See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/362628426

Towards a Students’ Dropout Prediction Model in Higher Education

Institutions Using Machine Learning Algorithms

Article  in  International Journal of Emerging Technologies in Learning (iJET) · September 2022

DOI: 10.3991/ijet.v17i18.25567

CITATIONS READS

0 18

3 authors, including:

Khalid Oqaidi Sarah Aouhassi

Ecole Normale Supérieure de l'Enseignement Technique Université Hassan II de Casablanca
3 PUBLICATIONS   0 CITATIONS    9 PUBLICATIONS   13 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Information System Quality View project

Towards an intelligent system to enhance higher education quality using data science techniques View project

All content following this page was uploaded by Khalid Oqaidi on 21 September 2022.

The user has requested enhancement of the downloaded file.

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

Towards a Students’ Dropout Prediction Model

in Higher Education Institutions Using Machine
Learning Algorithms
https://doi.org/10.3991/ijet.v17i18.25567

Khalid Oqaidi1(*), Sarah Aouhassi2, Khalifa Mansouri1

1
Laboratory SSDIA, ENSET of Mohammedia, University Hassan II of Casablanca,
Mohammedia, Morocco
2
Laboratory SSDIA, ENSAD, University Hassan II of Casablanca, Mohammedia, Morocco
[email protected]

Abstract—Using machine learning to predict students’ dropout in higher

education institutions and programs has proven to be effective in many use cases.
In an approach based on machine learning algorithms to detect students at risk of
dropout, there are three main factors: the choice of features likely to influence a
partial or total stop of the student, the choice of the algorithm to implement a pre-
diction model, and the choice of the evaluation metrics to monitor and assess the
credibility of the results. This paper aims to provide a diagnosis of machine learn-
ing techniques used to detect students’ dropout in higher education programs, a
critical analysis of the limitations of the models proposed in the literature, as well
as the major contribution of this article is to present recommendations that may
resolve the lack of global model that can be generalized in all the higher educa-
tion institutions at least in the same country or in the same university.

Keywords—student retention, dropout prediction, higher education, machine

learning, classification

1 Introduction

Through a bibliographical study we underline the essential elements in students’

dropout prediction study. Then we highlight some issues in the literature such as:

• The lack of a global model that brings together all the possible variables that concern
the students subject of the study;
• The lack of an ease of choice of one machine learning algorithm over another accord-
ing to the objective of the study: in fact, the algorithms used in different use cases
differ according to the variables and the data available;
• Lack of clear criteria of the choice of the evaluation metric to evaluate the machine
learning algorithm chosen.

This article is organized in sections. In section 2 we present the literature review and
comment the results extracted. In section 3 we discuss and highlight the missing points

iJET ‒ Vol. 17, No. 18, 2022 103

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

in the literature and we present our recommendations to complement them in section 4.

We conclude with section 5.
In our methodology we selected 29 recent articles, most of them dated between
2018 and 2020. In order to be located within the subject, we started with 2 articles on
bibliographic studies of works on the dropout and performance prediction of students
in higher education institutions. Then 1 article on the multi-criteria aspect of the choice
of the algorithm, input variables and metrics for evaluating machine learning models
in different contexts. The following 25 articles are use cases of implementing machine
learning models to predict dropout in the context of higher education, presented in
chronological order of publication. The 29th article is a reference of the confusion
matrix that we present in Table 2. The choice of the articles on dropout and performance
is based on the fact that all the articles must focused on higher education programs
which are degree oriented. This has been achieved by avoiding all studies that are inter-
ested in business-oriented online courses that are not geared towards issuing a univer-
sity degree, and all studies on primary or secondary education. The articles were found
using the key words: students’ dropout, higher education, machine learning, prediction
and students’ performance in databases as Elsevier, IEEE Xplore, ACM, Springer and
other journals.

2 Literature review

2.1 Related works

In this survey the authors analyze different contributions of students’ dropout pre-
diction in India between 2009 and 2016, and try to localize the missing elements that
make the gaps between the previous studies. They stressed four kinds of studies in
Educational Data Mining: Classification, Clustering, Prediction and Association Rule
mining. The machine learning classifiers found in the literature are varied, we note
the most used: Support Vector Machine, Decision Tree algorithms, Artificial Neural
Networks, Logistic Regression, Naïve Bayes, Random Forest and others. The data vari-
ables used to implement the models are diversified, we note some of them: grade in
high school, secondary school and other related education, Gender, Family structure,
Parents Qualification, Parents Occupation, Required for Household work, Addictions
(Alcohol, Smoke, Pills, Solvents, Drugs etc.), Basic facility in the education institution
different for boys and girls, Poor Teaching methodology adopted, Got married [1].
1681 identified papers 67 selected ones to write a systematic literature according to
the Organization for Economic Cooperation and Development (OECD), the European
dropout rates ranged between 30% and 50%, while in the USA the student dropout
rate was 37%. They identified the techniques used for data pre-processing, the factors
affecting dropout, the techniques used for factor selection, the techniques used for pre-
diction, their levels of reliability and the tools used [2].
The selection and evaluation of the machine learning algorithm between the large
choices of possibilities consume lot of time when it is done manually, and usually give
not the results wanted. A classifier must be chosen regarding not only the accuracy, but
also time complexity and consistency [3].

104 http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

Based on the assumption that a single algorithm may fail to detect accurately the
dropout of e-learning courses, the authors have used three machine learning algorithms
for the same purpose. The three methods have been combined and evaluated through
accuracy, sensitivity and precision. The results gave more satisfaction than each of the
algorithms alone [4].
Models that classify students at risk of abandoning a degree were implemented on
27 university degrees using logistic regression. The goal was to locate the variables
responsible of dropout, the most important among them are: start age, parents’ studies,
academic performance, success, average mark in the degree and others. The authors
have noticed that at the same university, two different degrees may have different rea-
sons of dropout [5].
A comparative study to know the optimal Decision Tree to predict students’ perfor-
mance was made. This study is on only just one machine learning algorithm and its
goal was the comparison between these variants of decision tree using metrics such as
accuracy and computational time. The results state that J48 tree was the most suitable
algorithm [6].
A predictive model in e-learning context was developed using three machine learn-
ing algorithms: Artificial Neural Network, Decision Tree and Bayesian Networks,
with two categories of features: students’ personal characteristics and academic per-
formance. The metrics used to evaluate the models are accuracy, precision, recall and
F-measure. Decision tree gave the best results in this case and the students at risk of
dropout have been addressed to the department of the university concerned to take the
necessary measures [7].
A personalized course recommendations based on machines learning techniques ori-
ented system was proposed. The system makes the prediction of students’ dropout as
well as the of the students course grades. The data used has the particularity of contain-
ing only the subject marks obtained in the current year and the entry grade score to the
current institution [8].
An Early Detection System (EDS) was developed in order to detect students at risk of
dropping out by predicting their performance in an early stage. The algorithms Regres-
sion Analysis, Neural Networks, Decision Trees and the AdaBoost are used to identify
the factors responsible for a probable dropout in German universities. Two categories
of data were extracted to train the model: demographic, historical performance and cur-
rent academic performance. The accuracy was higher at the end of the fourth semester
than the first semester, and the AdaBoost performed better than the other algorithms [9].
Two machine learning algorithms have been implemented, Decision Tree and ran-
dom forest, to predict students’ dropout between first and second years in an institute of
technology. The model focused on academic and institutional features only [10].
The authors collected data via questions survey on different factor classes like Aca-
demics, Demographical factors, Psychological factors, Health issues and Behavioral
factors. They implemented the model using Naïve Bayes classifier as only algorithm
and evaluated it with metrics such as Recall and Precision [11].
Unlike most studies, the authors considered 37 non-academic factors for the dropout
prediction. These factors are grouped in 5 categories: demographical, social interaction,
finance, motivation and personal. The data analysis revealed two influencing factors,

iJET ‒ Vol. 17, No. 18, 2022 105

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

number of family members and the relationship with the lecturer, using Decision Tree
algorithm with time and accuracy better than SVM and KNN algorithms [12].
Four machine learning algorithms were compared with 10 variables randomly sam-
pled. The results show that Random Forest did well comparing to Neural Networks,
Support Vector Machines and Logistic Regression in terms of the correctly predicted
dropouts and sensitivity. It has been proved that the prediction is better when the
students’ current data are used to train the model than with previous data [13].
The authors used the classification trees techniques (CART and QUEST) to predict
students at risk of dropout. The predictor variables were the students’ socio-demographic
and academic data. Three models have been developed, the induction-week model,
the first 6–7 weeks’ model and the end first semester model. These models are named
according to the time to extract the data. The results revealed that the predictor factors
are only academic performance while studying in the program and not previous perfor-
mance before entering the university [14].
Gradient Boosted Trees and Deep Learning evaluated with (CV) Cross validation
and (AUC) Area Under Curve metrics gave the best results of under-graduated students
dropout prediction. Special attention was given to the previous high school perfor-
mance in feature selection and feature extraction steps to implement the models [15].
The authors focused on early detection of students at risk of dropout in order to make
an intervention. The class labels to predict by the machine learning algorithms are Pass,
Fail, Conditional Fail, Repeat the Year and Repeat a Single Semester. The experiment
is based on 32 features at the start point, but more features in other experiments were
added progressively in the weeks after. The goal in this study was the detection of at
risk students as soon as possible more than the search for the best accuracy; however, in
less than 4 weeks it gave an accuracy of 97%. This model is ongoing and its data must
be modified through time. It might not be applicable in other institutions or contexts as
it is, but the idea of editable model that provide new results at any moment through the
year still applicable [16].
The authors classified the student representations into three: Global Features-Based,
Local Features-Based and Time Series with the appropriate learning algorithm for each
of them. The experimental implementation showed that the Local Feature-Based was
the best approach to predict the dropout, and the more the model is complex the more
the computational costs increase [17].
The authors predict students’ dropout in the case of programming courses using
Online Judjes with Decision Tree algorithm. They used specific features related espe-
cially to the student’s behavior and performance while trying to answer and submit
programming codes such as number of student logins between the beginning and of a
session [18].
Data of about 28000 students collected from 8 universities in Chile was subject of a
study to identify high-risk students of leaving their studies. Two models are deployed
through the use of Logistic Regression, one with data collected at the beginning of
the semester; the second included the GPA of the first semester. The models were fit
separately to each university. The low engagement and performance in the first weeks
in online courses, attendance sheets and students’ forum activity are signs that requires
urgent intervention [19].

106 http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

A case study of deep learning approach for predicting students’ dropout at the
University of Roma has been realized. The study concern 6000 students enrolled
between 2009and 2014 and used Conventional Neural Network (CNN). The features
were divided into three lists: the first and second one contains the administrative attri-
butes of the students, and the third one includes the students’ career attributes. Three
models were implemented using CNN algorithm, and also Bayesian Networks just to
compare and understand CNN. F1-measure was chosen as the adequate metric indica-
tor into other evaluation metrics such as: Accuracy, Precision and Recall. The perfor-
mance of the prediction has been improved from semester to another from about 67%
for the first year’s students to more than 94% for the third year students. That’s why it
is recommended to develop a real time and permanent predictive model [20].
The authors noticed that the dropout prediction in online higher education is basically
a sequence labeling or time series prediction problem. They extracted seven behavioral
features considered relevant the literature and added five others like Resource, Forum,
and Subpage. They used the classifiers LR, SVM, RF, and DT and different evalu-
ation metrics. The results classify LR as the best in prediction in terms of AUC and
Accuracy [21].
A comparison between Logistic Regression and Decision Tree algorithms students’
dropout has been done in order to predict university students’ dropout. Both classifiers
yield high accuracies with an advantage to decision tree. The data was collected from a
higher education institute in Germany (KIT), and concentrate on different factors like
performance in examinations, financial situation, motivation, health, and family issues.
After the first semester the accuracy has reached 83%, after the third semester it has
reached 95% [22].
In order be able to self-adjust results in an e-learning degree oriented program, the
authors proposed a rule-based classification techniques JRip (JR), PART (PA), OneR
(OR) and Ridor (RI) to stress the factors leading to the dropout of the students at dif-
ferent moments [23].
Three data mining classification techniques are used to provide university students
dropout prediction. The algorithms are Machine Learning Based Decision Stump,
NDTREE algorithms and Enhanced Machine Learning algorithms (EMLA). The data
composed from 407 instances with 5 features and gave the following accuracies in
order 78.3%, 70% and 30.7% [24].
The authors implemented an academic performance predictive model in private
higher education institutions. They compared the results of the algorithms: Artificial
Neural Networks (ANN), eXtreme GBoost, Linear Regression, Support Vector
Machine, Naïve Bayes, and Random Forest. The models were fit with data composed of
10 features including marks, financial situation, study hours per week and English pro-
ficiency. The accuracy shows that ANN performed better than the five other algorithms
in terms of classification of the students in two categories: those who will probably pass
and those predicted to fail [25].
58% of the dropouts noticed have never failed a course, so there are other pre-
dictors. To identify them the authors employed three machine learning algorithms:
Logistic Regression, Neural Networks and Decision Tree. They classify the dropouts
without clear academic reasons as the existential dropouts, and analyzed the data of
45752 students between 2003 and 2015. The variables contain socio-economic status,

iJET ‒ Vol. 17, No. 18, 2022 107

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

demographic, high school performance and other non-academic and non-academic

predictors. The models have successfully predicted the graduates; however, the exis-
tential dropouts still a riddle that even non-academic pre-university performance pre-
dictors failed to foretell [26].
Data from 11 schools of a major university have been used to predict dropout of
the first-year undergraduate students. The machine learning algorithms considered to
choose the best one in terms of performance are Linear Discriminant Analysis (LDA),
Support Vector Machines SVM and Random Forest (RF). The features vary between
demographic previous academic performance and current performance. To evaluate the
algorithms, the authors prioritized accuracy, sensibility and sensitivity than the met-
rics TP (true positive), TN (true negative), FP (false positive) and FN (false negative),
because in this case it is about binary classification [27].
To predict wither a student will pass or fail and to evaluate the effectiveness of
the predictive strategies in the literature, the authors used (ANN) Artificial Neural
Networks and five other algorithms. ANN performed better and the results showed that
the students’ performance6 is correlated with features like group assignments and if the
student is bursary or not [28].

2.2 Comment and syntax presentation

The different works of literature present the different ways of seeing the problem of
students’ dropout in the different institutions of higher education in the world. We can
classify these works according to the algorithms, the type of variables used, or the time
a student spent in the institution.
There are articles that have opted for a single Machine Learning algorithm to imple-
ment a model for predicting students’ abandonment and performance. These kinds of
studies sometimes focus on a single curriculum or different programs or even different
universities to compare the effectiveness of the prediction among them as it is in [5],
[6], [14], [19], [20], [27].
Other works have used several algorithms either to compare between them to choose
the best [9], [12], [13], [26] or to combine them into one [4].
Online higher education or e-learning that is oriented to university degree is also
concerned by several works as in [7], [18], [21], [23].
Some authors have considered that early detection plays an important role in ensur-
ing timely action to avoid dropout as [9], [16].
There are dozens of variables used for training classification models, between those
who have distinguished between local variables that are more specific to the program to
which the student belongs, and global variables that do not take into account program
data [17]. Others have used questions surveys to collect the data needed [11].
In general, we have noted in the literature the most widely used categories, and
we list the variables belonging to each category in Table 1. They are divided into six
classes: (A1) Academic features in the current study program, (A2) previous Academic
features before joining the current program, (SD) Socio-Demographical features,
(In) the features of the Institution and the program; (Bh) Behavioral characteristics and
personality of the students in the program; (Fn) Financial features of the student.

108 http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

Table 1. The features used in machine learning algorithms divided into categories
A1: A2: SD: In: Bh: Fn:
– Access year – Access mark – Marriage – End degree – Section activity – Student job
– Academic – Access form status – Mode – Project – Father’s job
performance – Prior – Sex round submission – Mother’s
rate academic – Birth – Start year date job
– Success rate performance – Family – End year – Number of – Tuition fee
– Average – Educational city – Degree click stream source
mark background – Country – Certified – Play video – Received
– Mode round before – Father’s program event a financial
– Exam round enrollment education – Start – Interact with grant
– Educational – Total failed level course chapter during the
level courses – Mother’s date – Study semester
– English – Total passed education – End course motivation – Scholarship
language courses level date – Suitability of the
literacy – Lowest – Residency – Number expectation to institution
– Multiple previous – Working of unique the majors during the
choice test assessment experience days – Relationship semester
grade results – Political – Study with students – Main
– Project – Highest status program of – Relationship source of
grade previous – Number enrollment with family household
– Lowest assessment of family – School – Relationship income
results results members shift with lecturers
– Highest – Average – Waiting – Inter- – Order
results previous time to Itinerary – Achievement
– Number of assessment study – Pupil – Autonomy
correct tests results – Internet in classroom – Change
– Year of the house ratio – Endurance
admission – Health – Pupil – Aggression
to the – Health teacher – Consistent
university Insurance ratio – Heterosexuality
– Semesters – Home – Years in – Number of
not enrolled language system chapter read
in the past – Family – Strict – Forum
– Courses problems institution Interaction
needed to rules – User Web
graduate – Facility access
– Modules different – Time used to
first time for boys submit answers
and girls – Addictions
(Alcohol,
Smoke, Pills,
Solvents,
Drugs),

We identified the most commonly used machine learning algorithms in the liter-
ature, namely (LR) Logistic Regression, (DT) Decision Tree, (RF) Random Forest,
(ANN) Artificial Neural Networks, (kNN) k-Nearest Neighbors, (SVM) Support
Vector Machine, (NB) Naive Bayes, and (Oth1) the other less used algorithms like
Linear regression, AdaBoost, Enhanced ML Algorithm, Linear Discriminant Analysis
oreXtremeGBoost.

iJET ‒ Vol. 17, No. 18, 2022 109

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

To evaluate the models, the authors used several evaluation metrics, which are the
measures by which we can rate and understand the performance of a machine learning
model. A quick way of visualizing the performance of a model is by a confusion matrix
[29]. For our purposes, the “positive” class is considered the students who have gradu-
ated; and the “negative” class is the non-graduated students.

Table 2. General confusion matrix

Actual Positive True Positive (TP) False Negative (FN) Type I error
Actual Negative False Positive (FP) True Negative (TN)
Type II error
Predicted Positive Predicted Negative

Accuracy (Acr): is the percentage of correctly classified samples. The accuracy of

a machine learning classification algorithm is one way to measure how often the algo-
rithm classifies a data point correctly. Accuracy is the number of correctly predicted
data points out of all the data points. More formally, it is defined as the number of true
positives and true negatives divided by the number of true positives, true negatives,
false positives, and false negatives. A true positive or true negative is a data point that
the algorithm correctly classified as true or false, respectively. A false positive or false
negative, on the other hand, is a data point that the algorithm incorrectly classified.
For example, if the algorithm classified a false data point as true, it would be a false
positive. Often, accuracy is used along with precision and recall, which are other met-
rics that use various ratios of true / false positives / negatives. Together, these metrics
provide a detailed look at how the algorithm is classifying data points.
TP  TN
Accuracy  (1)
TP  TN  FP  FN
Precision and Recall (PR): are two numbers which together are used to evaluate
the performance of classification or information retrieval systems. Precision is defined
as the fraction of relevant instances among all retrieved instances. Recall, sometimes
referred to as ‘sensitivity’, is the fraction of retrieved instances among all relevant
instances. Perfect classifiers have precision and recall both equal to 1.
TP
Recall  (2)
TP  FN
TP
Precision  (3)
TP  FP
It is often possible to calibrate the number of results returned by a model and improve
precision at the expense of recall, or vice versa.
Specificity (Sp): describes the probability of the prediction being false when the
actual class is false.
In simple terms, it describes how specific the model is when predicting negative
instances. It is calculated as the ratio of true negatives to the actual negative cases. It
can be calculated by using the values in the confusion matrix as the ratio of true nega-
tives to the sum of true negatives and false positives.

110 http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

TN
Sensitivity  (4)
TN  FP
F-Score (F-S): also called the F1-Score is a measure of a model’s accuracy on a
dataset. It is used to evaluate binary classification systems, which classify examples
into “positive” or “negative”.
The F-score is a way of combining the precision and recall of the model, and it is
defined as the harmonic mean of the model’s precision and recall.
The F-score is commonly used for evaluating information retrieval systems such as
search engines, and also for many kinds of machine learning models, in particular in
natural language processing.
It is possible to adjust the F-score to give more importance to precision over recall,
or vice versa. Common adjusted F-scores are the F0.5-score and the F2-score, as well
as the standard F1-score.
2*precision*recall
F  Score  (5)
precision  recall
Area Under Curve (AUC): In binary classification problems, the general rule of
thumb is to use a probability threshold of 0.5 to make classification predictions. But for
few scenarios, this threshold might not hold good and using a different threshold would
be more appropriate. A Receiver Operating Characteristic (ROC) curve is the most
commonly used method to visualize the performance of a binary classifier for different
thresholds. It is obtained by plotting the True Positive Rate against the False Positive
Rate. False positive rate is calculated as (1 - Specificity). From the ROC plot, we can
calculate the Area Under the Curve (often referred to as simply the AUC) which is the
probability that a classifier will rank a randomly chosen positive instance higher than a
randomly chosen negative one.
Oth2: other less used evaluation metrics as Cross Validation or time of execution.
We present the essence of the literature review in Table 3 in two parts to facilitate the
readability of the results. The columns indicate the articles in the references at the end
of this article; the rows represent (F) Features, (MLA) Machine Learning Algorithms
and (EM) Evaluation Metrics.
In the features columns we used the number of variables which belongs to each
class category. When the details of the category are not indicated in the literature we
make “X”.

iJET ‒ Vol. 17, No. 18, 2022 111

 Table 3. Factors and tools involved in the prediction of dropout in the literature

112
Articles [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28]
A1 3 6 X 11 X 4 10 17 0 0 X 5 X 8 3 3 8 0 7 10 7 1 2 2 6
A2 1 2 X 3 1 5 0 X 0 7 X X X 4 0 2 0 0 3 6 5 0 1 3 0
SD 4 7 X 9 0 7 0 13 9 5 X 3 X 0 0 7 3 0 2 0 5 1 2 5 0
F
In 0 5 0 1 0 2 2 0 0 7 X 3 X 0 0 2 1 0 0 0 6 2 2 3 1
Bh 2 0 0 1 0 0 0 19 23 0 0 0 0 0 2 1 0 12 1 0 0 1 0 1 1
Fn 0 3 0 1 0 0 0 X 5 0 0 0 0 0 0 1 0 0 2 0 3 0 0 1 2
LR X X X X X X X X
DT X X X X X X X X X X X X X X
RF X X X X X X X X X
ANN X X X X X X X X
MLA
kNN X X X X
SVM X X X X X X X
NB X X X X X X X
Oth1 X X X X X X X X X X
Acr X X X X X X X X X X X X X X X X
PR X X X X X X X X X X X X X X X X X X X X
Sp X X X X X X X
EM
F-S X X X X X X X X X
AUC X X X X X X
Oth2 X X X X
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

3 Discussion

In the solutions proposed in the literature most of the studies are based on academic,
socio-demographic, personal, behavioral financial and institutional data. Some of them
consider that data about higher education degree oriented e-learning must be treated
differently from face to face education. Given a program or an institution, the studies do
not decide for the dropout and performance prediction the choice of the best features,
the best algorithm and the best evaluation metrics that can be generalized for other sim-
ilar programs or institutions. Indeed, in each case study, we find a different algorithm
with different variables and different evaluation metrics.
The problem in the solutions proposed in the literature lies in the fact that there is no
good model for all kind of higher education institutions, even at the same institution the
results differ from one program to another. In each dropout prediction study the authors
start from the beginning and manually test different machine learning algorithms with
different training features. That costs time especially in the collection of data that we
don’t even know if it may have an impact on the dropout. The same university or the
same country does not benefit from previous studies. One cannot generalize a model
on all the other programs. When we talk about one algorithm is better than the oth-
ers, this comparison is not absolute and it depends on the metrics we used to evaluate
the algorithm.
To implement a machine learning algorithm with the goal: prediction students’ drop-
out in higher educational context there are three challenges:
The first is forced by the availability of data, but using all the data we have is costly
in terms of execution time and in terms of the difficulty of taking action after detect-
ing the factors leading to dropout. In one side, the more variables are available in the
outset, the more effective the study will be. In the other side, the more minimalist the
model, the more generalized it can be on many institutions, even if it requires making
some slight modifications.
The second challenge is the machine learning algorithm, it should be chosen in
respect to the university recourses, data type, data size, time of execution authorized
and more criteria.
The third level is the evaluation metrics that should judge the efficiency of the
algorithm. There is no algorithm that is better than another always and in all metrics,
so there is a decision to be done.
These challenges prove the multi-criteria aspect of the issue.

4 Recommendations

We raise the need for a global model (Figure 1) that takes into account the particu-
larities of the study program and the data available in the concerned Higher Education
Institution (HEI), in order to facilitate the choice of the best Machine Learning Algo-
rithm (MLA) for the best available Variables (V) to consider, and the best Evaluation
Metric (EM) to predict dropout as effectively as possible.

iJET ‒ Vol. 17, No. 18, 2022 113

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

Fig. 1. Global model for higher education dropout prediction

To meet these needs, we put in recommendations the following actions:

a) Identification of Higher Education Institutions Classes (HEIC1, HEIC2, HEIC3…):

Categories of institutions that have similarities in the prediction factors of the drop-
out. The objective is to be able to generalize the prediction models as much as
possible in order to save time and effort, and to enable all institutions in the same
category to benefit from a good model already successful in a similar institution.
b) Given the multi-criteria aspect of the problem identified in this work, a model of a
(MCDM) Multi-Criteria Decision Making approach to choosing the three essential
axes responsible for the prediction of the dropout must be the subject of next steps.
c) Development and testing of different Machine Learning Algorithms in a given
higher education establishment in Morocco, then in different institutions and dif-
ferent programs and options, taking into account all the features of the literature to
validate and judge the previous classification and model in a) and b).
d) Once all this work is done, the effort made can be exploited for predicting work
integration. Because after successfully reducing the dropout rate, graduate students
are faced with a new challenge, which is coming out of the unemployment phase.

5 Conclusion

Based on the literature review of the specific topic: dropout prediction in higher edu-
cational programs using Machine Learning techniques. We have classified the works
according to the variables, algorithms and evaluation metrics used. The multi-criteria
aspect of the problem has been proven. Given a higher educational institution, how
to implement a machine learning algorithm to predict the dropout of student knowing
that the larger the number of variables the more time we waste, but the performance of
the algorithm would be better. In the other side if the number of variables is small, the
algorithm can be generalized to other institutions in the same country, in the same city
or at least in the same university.
The result of this work is a clear vision of research tracks that we have cited in the
recommendations and that will serve all researchers who are interested in the quality of
higher education, in the application of artificial intelligence or even to decision-makers

114 http://www.i-jet.org
Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

in the field of higher education, as a diagnostic and guide to getting to the point and
solving one of the major problems of higher education worldwide which is students’
dropout.

6 References
[1] Mukesh Kumar, A.J. Singh, & DishaHanda. (2017). Literature Survey on Educational
Dropout Prediction. International Journal of Education and Management Engineering
(IJEME), 7(2), pp. 8–19. https://doi.org/10.5815/ijeme.2017.02.02
[2] Mayra Alban & David Mauricio. (2019). Predicting University Dropout through Data
Mining: A Systematic Literature. Indian Journal of Science and Technology, 12(4), pp. 1–12.
https://doi.org/10.17485/ijst/2019/v12i4/139729
[3] Rahman Ali, Sungyoung Lee, & Tae Choong Chung. (2016). Accurate Multi-criteria
Decision Making Methodology for Recommending Machine Learning Algorithm. Expert
Systems with Applications. https://doi.org/10.1016/j.eswa.2016.11.034
[4] Ioanna Lykourentzou, Ioannis Giannoukos, Vassilis Nikolopoulos, George Mpardis, &
Vassili Loumos. (2009). Dropout Prediction in E-learning Courses Through the Combina-
tion of Machine Learning Techniques. Computers & Education, 53(3), pp. 950–965. https://
doi.org/10.1016/j.compedu.2009.05.010
[5] Francisco Araque, Concepción Roldán, & Alberto Salguero. (2009). Factors Influenc-
ing University Dropout Rates. Computers & Education, 53(3), pp. 563–574. https://doi.
org/10.1016/j.compedu.2009.03.013
[6] Sulaiman Abdulsalam, Akinbowale Babatunde, Ronke Babatunde, & Moshood Hambali.
(2015). Comparative Analysis of Decision Tree Algorithms for Predicting Undergradu-
ate Students’ Performance in Computer Programming. Journal of Advances in Scientific
Research & Its Application (JASRA), 2, pp. 79–92.
[7] Mingjie Tan & Peiji Shao. (2015). Prediction of Student Dropout in E-Learning Program
Through the Use of Machine Learning Method. International Journal of Emerging Technol-
ogies in Learning (iJET), 10(1), p. 11. https://doi.org/10.3991/ijet.v10i1.4189
[8] Sergi Rovira, Eloi Puertas, & Laura Igual. (2017). Data Driven System to Predict Academic
Grades and Dropout. PLoS ONE, 12(2), p. e0171207. https://doi.org/10.1371/journal.
pone.0171207
[9] Johannes Berens, Simon Oster, Kerstin Schneider, & Julian Burghoff. (2018). Early
Detection of Students at Risk – Predicting Student Dropouts Using Administrative Student
Data and Machine Learning Methods, Schumpeter Discussion Papers, No. 2018-006,
University of Wuppertal, Schumpeter School of Business and Economics, Wuppertal,
https://doi.org/10.2139/ssrn.3275433
[10] K. Limsathitwong, K. Tiwatthanont, & T. Yatsungnoen, “Dropout prediction system to
reduce discontinue study rate of information technology students,” 2018 5th International
Conference on Business and Industrial Research (ICBIR), 2018, pp. 110–114, https://doi.
org/10.1109/ICBIR.2018.8391176
[11] V. Hegde & P. P. Prageeth, “Higher education student dropout prediction and analysis
through educational data mining,” 2018 2nd International Conference on Inventive Systems
and Control (ICISC), 2018, pp. 694–699, https://doi.org/10.1109/ICISC.2018.8398887
[12] T. Dharmawan, H. Ginardi, & A. Munif, “Dropout detection using non-academic data,”
2018 4th International Conference on Science and Technology (ICST), 2018, https://doi.
org/10.1109/ICSTC.2018.8528619

iJET ‒ Vol. 17, No. 18, 2022 115

Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

[13] M. Solis, T. Moreira, R. Gonzalez, T. Fernandez, & M. Hernandez, “Perspectives to pre-

dict dropout in university students with machine learning,” 2018 IEEE International
Work Conference on Bioinspired Intelligence (IWOBI), 2018, https://doi.org/10.1109/
IWOBI.2018.8464191
[14] José María Ortiz-Lozano, Antonio Rua-Vieites, Paloma Bilbao-Calabuig, &
MartíCasadesús-Fa. (2018). University Student Retention: Best Time and Data to Iden-
tify Undergraduate Students at Risk of Dropout. Innovations in Education and Teaching
International, https://doi.org/10.1080/14703297.2018.1502090
[15] M. Nagy & R. Molontay, “Predicting dropout in higher education based on secondary school
performance,” 2018 IEEE 22nd International Conference on Intelligent Engineering Sys-
tems (INES), 2018, pp. 000389–000394, https://doi.org/10.1109/INES.2018.8523888
[16] Cameron C. Gray & Dave Perkins. (2019). Utilizing Early Engagement and Machine Learn-
ing to Predict Student Outcomes. Computers & Education, 131, pp. 22–32. https://doi.
org/10.1016/j.compedu.2018.12.006
[17] Rubén Manrique, Bernardo Pereira Nunes, Olga Marino, Marco Antonio Casanova, & Terhi
Nurmikko-Fuller. (2019). An Analysis of Student Representation, Representative Features
and Classification Algorithms to Predict Degree Dropout. LAK19: Proceedings of the 9th
International Conference on Learning Analytics & Knowledge, pp. 401–410. https://doi.
org/10.1145/3303772.3303800
[18] Pereira F.D. et al. (2019). Early Dropout Prediction for Programming Courses Supported
by Online Judges. In: Isotani S., Millán E., Ogan A., Hastings P., McLaren B., Luckin R.
(eds) Artificial Intelligence in Education. AIED 2019. Lecture Notes in Computer Science,
vol. 11626. Springer, Cham. https://doi.org/10.1007/978-3-030-23207-8_13
[19] Gad Yair, Nir Rotem, & Elad Shustak. (2020). The Riddle of the Existential Dropout: Les-
sons From an Institutional Study of Student Attrition. European Journal of Higher Educa-
tion, 10(4), pp. 436–453. https://doi.org/10.1080/21568235.2020.1718518
[20] Lorenz Kemper, Gerrit Vorhoff, & Berthold U. Wigger. (2020). Predicting Student Dropout:
A Machine Learning Approach. European Journal of Higher Education, 10(1), 28–47.
https://doi.org/10.1080/21568235.2020.1718520
[21] Ahmed A. Mubarak, Han Cao & Weizhen Zhang. (2020). Prediction of Students’ Early
Dropout Based on Their Interaction Logs in Online Learning Environment. Interactive
Learning Environments, https://doi.org/10.1080/10494820.2020.1727529
[22] Paul T. Von Hippela & Alvaro Hofflinger. (2020). The Data Revolution Comes to Higher
Education: Identifying Students at Risk of Dropout in Chile. Journal of Higher Education
Policy and Management, 43(1), pp. 2–23. https://doi.org/10.1080/1360080X.2020.1739800
[23] Clauirton Siebra, Ramon Santos, & Natasha Lino. (2020). A Self-Adjusting Approach for
Temporal Dropout Prediction of E-Learning Students. International Journal of Distance
Education Technologies, 18, pp. 19–33. https://doi.org/10.4018/IJDET.2020040102
[24] Francesco Agrusti, Mauro Mezzini, & Gianmarco Bonavolonta. (2020). Deep Learning
Approach for Predicting University Dropout: A Case Study at Roma Tre University. Journal
of E-Learning and Knowledge Society, 16, pp. 44–54.
[25] Sallan, G. and Behal, S. (2020). Prediction of Student Dropout Using Enhanced Machine
Learning Algorithm. Advances in Mathematics: Scientific Journal, 9(6), pp. 3821–3826.
https://doi.org/10.37418/amsj.9.6.61
[26] Francesca del Bonifro, Maurizio Gabbrielli, Giuseppe Lisanti, & Stefano Zingaro. (2020).
Student Dropout Prediction. Artificial Intelligence in Education, pp. 129–140. https://doi.
org/10.1007/978-3-030-52237-7_11

116 http://www.i-jet.org
 Paper—Towards a Students’ Dropout Prediction Model in Higher Education Institutions Using Machine…

[27] Roderick Lottering, Robert Hans, & Manoj Lall. (2020). A Machine Learning Approach to
Identifying Students at Risk of Dropout: A Case Study. International Journal of Advanced
Computer Science and Applications, 11(10). https://doi.org/10.14569/IJACSA.2020.0111052
[28] Francis Makombe & Manoj Lall. (2020). A Predictive Model for the Determination of
Academic Performance in Private Higher Education Institutions. International Jour-
nal of Advanced Computer Science and Applications, 11(9). https://doi.org/10.14569/
IJACSA.2020.0110949
[29] Ting K.M. (2011). Confusion Matrix. In: Sammut C., Webb G.I. (eds) Encyclopedia of
Machine Learning. Springer, Boston, MA. https://doi.org/10.1007/978-0-387-30164-8_157
[Accessed: December 10, 2021]

7 Authors

Khalid Oqaidi is a PhD student, IT engineer, and founder of a startup in the field
of data science and market research and a company in computer science engineering
(email: [email protected]).
Sarah Aouhassi is a professor engineer at the University Hassan II Casablanca,
ENSAD Institute with a doctorate in Information Systems and a state engineering
diploma from the national institute of statistics and applied economy (INSEA). Having
experience of ten years within Hassan II University in Casablanca as a statistic chief
officer. Her research focuses on the quality of Information System and its interaction
with the quality of higher education. (email: [email protected]).
Khalifa Mansouri is a professor of computer science and researcher at the University
Hassan II Casablanca, ENSET Institute. His research is focused on Real Time Systems,
Information Systems, e-Learning Systems, Industrial Systems (Modeling, Optimization,
and Numerical Computing). PhD (Calculation and optimization of structures) in 1994
to Mohammed V University in Rabat, HDR in 2010 and PhD (Computer Science) in
2016 to Hassan II University in Casablanca. (email: [email protected]).

Article submitted 2021-09-28. Resubmitted 2022-06-22. Final acceptance 2022-07-22. Final version
published as submitted by the authors.

iJET ‒ Vol. 17, No. 18, 2022 117

View publication stats