Higher Educational... 565
Table 2 shows the minimum thresholds defined for
each ADSM programs, the measurements vary
according to the number of the students enrolled in each
program.
Table 2. Minimum association rules thresholds defined for each
ADSM program in stage 2.
Program
Minimum
Support
Minimum
Confidence
Rule Generated
MBA (Master of Business
Administration)
0.5
0.95
20
MSBA (Master of Science of
Business Analytics)
0.05
0.90
10
MSLOD (Master of Science in
Leadership and Organizational
Development)
0.4
0.90
15
MSQBE (Master of Science in
Quality and Business Excellence)
0.3
0.95
10
Table 3 shows the sample of the association rules that
are revealed from our dataset after specifying the
previous threshold for the student’s characteristics in
MBA Program.
Table 3. Sample of rules generated using apriori algorithm for the
student's characteristics in the MBA Program.
LHS
RHS
Support
Confidence
{Status=Graduated,
workplaceSector=Government}
{MBA}
0.6708861
0.9814815
{UGCountry=United Arab Emirates,
workPlace=OilGasSector}
{MBA}
0.5063291
0.9756098
{UGSchool=UAE
UNIVERSITY,workplace=Education}
{MBA}
0.5000000
0.9753086
{gender=male, Nationality=United Arab
Emirates, position=Admin, age = [30-35] }
{MBA}
0.6582278
0.9904762
Figure 4 is the sample of the association rules that are
revealed from our dataset after specifying the previous
threshold for MBA.
Figure 4. Sample of rules generated using apriori algorithm for the
students in the master of Business Administration Program (MBA)
program.
 Stage 3: For the aim of detecting the students who are
under the risk of attrition, association rule Apriori
algorithm is applied to the given dataset with the
following minimum threshold defined in Table 4.
The measurements vary according to the number of
the students drop out from each program in the past
five years. Stage 3 framework is shown in Figure 5
Table 4. Minimum association rules thresholds defined for each
ADSM program in stage 3 for attrition rate detection.
Program
Minimum
Support
Minimum
Confidence
Rule Generated
MBA (Master of Business
Administration)
0.01
0.8
28
MSBA (Master of Science of
Business Analytics)
0.05
0.90
10
MSLOD (Master of Science in
Leadership and Organizational
Development)
0.4
0.90
15
MSQBE (Master of Science in
Quality and Business
Excellence)
0.3
0.95
10
Figure 5. Stage 3 framework.
Table 5 shows the sample of the association rules that
are revealed from our dataset after specifying the
previous thresholds for the student’s attrition in MBA
program.
Table 5. Sample of rules generated using apriori algorithm for the
students attrition rate in MBA program.
LHS
RHS
Support
Confidence
{Age=30-40,Experience.Years=<
2,UGCountry=United Arab
Emirates,UGGPA=2.5-
3,ELR.Score=4-
5,UGcollege=College of Business
and Economics}
{AttritionStatus=
Attrition}
0.012
1.0
{UGGPA=2.5-3,ELR.Score=4-
5,UGSchool=UAE
UNIVERSITY,UGcollege=College
of Business and Economics}
{AttritionStatus=
Attrition}
0.012
1.0
{Programs=MBA,Age=30-
40,gender=Male,UGGPA=2.5-
3,ELR.Score=4-5,position=NOT
EMPLOYEE}
AttritionStatus=
Attrition}
0.014
0.818
3.3. Results Comparison
Given the fact that the admission cycle for year 2019
and 2020 is over. In this section the boosted regression
model accuracy is computed by comparing the value of
predicted and the actual enrolments that is done over
2019 and 2020 academic year. The boosted regression
model that is developed for the aim of predicting the
number of enrolments students for the academic years
(2019 to 2022) and explained in section 3.1 stage 1.
ADSM
dataset
Data
Description &
Cleaning
Apply Algorithm
Generate Rules
for Attrition
Rate
Output

566 The International Arab Journal of Information Technology, Vol. 18, No. 4, July 2021
It is found out that the model results in 86% accuracy
in 2019, and 78% accuracy in 2020. The drop in the
model accuracy in 2020 may occurred because of
unexpected corona virus incident that effects the
student’s enrolment pattern.
4. Conclusions
The models we presented in this paper have broad
significance for the higher educational institutions.
They attempt to answer two important questions facing
almost every institution of higher education, these are:
What type of student the academic institution should
attract and maintain to keep their academic standards?”
and “What type of students are most likely to drop out
during their academic journey?”
In the context of our work, we managed to answer
these two questions using a set of machine learning
approaches. The work described in this paper was held
at ADSM and is aimed to identify and prioritize students
who are at risk of attrition. Although the work in this
paper is aimed to predicting students who are likely to
enrol at specific academic program, we believe that this
solution (problem formulation, the feature extraction
process, data pre-processing, predictive analysis, and
evaluation) applies and generalizes to other academic
outcomes as well, such as predicting student behaviour.
We differentiate our work in this paper from the
above other researches in enrolment and attrition rate
prediction by predicting the number of enrolment of
students in the coming three years and also we managed
to generalize interesting patterns for the student who is
interested in enrolling in a specific program. Besides,
the size of our dataset is significantly larger (1600) than
those used in previous studies, and the model
performance is significantly improved by using the
ensemble approach.
References
[1] Agrawal R., Mannila H., Srikant R., Toivonen H.,
and Verkamo A., “Fast Discovery of Association
Rules,” Advances in Knowledge Discovery and
Data Mining, vol. 12, no. 1, pp. 307-328, 1996.
[2] Al-Shabandar R., Hussain A., Liatsis P., and
Keight R., “Detecting At-Risk Students with Early
Interventions Using Machine Learning
Techniques,” IEEE Access, vol. 7, pp. 149464-
149478, 2019.
[3] Freund Y., Schapire R., and Abe N., “A Short
Introduction to Boosting,” Journal-Japanese
Society for Artificial Intelligence, vol. 14, pp.
1612, 1999.
[4] Kemper L., Vorhoff G., and Wigger B.,
“Predicting Student Dropout: A Machine
Learning Approach,” European Journal of Higher
Education, vol. 10, no. 1, pp. 28-47, 2020.
[5] Mulugeta M. and Borena B., “Higher Education
Students’ Enrolment Forecasting System Using
Data Mining Application in Ethiopia” HiLCoE
Journal of Computer Science and
Technology, vol. 2, no. 2, pp. 37-43, 2013.
[6] Nakhkob B. and Khademi M., “Predicted Increase
Enrolment In Higher Education Using Neural
Networks and Data Mining Techniques,” Journal
of Advances in Computer Research, vol. 7, no. 4,
pp. 125-140, 2016.
[7] Petkovski A., Stojkoska B., Trivodaliev K., and
Kalajdziski S., “Analysis of Churn Prediction: A
Case Study on Telecommunication services in
Macedonia,” in Proceedings of 24th
Telecommunications Forum, Belgrade, pp. 1-4,
2016.
[8] Rai S. and Jain A., “Students' Dropout Risk
Assessment in Undergraduate Courses of ICT at
Residential University-A Case Study,”
International Journal of Computer
Applications, vol. 84, no. 14, pp. 31-36, 2013.
[9] Sadatrasoul S., Gholamian M., and Shahanaghi
K., “Combination of Feature Selection and
Optimized Fuzzy Apriori Rules: The Case of
Credit Scoring,” The International Arab Journal
of Information Technology, vol. 12, no. 2, pp. 138-
145, 2015.
[10] Schapire R. and Freund Y., “Boosting:
Foundations and Algorithms,” Kybernetes, 2013.
[11] Waters A. and Miikkulainen R., “Grade: Machine
Learning Support for Graduate Admissions,” AI
Magazine, vol. 35, no. 1, pp. 64-75, 2014.
[12] Yukselturk E., Ozekes S., and Türel Y.,
“Predicting Dropout Student: An Application of
Data Mining Methods in an Online Education
Program,” European Journal of Open, Distance
and e-learning, vol. 17, no. 1, pp.118-133, 2014.

Predicting Student Enrolments and Attrition Patterns in Higher Educational... 567
Samar Shilbayeh is Assistant
Professor in Business Analytics
program in Abu Dhabi School of
Management, she worked as a senior
data scientist and head of research
Centre in Cognitro Analytics
company experienced in extracting
data, analyzing, findings and applying predictive
modeling techniques, assisted in the development of an
analytics framework for smart construction of a new
type of predictive indictors “Cognitive Indicators”.
Developed active learning algorithms. Developed a cost
effective, expert intelligent system, which guides the
data miner to optimize models selection. Her research
interests include machine learning and AI approaches,
cost sensitive machine learning algorithms, applying
machine learning solutions in health, finance, telecom
and insurance. Dr Samar has a PhD in Machine learning
and AI from the university of Salford, Computer science
and Engineering department, Manchester, UK.
Abdullah Abonamah is the
President and Provost of the Abu
Dhabi School of Management. From
August, 2000 to October, 2007, he
was a Professor and Director of the
Institute for Technological
Innovation at Zayed University and
the Assistant Dean of the College of
Information Systems. Before coming to the UAE, he
was a Professor of Computer Science and the Computer
Science Division Head at the University of Akron. Dr.
Abonamah has a PhD in Computer Science from the
Illinois Institute of Technology, Chicago, Illinois and an
Executive Management Graduate Degree from Yale
University School of Management. His research and
teaching interests include strategy, technology
management, and entrepreneurship and innovation.
With over fifty publications in international journals and
conferences and a US patent in reliable systems, Dr.
Abdullah remains a strong advocate of strategic
management, innovation, entrepreneurship, and proper
technology-business integration.