Exploiting LSTM Neural Network Algorithm Potentiality for Early
Identification of Delayed Graduation in Higher Education
THEODOROS ANAGNOSTOPOULOS1, DIMITRIS PAPAKYRIAKOPOULOS1,
YANNIS PSAROMILIGKOS1, SYMEON RETALIS2
Department of Business Administration,
University of West Attica,
250 Thivon & P. Ralli Str, Egaleo, 12241 Athens,
GREECE
2Department of Digital Systems,
University of Piraeus,
150 Androutsou Str, 18532 Piraeus,
GREECE
Abstract: - Adoption of deep learning classification algorithms in the domain area of higher education provides
exploratory predictive data analytics able to exploit students’ academic behavior. Concretely, student retention
and success are critical concerns in higher education globally. Timely identification of potential delays in
graduation is essential for universities to provide effective interventions and support, ensuring students’
progress efficiently and maintaining high graduation rates, thereby enhancing institutional reputation. This
study examines data from a typical computer science department of a central Greek university, covering student
performance for almost two decades (1999-2018). Through extended data preprocessing, we developed a robust
dataset focusing on key courses indicative of students' likelihood to graduate on time or experience delays. We
employed a deep learning Long Short-Term Memory (LSTM) Neural Network algorithm, leveraging this
dataset to classify and predict students' final academic outcomes. Our findings reveal that early-semester
performance data can successfully forecast graduation timelines, enabling proactive educational strategies to
support student success during their studies at the university.
Key-Words: - Deep Learning LSTM Neural Network Algorithm, Data Preprocessing, Predictive Data
Analytics, Binary Classification, Evaluation Method and Metrics, Early Identification, Delayed
Graduation in Higher Education.
Received: March 7, 2024. Revised: September 3, 2024. Accepted: October 6, 2024. Published: November 8, 2024.
1 Introduction
The complex phenomenon of a timely or a delayed
graduation as well as university dropout has
garnered significant attention within higher
education research communities due to its profound
impact on students' lives and societal development,
[1]. Timely graduation refers to the ability of
students to complete their degrees within an
expected duration of study, while delayed
graduation refers to the prolonged time taken by
students to complete their academic programs
beyond the expected duration. Concretely, dropout
denotes the premature withdrawal of students from
their educational endeavors before achieving their
intended qualifications, [2]. These phenomena not
only impact individual students but also have
broader implications for society, affecting
workforce readiness, economic productivity, and
social mobility.
According to the United States of America
(USA) National Center for Education Statistics, [3],
[4], only a small percentage of students 64% who
began seeking a bachelor’s degree at a 4-year
institution in the fall 2014 completed that degree at
the same institution within 6 years which is
comparable to rates seen in other countries
worldwide, [5]. Subsequently, spanning time of
study in higher educational institutes reveals that in
the case of a 6-year studying process, the graduation
rate was higher for females than for males (67% vs.
60%, respectively).
Concretely, in the European Union (EU), [6],
[7], a study composed of 14 countries out of the 35
bachelor completion rates reported a range from
53% to 83%. However, these numbers should be
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
524
Volume 21, 2024
interpreted with caution due to the different
evaluation methods, which were used to calculate
bachelor completion rates in European higher
education institutes. Intuitively, the unique national
context of each European country needs to be taken
into consideration when comparing educational
indicators, which are applied in countries with
different education cultures.
In Greece, according to data from the Hellenic
Authority for Higher Education, [8], only one in five
students (19.62%) of Greek universities graduate on
time, i.e. within the minimum study time of their
school. It takes an average of six years for a student
to complete a four-year course of study, 7.5 years
for five-year courses, and nine years for medical
studies, which are six-year courses. In fact, the
phenomenon of late or no graduation has worsened
greatly in the last decade after the economic crisis.
Observed data indicate that graduates have been
declining continuously since then from 10.04% of
all enrolled students in 2013 (9.91% in 2015, 9.41%
in 2017, 8.91% in 2019, and 8.60% in 2020).
Compared to the EU, the corresponding average for
EU countries in 2020 was 23.15%, up from 22.13%
in 2018 and 22.28% in 2019.
In this paper, we conducted a research study in
the Department of Digital Systems at the University
of Piraeus, Greece to predict a timely or delayed
student’s graduation. Data sources incorporated
include detailed information on students’ efforts
during the fall and the spring semesters in an
extended total period of time containing two
decades of university studies. The examined data
cover educational experiences over a wide range of
periods from 1999 to 2018. Extensive data pre-
processing was performed on the initial data sources
to produce a detailed derived dataset containing
meaningful information.
Intuitively, the resulting derived dataset
comprises specific courses that describe students’
educational behavior from their first semesters of
study. This educational behavior provides insights
into students’ likelihood of achieving a timely or
delayed graduation. To assess the potential of
predictive analytics, a Long Short-Term Memory
(LSTM) Neural Network deep learning algorithm is
employed to perform classification and prediction
based on the derived dataset. Concretely, the
adopted LSTM Neural Network algorithm
successfully classifies students’ final learning
performance and efficiently predicts whether a
student will graduate on time or experience delays.
These results are primarily based on data from the
first semesters of study in the examined university.
The rest of the paper is organized as follows. In
Section 2 it is presented the prior work in the
research effort area. Section 3 defines the adopted
data model. In Section 4 evaluation parameters are
defined. In Section 5 experiments are performed and
results are observed. Section 6 discusses the
strengths and the weaknesses of the proposed
research effort, while Section 7 concludes the paper
and proposes future work.
2 Prior Work
Delayed graduation and overeducation are studied in
research efforts, [9], where the focus is given to the
social impact of delayed graduation and
overeducation on salaries observed after graduation.
Research on delayed time to get a degree with
regards to post-graduation earnings, [10], correlates
the graduation process along with capital spent by
students’ families to achieve graduation. Degree
completion prediction is analyzed, [11], where
special focus is given to a timely and a delayed
graduation while examining dropout possibilities
within a certain time horizon. An ensemble
prediction model is elaborated, [12], to define
students’ timely and/or delayed graduation assessing
the potentiality of students’ learning behavior
enhanced with university domain knowledge. Data
analytics uplift modeling, [13], is used to model a
timely or delayed graduation while simultaneously
preventing a possible student dropout from the
university.
Machine learning explainable modeling
technology is incorporated, [14], to predict student-
delayed graduation as well as prevent student
dropout. Students’ educational behavior is modeled
as a timestamped academic trajectory, [15], which is
able to predict proactively a timely and/or delayed
graduation during university studies. Learning
analytics can be provided, [16], to understand
students’ academic behavior thus explaining
delayed graduation by assessing the potentiality of a
randomized control experiment. A data mining
research approach is incorporated, [17], which
exploits the ensemble classification algorithm’s
capabilities to proactively predict students’ dropout
and delayed graduation behavior. An evaluation
method is proposed, [18], which aims to explain the
academic dropout rate with regard to labor market
conditions and salaries gained from dropout
university students.
The university graduation process is linked with
the salary earned by graduated students, [19], to
understand the conditions leading to a timely or
delayed graduation behavior in academia. Social
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
525
Volume 21, 2024
network parameters enhance the capabilities of a
pattern-augmented algorithm, [20], which is able to
accurately measure timely and/or delayed student
graduation process as well as a possible dropout
from universities. Delayed graduation and dropout
behavior are examined, [21], with regard to success
or failure rates in core and direction courses during
contemporary studies in higher education. The
impact of financial support on students from their
families, [22], is analyzed to specify the reasons for
timely and/or delayed graduation, which affect their
observed wages in the labor marketplace. Predicting
accurately the timely graduation of students, [23], is
possible by exploiting the impact of crowdfunding
and institutional accompaniment on the continuous
educational process within the university.
Understanding the reasons for delayed
graduation and university dropout resulted in the
creation of degree roadmaps, [24], which used to
provide supportive educational services for timely
graduation. The discovery of potential reasons
resulting in delayed graduation and dropout is
examined, [25], to propose countermeasures, which
aim to eliminate in large scale such inefficient
academic behavior. Social networking analysis from
data belonging to student university networks, [26],
is a promising method to predict proactively
observed delayed graduation and dropout in
universities. Timely graduation is able to be
predicted accurately based on input information
provided by academia [27], thus enabling the
manipulation and measurement of student activities
within the universities.
Research efforts examined in the literature face
timely and/or delayed graduation by using data
sources provided by the universities. Such data are
generated by core or discipline courses offered by
higher education institutes. Several machine
learning algorithms are incorporated to perform
predictive analytics to identify students’ further
activities during their studies. Subsequently, there is
also examined the relation between a timely or
delayed graduation as well as students’ dropout with
regards to the family financial support provided as
well as salaries observed by students after their
graduation in the open marketplace?
Concretely, data preprocessing in the majority
of the research efforts is rather straightforward thus
not exploiting the rich information inherent within
the provided data sources. Intuitively, in most of the
cases, data stem from a limited time horizon of
studies in higher education, which deteriorates
incorporated algorithms’ prediction accuracy since
the more the stochastic data provided to inference
models the more the prediction accuracy observed.
Subsequently, in the observed research efforts focus
is given mainly on predicting a timely or a delayed
graduation for the total number of semesters of a
university, thus underestimating the potentiality of a
specific department’s domain knowledge, which can
provide accurate prediction results from the first
semesters of study.
Subsequently, such inefficiencies are faced by
current research studies, which are able to
understand students’ behavior and predict
proactively a timely or delayed graduation based on
specific domain knowledge observed by the first
semesters of study in the university. Provided data
sources cover a stochastic time horizon of almost
twenty years of students’ activity. This piece of
information is exploited by assessing the potentiality
of an LSTM Neural Network deep learning
algorithm, which is used to perform classification
and prediction based on a meaningful derived data
source of students’ behavior in the Department of
Digital Systems at the University of Piraeus,
Greece.
3 Data Model
Data provided for analysis by the university
department needs an extensive data preprocessing
phase. Such data cleaning and transformation is
necessary to derive a data source meaningful for
performing valuable data analytics by the adopted
LSTM Neural Network deep learning algorithm.
3.1 Provided Data Source
The provided data source has a time span of almost
twenty years of undergraduate student activity in the
Department of Digital Systems at the University of
Piraeus from 1999 to 2018. The initial number of
students examined was 1278. The personal data of
students are preserved by the department, which has
anonymized their sensitive data (i.e., name and age)
with certain identifiers provided by incorporating a
hash function. The provided data source has
information for both core and discipline courses,
while the number of courses is 42. For each type of
course, the dataset includes three discrete data
quantities: (1) the number of times a student has
attempted the course [1-8], (2) the highest grade
achieved [0-10], and (3) the number of years the
course has been taken [1-12]. Total number of
predictive attributes for the provided data source is
calculated by multiplying the number of core and
discipline courses with observed attribute quantities
stored for each course, which results in 126
predictive attributes. The data source has a class
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
526
Volume 21, 2024
attribute, which denotes if a student has timely or
delayed graduation. Specifically, Class 1 denotes
students with delayed graduation, while Class 2
denotes students with timely graduation. It is
observed that for the provided data source delayed
graduation students (i.e., Class 1) are 386, while the
number of timely graduation students (i.e., Class 2)
is 892.
3.2 Derived Data Source
Since the time span of studies is wide it is observed
that undergraduate courses have changed through
time either their names or their context. To avoid
overlaps focus is given to core courses, thus
omitting discipline courses. A number of core
courses is 15 in the undergraduate program of
studies spanning from the first to the last semester
of the undergraduate program. Feature selection is
applied to the observed data courses and it is found
that only a subset of them was significant for further
experimentation, that is only 5 core courses, [28].
Concretely, such courses span only in the first four
semesters of study, that is a timely or a delayed
graduation could be predicted from the first two
years of study thus enabling proactive recovery of
students that will face some problems in finishing
their studies. Specifically, these courses are: (1)
Mathematical Analysis & elements of Linear
Algebra taught in the first semester, (2) advanced
mathematical Analysis taught in the second
semester, (3) discrete mathematics taught in the
second semester, (4) introduction to
telecommunications taught in the third semester,
and (5) algorithms & complexity taught in the fourth
semester of studies. Intuitively, it is derived that
only the maximum number of grades information
observed in the interval is significant for
performing data analytics with the adopted LSTM
Neural Network. Subsequently, grades aggregation
leads to better predictive results, thus initial grades
were transformed according tothe following rules:
(1) grade equals to for grade interval between
, (2) grade equals to for grade interval
between , (3) grade equals to 2 for grade
interval between , and (4) grade equals to 3
for grade interval between . After data
preprocessing it is observed that the final number of
students prepared for experimentation is 674.
Intuitively, derived delayed graduation students
(i.e., Class 1) are 257, while number of timely
graduation students (i.e., Class 2) is 417.
Concretely, derived data source class instances
are visualized as presented in Figure 1. Intuitively,
Class 1 instances data distribution (i.e., delayed
graduation students) is depicted in the bottom left
corner with blue ‘x’ marks. Subsequently, Class 2
instances data distribution (i.e., timely graduation
students) is depicted in the upper right corner of the
plot with red ‘x’ marks.
Fig. 1: Class instances of derived data source
4 Evaluation Parameters
Assessing the performance of the adopted LSTM
Neural Network deep learning algorithm, certain
valuation methods and evaluation metrics should be
incorporated to perform specific experiments and
observe derived results.
4.1 Evaluation Method
To evaluate the adopted LSTM Neural Network
deep learning algorithm there are used certain
evaluation methods. Authors adopt one of the
widely used evaluation methods, due to its
simplicity and optimum results, which is 10-fold
cross-validation. Specifically, such an evaluation
method divides the input dataset into 10 equal-sized
parts and then in a certain loop incorporates the first
9 parts to train the LSTM Neural Network
classification algorithm and the remaining 1 to test
the classifier. This process is repeated until all the
parts are used for training and testing.
4.2 Evaluation Metrics
Given the evaluation method, which is proposed to
support the experimental setup there is a need to
adopt specific evaluation metrics. Such metrics are:
(1) prediction accuracy, (2) correctly classified
instances, and (3) confusion matrix that are able to
assess the efficiency of a deep learning
classification algorithm, such as the adopted LSTM
Neural Network algorithm.
4.2.1 Prediction Accuracy
The effectiveness of the adopted LSTM Neural
Network algorithm is assessed by incorporating
prediction accuracy evaluation metric, ,
which is defined in the following mathematical
equation, (1):
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
527
Volume 21, 2024
(1)
Where,
, are the instances, which are
classified correct as positives, and,
, are the
instances, which are classified correct as negatives.
In addition, , are the instances, which are
classified false are positives, and, , are the
instances, which are classified false as negatives. A
low value of means a weak classifier while a high
value of a indicates an efficient deep learning
classifier.
4.2.2 Correctly Classified Instances
In deep learning literacy, such as in the adopted
LSTM Neural Network algorithm, it is common to
express prediction accuracy as a percentage thus
observed results being more easily interpreted and
presented. Concretely, it is used the term correctly
classified instances, , which is
defined according to the following mathematical
equation, (2):
(2)
Where, a value close to means that the
classification algorithm is not efficient, while a
value close to indicates that the deep learning
algorithm is able to classify instances optimally.
4.2.3 Confusion Matrix
We also evaluated the adopted deep learning LSTM
Neural Network classification algorithm with the
confusion matrix evaluation metric. A confusion
matrix is a special form of data matrix, which in the
case of a binary classification of 2 classes, (i.e.,
Class 1: denoting the categorical class data value of
delayed graduation students, and Class 2: denoting
the categorical class data value of timely graduation
students) has the following encoded form, as
described in Table 1.
Table 1. Confusion matrix evaluation metric
Where, “A” quantity depicts the number of
Class 1 instances, which are classified correctly as
instances of Class 1. “B” quantity depicts the
number of Class 1 instances, which are falsely
classified as instances of Class 2. “C” quantity
depicts the number of Class 2 instances, which are
falsely classified as instances of Class 1. “D”
quantity depicts the number of Class 2 instances,
which are correctly classified as instances of Class
2. It holds that “A”, “B”, “C”, and “D” are denoting
certain numerical data values, which are observed
during the evaluation process.
A given classification model, such as the
adopted deep learning LSTM Neural Network
algorithm, is considered efficient if it maximizes the
numerical data elements of the main diagonal of the
confusion matrix (i.e., “A”, and “D” should have
high numerical data values) and also minimizes the
other numeric data elements of the matrix (i.e., “B”,
and “C” should have low numerical data values).
5 Experiments and Results
Experiments are performed by the adopted LSTM
Neural Network deep learning algorithm, which is
evaluated based on certain evaluation method and
metrics. Results observed have an impact on the
potentiality of the selected deep learning model for
the stated problem of the current research effort.
Table 2. LSTM Neural Network tuning parameters
5.1 Experimental Setup
An LSTM Neural Network deep learning algorithm
is incorporated to perform predictive analytics based
on the derived data source. Since there exist two
classes, namely, Class 1 of delayed graduation
students and Class 2 of timely graduation students
this is a binary class classification problem. The
adopted inference model is provided by Weka
machine learning software, [29]. However, to
observe optimal results there is a need to fine-tune
the provided algorithm according to certain
parameters’ values. Specifically, the adopted model
requires one input layer of 1 node and one hidden
layer. The hidden layer is composed of 2 nodes,
while the output layer is also composed of 2 nodes
(i.e., one node for each of the two classes).
Intuitively, the hidden layer activation function is
ReLu, while the output layer activation function is
SoftMax. Tuning model parameters along with their
values are described in Table 2.
Class 1
Class 2
Classified as
A
B
Class 1
C
D
Class 2
Algorithm tuning parameter
Value
Input layer
1 node
Number of hidden layers
1
Hidden layer
2 nodes
Output layers
2 nodes
Hidden layer activation function
ReLu
Output layer activation function
SoftMax
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
528
Volume 21, 2024
5.2 Derived Results
A fine-tuned LSTM Neural Network learning
algorithm is then used for experimentation, which
provides certain results.
5.2.1 Observed Prediction Accuracy
The evaluation method incorporated to evaluate the
adopted LSTM Neural Network binary class
classification algorithm is 10-fold cross-validation.
According to this evaluation method observed
prediction accuracy is: , which is a
high value for prediction accuracy thus proving that
the adopted machine learning algorithm is suitable
for the examined binary class classification
problem.
5.2.2 Observed Correctly Classified Instances
According to the evaluation method of 10-fold
cross-validation correctly classified instances it
occurred to be: , which indicated that
the selected LSTM Neural Network algorithm is an
optimal choice for the examined classification
problem.
5.2.3 Observed Confusion Matrix
Confusion matrix results are derived based on the
10-fold validation evaluation method for the
examined binary class classification problem.
Derived results are presented in Table 3.
Table 3. Confusion matrix of observed results
It can be observed that most of the classified
instances are located in the main diagonal of Table
3. Specifically, the quantity of elements in the main
diagonal depicts the significant number of certain
instances, which are correctly classified. Concretely,
such an optimal prediction behavior indicates a
robust deep learning LSTM Neural Network
classification algorithm for the examined binary
class classification problem.
Intuitively, the adopted LSTM Neural Network
algorithm has certain classification errors, which
affects its prediction accuracy. Such classification
errors can be visualized as presented in Figure 2.
Subsequently, it can be observed that Class 1 error
instances (i.e., depicted with blue ‘squares’) have
been plotted within the Class 2 correct instances
area (i.e., depicted with red ‘x’ marks) in the upper
right corner of the visualized classification results.
In the bottom left corner, it can be also observed the
Class 1 correct instances area (i.e., depicted with
blue ‘x’ marks).
Fig. 2: Visualization of classification errors
6 Discussion
The current research effort has specific strengths
and certain weaknesses, which are analyzed to
assess qualitatively the potentiality of the study.
6.1 Strengths of the Study
The provided data source contains information on
almost twenty years of students’ activity at the
Department of Digital Systems at the University of
Piraeus, Greece. Concretely, a data preprocessing
phase is required to mine rich context, which in turn
can be used to further perform predictive analytics.
During this phase, there were separated cores from
discipline courses. Research focuses on core courses
that are fundamental for determining a timely or a
delayed graduation. Intuitively, courses of only the
first four semesters of studies are capable of
providing proactive predictions on students’
graduation. Subsequently, an LSTM Neural
Network deep learning algorithm is incorporated to
predict sufficiently a timely or a delayed graduation
of certain students contained in the derived data
source
6.2 Weaknesses of the Study
Discontinuity of undergraduate programs’ courses
during the time period span of almost twenty years
affects the quality of the derived data source.
Specifically, rich information is lost during the data
preprocessing phase, which leads to the deletion of
more precise information. However, such
inefficiencies are part of the provided data source
and are inherent to the initial data source, thus it
could not be feasible to exploit further their
potentiality. In addition, although prediction
accuracy is high and Class 2 has been successfully
predicted there is not the same for Class 1
prediction. This is explained due to the fact that
Class 1 students either have decided to follow a
Class 1
Class 2
Classified as
110
147
Class 1
0
417
Class 2
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
529
Volume 21, 2024
delayed graduation behavior or are in the process of
actively performing their studies at a more relaxed
pace, which affects the observed prediction
accuracy.
7 Conclusions and Future Work
Defining a timely or a delayed graduation of the
students population is of significant importance for
a university since such knowledge has an impact on
academic ratings worldwide. Concretely, proactive
prediction of students’ behavior can affect the
countermeasures incorporated by the university to
help students with their activities to have an
interesting and valuable daily activity. Intuitively,
timely graduation can positively affect the amount
of salary gained on the open market, while delayed
graduation might lead to low wages.
In this research effort, we exploit the
potentiality of students’ data from the Department
of Digital Systems at the University of the Piraeus,
Greece to proactively predict a timely or a delayed
graduation based mainly on core courses of the first
four semesters of study. Future work will be
towards the direction of being able to proactively
infer students’ academic behavior based on more
detailed information existing in the undergraduate
programs such as the discipline courses thus adding
more knowledge to the adopted LSTM Neural
Network deep learning algorithm, which is
incorporated in this research study.
Acknowledgement:
The authors would like to thank the Department of
Digital Systems at the University of Piraeus, Greece
for providing the initial data on students’
performance during their course of study in the
university, which are further exploited by the
current research effort.
Declaration of Generative AI and AI-assisted
Technologies in the Writing Process
During the preparation of this work the authors used
classification and visualization services of the Weka
AI-enabled open source software in order to analyse
provided data sources and exploit their visual
potentiality. After using the services of the AI-
enabled Weka tool, the authors reviewed and edited
the content as needed and take full responsibility for
the content of the publication.
References:
[1] What Matters to Students Success: A Review
of the Literature, National Postsecondary
Education Cooperative (NPEC), July 2006,
[Online].
https://nces.ed.gov/npec/pdf/kuh_team_report.
pdf (Accessed Date: February 12, 2024).
[2] B. Castleman, and K. Meyer, “Financial
Constraints & Collegiate Student Learning: A
Behavioral Economics Perspective”,
Daedalus, vol. 148, no. 4, 2019, pp. 195–216,
DOI: https://doi.org/10.1162/daed_a_01767.
[3] Report of the Condition of Education 2022,
National Center for Education Statistics
(NCES), [Online].
https://nces.ed.gov/pubs2022/2022144.pdf
(Accessed Date: February 12, 2024).
[4] Impact of the Coronavirus Pandemic on Fall
Plans for Postsecondary Education, National
Center of Education Statistics (NCES),
[Online].
https://nces.ed.gov/programs/coe/indicator/tpb
/covid-impact-postsecondary-plans (Accessed
Date: February 12, 2024).
[5] T. Pascarella, and P. Terenzin, “How College
Affects Students, A Third decade of
Research”, Journal of Student Affairs in
Africa, vol. 2, no. 2, 2014, pp. 47–50, DOI:
https://doi.org/10.14426/jsaa.v2i2.70.
[6] Dropout and completion in higher education
in Europe, Publications Office of the
European Union (POEU), [Online].
https://op.europa.eu/en/publication-detail/-
/publication/4deeefb5-0dcd-11e6-ba9a-
01aa75ed71a1/language-en (Accessed Date:
February 12, 2024).
[7] Delayed Graduation and University Dropout:
A review of Theoretical Approaches, Institute
of Labor Economics (IZA), [Online].
https://www.iza.org/publications/dp/12601/del
ayed-graduation-and-university-dropout-a-
review-of-theoretical-approaches (Accessed
Date: February 12, 2024).
[8] Quality Assurance, Hellenic Authority for
Higher Education (HAHE), [Online].
https://www.ethaae.gr/en/quality-assurance
(Accessed Date: February 12, 2024).
[9] C. Aina, and F. Pastore, “Delayed Graduation
and Overeducation in Italy: A Test of the
Human Capital Model Versus the Screening
Hypothesis”, Social Indicators Research, vol.
152, 2020, pp. 533–553, DOI:
https://doi.org/10.1007/s11205-020-02446-0.
[10] D. Witteveen, and P. Attewell, “Delayed
Time-to-Degree and Post-college Earnings”,
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
530
Volume 21, 2024
Research in Higher Education, vol. 62, 2021,
pp. 230–257, DOI:
https://doi.org/10.1007/s11162-019-09582-8.
[11] E. Demeter, M. Dorodchi, E. A. Hossami, A.
Benedict, L. S. Walker, and J. Smail,
“Predicting first-time-in-college students’
degree completion outcomes”, Higher
Education, vol. 84, 2022, pp. 589–609, DOI:
https://doi.org/10.1007/s10734-021-00790-9.
[12] A. A. Priyambada, T. Usagawa, and M. Er,
“Two-layer ensemble prediction of students’
performance using learning behavior and
domain knowledge”, Computers and
Education: Artificial Intelligence, vol. 5,
2023, pp. 1–12, DOI:
https://doi.org/10.1016/j.caeai.2023.100149.
[13] D. Olaya, J. Vasquez, S. Maldonado, J.
Miranda, and W. Verbeke, “Uplift Modeling
for preventing student dropout in higher
Education”, Decision Support Systems, vol.
134, 2020, pp. 1–11, DOI:
https://doi.org/10.1016/j.dss.2020.113320.
[14] J. G. C. Kruger, A. S. Britto, and J. P.
Barddal, “An explainable machine learning
approach for student dropout prediction”,
Expert Systems with Applications, vol. 233,
2023, pp. 1–9, DOI:
https://doi.org/10.1016/j.eswa.2023.120933.
[15] M. F. Musso, C. F. R. Hernandez, and E. C.
Cascallar, “Predicting key educational
outcomes in academic trajectories: a machine-
learning approach”, Higher Education, vol.
80, 2020, pp. 875–894, DOI: 10.1007/s10734-
020-00520-7.
[16] J. Hellings, and C. Haelermans, “The effect of
providing learning analytics on student
behavior and performance in programming: a
randomized controlled experiment”, Higher
Education, vol. 83, 2022, pp. 1–18, DOI:
https://doi.org/10.1007/s10734-020-00560-z.
[17] K. Okoye, J. T. Nganji, J. Escamilla, and S.
Hosseini, “Machine learning model (RG-
DMML) and ensemble algorithm for
prediction of students’ retention and
graduation in education”, Computers and
Education: Artificial Intelligence, vol. 6,
2024, pp. 1–13, DOI:
https://doi.org/10.1016/j.caeai.2024.100205.
[18] D. Mikola, and M. D. Webb, “Finish it and it
is free: An evaluation of college graduation
subsidies”, Economics of Education Review,
vol. 93, 2023, pp. 1–18, DOI:
https://doi.org/10.1016/j.econedurev.2023.102
355.
[19] L. Finamor, “Labor market conditions and
college graduation: Evidence from Brazil”,
Economics of Education Review, vol. 94,
2023, pp. 1–13, DOI:
https://doi.org/10.1016/j.econedurev.2023.102
403.
[20] B. S. Assis, E. Ogasawara, R. Barbastefano,
and D, Carvalho, “Frequent pattern mining
augmented by social network parameters for
measuring graduation and dropout time
factors: A case study on a production
engineering course”, Socio-Economic
Planning Sciences, vol. 81, 2022, pp. 1–12,
DOI:
https://doi.org/10.1016/j.seps.2021.101200.
[21] C. S. Sargent, T. Sullivan, and H. M. Alum,
“DFW in gateway courses not always a
graduation problem: A study in Intermediate
Accounting I from 2007 to 2018”, Journal of
Accounting Education, vol. 60, 2022, pp. 1–7,
DOI:
https://doi.org/10.1016/j.jaccedu.2022.100795
[22] V. Rattini, “The effects of financial aid on
graduation and labor market outcomes: New
evidence from matched education-labor data”,
Economics of Education, vol. 96, 2023, pp. 1–
15, DOI:
https://doi.org/10.1016/j.econedurev.2023.102
444.
[23] C. A. Silvera, J. C. Nino, J. D. Manotas, N. S.
Quintero, and Z. H. Fontalvo, “Model based
in Crowdfunding and institutional
accompaniment for the timely graduation in
undergraduate students”, Procedia Computer
Science, vol. 220, 2023, pp. 991 – 997, DOI:
https://doi.org/10.1016/j.procs.2023.03.137.
[24] X. Su, M. Chen, J. Y. Austin, and Y. Liu,
“Restructuring degree roadmaps to improve
timely graduation in higher education”,
International Journal of Educational
Management, vol. 34, no. 2, 2020, pp. 432-
449, DOI: 10.1108/IJEM-07-2019-0257.
[25] V. Bocsi, T. Cegledi, Z. Kocsis, K. E. Kovacs,
K. Kovacs, A. Muller, K. Pallay, B. E. Szabo,
F. Szigeti, and D. A. Toth, “The discovery of
the possible reasons for delayed graduation
and dropout in the light of a qualitative
research study”, Journal of Adult Learning,
Knowledge and Innovation, vol. 3, no. 1,
2019, pp. 27–38, DOI:
https://doi.org/10.1556/2059.02.2018.08.
[26] N. Nur, N. Park, M. Dorodchi, W. Dou, M. J.
Mahzoon, X. Niu, and M. L. Maher, “Student
Network Analysis: A Novel Way to Predict
Delayed Graduation in Higher Education”,
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
531
Volume 21, 2024
Lecture Notes in Computer Science (LNAI),
vol. 11625, 2019, pp. 370–382, DOI:
https://doi.org/10.1007/978-3-030-23204-
7_31.
[27] N. L. Wade, “Measuring, Manipulating, and
Predicting Student Success: A 10-Year
Assessment of Carnegie RI Doctoral
Universities Between 2004 and 2013”,
Journal of College Student Retention:
Research, Theory & Practice, vol. 21, no. 1,
2019, pp. 119–141, DOI:
https://doi.org/10.1177/1521025119831456.
[28] M. Hall, E. Frank, G. Holmes, B. Pfahringer,
P. Reutemann, and I. H. Witten, “The WEKA
Data Mining Software: An Update”, ACM
SIGKDD Explorations Newsletter, vol. 11, no.
1, 2009, DOI:
https://doi.org/10.1145/1656274.1656278.
[29] I. H. Witten, E. Frank, M. A. Hall, and C. J.
Pal, Data Mining Practical Machine Learning
Tools and Techniques”, Morgan Kaufmann,
Fourth Edition, 2017, DOI:
https://doi.org/10.1016/C2015-0-02071-8.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This research work was supported in part by the
University of Piraeus Teaching and Learning Center
(TLC), Greece under Grant C.947.
Conflict of Interest
The authors have no conflicts of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2024.21.48
Theodoros Anagnostopoulos,
Dimitris Papakyriakopoulos,
Yannis Psaromiligkos, Symeon Retalis
E-ISSN: 2224-3402
532
Volume 21, 2024
