Revista Brasileira de Informática na Educação – RBIE

Brazilian Journal of Computers in Education

(ISSN online: 2317-6121; print: 1414-5685)
https://sol.sbc.org.br/journals/index.php/rbie

Submission: 19/Jul/2023; 1st round notif.: 02/Jan/2024; New version: 05/Jan/2024; 2nd round notif.: 04/Mar/2024;
Camera ready: 20/May/2024; Edition review: 20/May/2024; Available online: 20/May/2024; Published: 20/May/2024;

Educational Data Mining for Dropout Prediction: Trends,

Opportunities, and Challenges

Miriam Pizzatto Colpo Tiago Thompsen Primo

Programa de Pós-Graduação em Computação Programa de Pós-Graduação em Computação
Universidade Federal de Pelotas (UFPel) Universidade Federal de Pelotas (UFPel)
Diretoria de Tecnologia da Informação ORCID: 0000-0003-3870-097X
Instituto Federal Farroupilha (IFFar) [email protected]
ORCID: 0000-0002-6477-3227
[email protected]

Marilton Sanchotene de Aguiar Cristian Cechinel

Programa de Pós-Graduação em Computação Programa de Pós-Graduação em Computação
Universidade Federal de Pelotas (UFPel) Universidade Federal de Pelotas (UFPel)
ORCID: 0000-0002-5247-6022 Centro de Ciências, Tecnologias e Saúde
[email protected] Universidade Federal de Santa Catarina (UFSC)
ORCID: 0000-0001-6384-409X
[email protected]

Abstract
Today, we face academic, social, and economic losses associated with student dropouts. Several studies have applied
data mining techniques to educational datasets to understand dropout profiles and recognize at-risk students. To
identify the contextual (academic levels, modalities, and systems), technical (tasks, categories of algorithms, and
tools), and data (types, coverage, and volume) characteristics related to these works, we performed a systematic
literature review, considering institutional and academic degree dropout. Internationally recognized repositories were
searched, and the selected articles demonstrated, among other characteristics, a greater exploration of educational,
demographic, and economic data of undergraduate students from classification techniques of decision tree ensembles.
In addition to not having identified any study from underdeveloped countries among the selected ones, we found
shortcomings in the application of predictive models and in making their predictions available to academic managers,
which suggests an underutilization of the efforts and potential of most of these studies in educational practice.
Keywords: Student Dropout; Dropout Prediction; Educational Data Mining; Systematic Literature Review.

Cite as: Colpo, M. P., Thompsen Primo, T., Aguiar, M. S. & Cechinel, C. (2024). Educational Data Mining for
Dropout Prediction: Trends, Opportunities, and Challenges. Revista Brasileira de Informática na Educação,
32, 220-256. https://doi.org/10.5753/rbie.2024.3559.
Colpo et al. RBIE v.32 – 2024

1 Introduction

During the COVID-19 pandemic and the need for social distancing, the formal education, until
then on-site, changed to remote and virtual classes. Although essential to the pandemic, the
emergency remote education posed challenges to students, professors, and institutions, which
needed to be methodologically and structurally prepared for this alternative form of teaching (J. R.
Santos & Zaboroski, 2020). In this scenario of difficulties and changes, student dropout caused an
even more significant concern (Colpo et al., 2021).
The student dropout, characterized by the student’s permanent departure from your degree
of origin without proper completion (Brasil, 1996), represents a severe educational problem that
causes academic, social, and economic damage to the individual who evades and to society. This
fact is because there is a correlation between the education level and the salary gains of the popu-
lation, which directly affects the socio-economic development of a country (Pontili et al., 2018).
In addition, shortages of skilled labor can negatively affect the productive capacity of a nation, and
students who have not completed their degrees are more likely to be unemployed and need welfare
benefits (Lee & Chung, 2019). When considering public education, student dropout also means a
waste of state resources, not only financial but also personnel and infrastructure (Silva Filho et al.,
2007).
Identifying at-risk students, based on knowledge of the previous dropout patterns, can sup-
port educational institutions in the decision-making and implementing institutional policies to
prevent dropout (Nagy & Molontay, 2018). With this purpose, some studies try to draw dropout
profiles from manual analyses based on data collected from interviews or surveys. However, be-
cause of the difficulty of contacting a significant portion of the evaded population, the results are
susceptible to biases (Silva, 2013). Moreover, the large volumes of educational data provided by
learning and academic management systems also make manual explorations unfeasible, although
they store information of high strategic potential (Romero & Ventura, 2020).
In this context, Educational Data Mining (EDM) techniques have been adopted to automate
dropout analysis. EDM is a research area that focuses on developing methods capable of exploring
large volumes of educational data to understand more effectively the students’ behavior and other
factors related to learning (Baker et al., 2011). As a multidisciplinary area, data mining (academic
or not) incorporates techniques from different domains, such as machine learning (Han et al.,
2012). Supervised machine learning, for example, is the basis of classification, which is the
EDM task often used in dropout prediction for automatically discovering patterns and attributes
related to this phenomenon and for automatically identifying at-risk students. In the first scenario,
classification models are developed descriptively to improve the understanding of dropout-related
patterns and help professionals identify at-risk students through pedagogical monitoring (manual
prediction). In the second one, the models are intended to identify potential dropouts (automatic
prediction) (Colpo et al., 2020).
Several secondary studies have already been conducted on applying data mining and ma-
chine learning techniques to educational problems. Specifically considering the context of student
dropout, Mduma et al. (2019a) presented a survey on machine learning techniques used to pre-
dict dropout. Marques et al. (2019) and Rondado de Sousa et al. (2021) carried out systematic
mappings on data mining technologies in identifying the causes of dropout and addressing the

221
Colpo et al. RBIE v.32 – 2024

problem of face-to-face students dropout, respectively. Unlike Mduma et al. (2019a), Marques
et al. (2019) and Rondado de Sousa et al. (2021), in this work, we analyze studies that apply EDM
techniques in the dropout prediction context through a Systematic Literature Review (SLR). In
addition, the scope of this research is broader than that considered by Marques et al. (2019) since
it is not restricted to identifying the causes of dropout, and our RSL includes contextual, technical,
and data aspects not covered by Mduma et al. (2019a) and Marques et al. (2019). SLRs also were
conducted by Colpo et al. (2020), de Oliveira et al. (2021), and Agrusti et al. (2019). The first
study focused on the Brazilian research scenario, while the latter two concentrated on university
dropouts. In contrast, our RSL covers the international research scene without restrictions on the
level or modality of education. However, we have limited our scope to studies that address the
problem of breaking the link with formal education.
In more detail, this SLR analyzes studies that use EDM to predict institutional and degree
dropout. The aim is to identify (i) contextual characteristics, including educational modalities,
levels, and systems; (ii) technical characteristics as tasks, categories of algorithms, and tools; and
(iii) data characteristics considering types, coverage, and volume. We do not view works that deal
with dropout in subjects/courses, as it does not reflect the closure of students’ educational bonds.
The planning of this SLR, including the method and definition of research questions, search
strategies, and the selection and quality criteria, was described in Section 2. At the same time,
we explained the details of its conduct in Section 3. In Section 4, we present the results obtained,
answer previously established research questions, and describe many of the studies analyzed.
Then, in section5, we summarize some trends, opportunities, and challenges that we have observed
among the studies that apply EDM in predicting student dropout, as well as possible threats to the
validity of our research. Last, conclusions and final remarks are presented in Section 6.

2 Methodology and Planning

According to Kitchenham and Charters (2007), developing an SLR is identifying and evaluating a
broad range of research in an area or topic of interest using a suitable and reliable method. In this
paper, we used the SLR method proposed by Kitchenham and Charters (2007) to investigate the
international research scene regarding the use of EDM in predicting student dropout. More specif-
ically, as the primary research question, we sought to examine these studies’ contextual, technical,
and data characteristics, considering dropout at the institutional and degree levels/scopes. To this
end, we believe the following specific research questions in this SLR:

• Q1. What are the objectives of these studies?

• Q2. What educational levels, modalities, and systems were investigated?
• Q3. What is the nature, coverage, and volume of the data used?
• Q4. What tasks, techniques, and tools are being focused on?

This SLR considered works published from 2016 to 2022 in the following scientific bases:
ACM Digital Library (https://dl.acm.org), IEEE Xplore (https://ieeexplore.ieee.org), Web of Sci-
ence (https://www.webofscience.com), and Scopus (https://www.scopus.com). We contemplated
222
Colpo et al. RBIE v.32 – 2024

publications from the last seven years before the execution of the SLR since we consider this
sample to be reasonably representative of the advances and the current research scenario. In addi-
tion, we chose the scientific databases based on their international relevance to Informatics in the
Education community.
The expressions used in the search engines were previously refined through the execution
and evaluation of preliminary queries. They included terms related to the Problem ("drop*out" OR
"dropout"), Population ("student" OR "school" OR "college" OR "university”), and Intervention
("data mining" OR "machine learning"). However, as the repositories use different default search
forms, we have standardized that the query should be performed on the publication’s metadata
(title, abstract, and keywords). We achieved this by creating advanced queries adapted to each
repository’s different options and syntaxes, as shown in Table 1.

Table 1: Advanced search expressions for each scientific database.

Database Advanced Search Expressions

ACM Abstract:(("drop*out" OR "dropout") AND ("student" OR "school" OR "college" OR
"university") AND ("data mining" OR "machine learning")) OR Title:(("drop*out" OR
"dropout") AND ("student" OR "school" OR "college" OR "university") AND ("data min-
ing" OR "machine learning")) OR Keyword:(("drop*out" OR "dropout") AND ("student"
OR "school" OR "college" OR "university") AND ("data mining" OR "machine learning"))
IEEE (("All Metadata": "drop*out" OR "All Metadata": "dropout") AND ("All Metadata": "stu-
dent" OR "All Metadata": "school" OR "All Metadata": "college" OR "All Metadata":
"university") AND ("All Metadata": "data mining" OR "All Metadata": "machine learn-
ing"))
Scopus TITLE-ABS-KEY(("drop*out" OR "dropout") AND ("student" OR "school" OR "college"
OR "university") AND ("data mining" OR "machine learning"))
WoS TS=(("drop*out" OR "dropout") AND ("student" OR "school" OR "college" OR "univer-
sity") AND ("data mining" OR "machine learning"))

For the selection of studies, we established five criteria:

• the article must be written in Portuguese1 or English, following the format of a scientific
paper;

• the study should describe the application of a specific solution, thus excluding literature
reviews;

• the article must adopt EDM techniques in solutions that contribute to the prediction of stu-
dent dropout;

• the work should address dropout in an institutional or academic degree scope, discarding
research aimed at dropout in subjects/courses or short-term training; and
1 Although we take into account the international research scene, we kept the Portuguese language among the

acceptance/selection criteria because it is also in our domain and because we felt that this decision would not signifi-
cantly influence the results, considering that the searches were carried out with keywords in English and international
scientific databases. This judgment was confirmed during the RSL process, as only two of the selected studies (to be
presented in the next section) are written in Portuguese.

223
Colpo et al. RBIE v.32 – 2024

• The article must provide details of its development and results, transposing the proposal
sphere.

Finally, to refine the selection/exclusion process, we also observed three quality criteria:

• the paper should be more than five pages long (whether single or double column) to present
a minimum of detail about its solution;

• in its validations, the research must consider a dataset of at least 500 instances/students to
guarantee a minimum representativeness/quality to the experiments; and

• The data used in the research should not be collected mostly by questionnaires and surveys
due to the higher propensity of bias in this approach.

3 Conducting the RSL

We conducted the RSL in 2023 with the support of the Parsifal tool2 , considering three phases.
Initially, in Step 1, we query the scientific databases3 and export their results’ metadata (including
abstracts) in BibTeX format. Then, we import the BibTeX files into Parsifal to expedite the dupli-
cate check and streamline the article review process. After eliminating exact duplicates, in Step
2, the studies resulting from Step 1 had their abstracts and metadata (such as size, format, and
language) analyzed, and papers that did not meet the previously established selection or quality
criteria got removed from the analysis stage. Note that although the abstract is present, not all
publication metadata includes pages, language, and format information. Consequently, some arti-
cles that may have been excluded at this stage proceeded to the next phase. It is also important to
point out that we did not conduct a peer review in the analysis of the publications due to the large
volume of articles collected. That is, a single researcher analyzed each article. Finally, in Step 3,
we extracted the full texts of the studies resulting from Step 2, when available4 , and analyze them
based on our selection and quality criteria. Furthermore, we only retained the most recent studies
when detecting similar publications by the same authors in our reading. These articles represent
progress in the same research, even though they are not identical duplicates.
Figure 1 shows, by repository, the number of results returned in the searches and the number
of exclusions for each criterion analyzed in each step. Notice that in Step 1, the searches yielded
1275 results, and many of the studies returned by the WoS (99) and Scopus (288) databases were
identified as duplicates and removed. This is because these databases were the last to be searched,
and the other repositories had already retrieved many of their results. It is also possible to observe
that during the analysis of the abstracts in Step 2, many studies were removed because they did
not use EDM and machine learning techniques or because they did not address the problem of
2 https://www.parsif.al/
3 The scientific databases were accessed through the Portal de Periódicos (https://www.periodicos.capes.gov.br) of
the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES), a foundation linked to the Brazilian
Ministry of Education.
4 Notice that the availability of the full text is not among the selection criteria presented in Section 2. This restric-

tion is a standard collection criterion since selection depends on the feasibility of the analysis. The same applies to
the duplicate limitation, as there is no reason to analyze an article that is identical/similar to another.

224
Colpo et al. RBIE v.32 – 2024

institutional or degree dropout. More specifically, many studies dealt with dropouts from cours-
es/subjects, Massive Open Online Courses (MOOCs), or other educational problems such as grade
prediction. Furthermore, especially considering the IEEE database, many articles were entirely
outside the educational domain and related to the use of the dropout technique in the regularization
of neural network models. Finally, in Step 3, the most significant discards were publications whose
files/full texts were not available/accessible. Since the ACM and IEEE Digital Libraries mostly
hold their publications, these occurrences were concentrated in the WoS and Scopus databases.

Figure 1: Progress of the article analysis steps, taking into account the selection and quality criteria specified in the RSL protocol.

Figure 2(a) graphically shows the number of articles resulting from each selection phase,
considering the scientific databases separately. We selected 71 works identified in Tables 2 and 3.
Figure 2(b) shows the distribution of the selected studies by year of publication. It can be seen
that no article published in 2016 was selected and that there has been a progressive increase in
the number of selected papers since 2019. These results suggest a recent growth in the interest
to apply EDM techniques to face the student dropout problem, confirming the potential of this
research topic.
Regarding the origin of publications, Figure 3 shows the mapping of the number of selected
225
Colpo et al. RBIE v.32 – 2024

Table 2: List of selected articles.

# Reference Country(ies) of Affiliations Database

1 Baranyi et al. (2020) Hungary ACM
2 Chen et al. (2018) USA ACM
3 Kang and Wang (2018) USA ACM
4 Shiau (2020) China ACM
5 Vasquez Verdugo et al. (2022) Chile, USA ACM
6 Xu and Wilson (2021) USA ACM
7 Yu et al. (2021) USA ACM
8 Aguirre and Pérez (2020) Ecuador, Spain IEEE
9 Bassetti et al. (2022) Italy IEEE
10 Böttcher et al. (2021) Germany IEEE
11 Costa et al. (2021) Brazil, Chile IEEE
12 da Silva et al. (2019) Brazil IEEE
13 Fernández-García et al. (2021) Spain IEEE
14 Kiss et al. (2019) Hungary IEEE
15 Kurniawati and Maulidevi (2022) Indonesia IEEE
16 Lottering et al. (2020) South Africa IEEE
17 Masood and Begum (2022) India IEEE
18 Mduma and Machuve (2021) Tanzania IEEE
19 Nagy and Molontay (2018) Hungary IEEE
20 Orooji and Chen (2019) USA IEEE
21 Ortigosa et al. (2019) Spain IEEE
22 Pachas et al. (2021) Brazil, Peru IEEE
23 Prada et al. (2020) Italy, Poland, Portugal, Romania, Spain IEEE
24 G. Santos et al. (2020) Brazil IEEE
25 Solis et al. (2018) Costa Rica IEEE
26 Yang et al. (2021) USA IEEE
27 Yoo et al. (2017) USA IEEE
28 Alturki et al. (2022) Germany WoS
29 Barros et al. (2019) Brazil WoS
30 Beaulac and Rosenthal (2019) Canada WoS
31 Berka and Marek (2021) Czech Republic WoS
32 Chung and Lee (2019) South Korea, USA WoS
33 Crespo (2020) United Kingdom WoS
34 de Assis et al. (2022) Brazil WoS
35 Deho et al. (2022) Australia WoS
36 Del Bonifro et al. (2020) Italy, France WoS
37 Demeter et al. (2022) USA WoS
38 Flores et al. (2022) Peru, Spain WoS
39 Fontana et al. (2021) Italy WoS
40 Freitas et al. (2020) Brazil, Saudi Arabia, Canada WoS
41 Hannaford et al. (2021) USA WoS
42 Hoffait and Schyns (2017) Belgium WoS
43 Iam-On and Boongoen (2017a) Thailand WoS
44 Iam-On and Boongoen (2017b) Thailand WoS
45 Karimi-Haghighi et al. (2022) Spain WoS

226
Colpo et al. RBIE v.32 – 2024

Figure 2: Status by database and step (a); and temporal evolution of selected articles (b).

Table 3: List of selected articles (continuation of Table 2).

# Reference Country(ies) of Affiliations Database

46 Kuzilek et al. (2021) Czech Republic, Germany WoS
47 Lee and Chung (2019) USA, South Korea WoS
48 Opazo et al. (2021) Chile WoS
49 Palacios et al. (2021) Chile WoS
50 Perchinunno et al. (2021) Italy WoS
51 Perez et al. (2018) Colombia WoS
52 Queiroga et al. (2022) Brazil, Uruguay WoS
53 Queiroga et al. (2020) Brazil, Chile WoS
54 Segura et al. (2022) Spain, Paraguay WoS
55 Shilbayeh and Abonamah (2021) United Arab Emirates WoS
56 Sorensen (2019) USA WoS
57 Tsai et al. (2020) Taiwan WoS
58 Urbina-Najera and Mendez-Ortega (2022) Mexico WoS
59 Villegas-Ch et al. (2020) Ecuador, Spain WoS
60 Viloria et al. (2019) Colombia WoS
61 Agrusti et al. (2020) Italy Scopus
62 Bitencourt et al. (2022) Brazil Scopus
63 Gamao and Gerardo (2019) Philippines Scopus
64 Hutagaol and Suharjito (2019) Indonesia Scopus
65 Mduma et al. (2019b) Tanzania Scopus
66 Naseem et al. (2022) Fiji Scopus
67 Nuanmeesri et al. (2022) Thailand Scopus
68 Oreshin et al. (2020) Russia Scopus
69 Park and Yoo (2021) South Korea Scopus
70 Rovira et al. (2017) Spain Scopus
71 Vega et al. (2022) Peru Scopus

studies by country, considering the affiliations of the authors of each article. For reference, the
graphic also presents the continent, the Gross National Income (GNI) per capita (gray line), and
the Human Development Index – HDI (blue line)5 for each country. Although the HDI considers
5 TheHDI aims to assess a country’s development based on the capabilities of its population, not just its economic
growth. Therefore, it considers three critical dimensions of human development: income, education, and health.

227
Colpo et al. RBIE v.32 – 2024

multiple dimensions, it is possible to observe an alignment between the two indicators so that
a country with a lower GNI also tends to have a lower HDI. A greater concentration of works
is observed in Europe and South and North America6 , with the United States (USA), Brazil, and
Spain standing out as the countries with the most considerable quantities of publications. Although
most studies have been conducted in developed countries with higher GNI and very high HDI
(equal to or greater than 0.8), the interest in developing countries with high HDI (between 0.7
and 0.8) in this area of research is also notable, especially in South America. However, only two
and one of the selected articles concern countries with medium (HDI below 0.7) and low (HDI
below 0.55) levels of human development, respectively, and with lower GNIs. This fact shows
that the growth of research aimed at applying EDM to the problem of student dropout is not
confirmed in the poorest and underdeveloped countries, precisely those that would most need this
type of initiative. As Section 1 mentions, reducing student dropout is essential for socioeconomic
development. Therefore, there is still a need to publish and disseminate knowledge on this topic
to instigate the development of new studies, especially in countries lacking it.

Figure 3: Distribution of selected articles by country and continent.

4 Results and Discussion

At the end of the selection process, we extracted information from the selected articles to answer
the research questions established in Section 2. The results were synthesized and discussed in
this section, considering the following organization: Subsection 4.1 will present the results from
The GNI per capita and HDI values shown in Figure 3 were taken from the last HDI dataset, available in December
2023 at https://hdr.undp.org/data-center/human-development-index#/indicies/HDI. Notice that the Taiwan/Republic
of China does not have GNI and HDI information in Figure 3, as the United Nations does not recognize it.
6 Notice that the total number of articles from each continent is presented with its respective legend, in parentheses.

228
Colpo et al. RBIE v.32 – 2024

contextual perspectives related to Q1 and Q2. Subsections 4.2 and 4.3 will describe characteristics
of the data and techniques used, answering Q3 and Q4, respectively.

4.1 Context

4.1.1 Objectives

Regarding the first research question, the works that use EDM techniques in the context of student
dropout, in general, aim at (i) the automatic discovery of patterns and attributes related to this
phenomenon to understand its causes better and thus help professionals to predict at-risk students
through pedagogical monitoring (“manual” prediction); (ii) the development of predictive models
that can automatically identify students with dropout patterns (automatic prediction); or (iii) both
of the above objectives, that is, they seek to provide predictive models of dropout, in addition
to investigating patterns or indications of the attributes most strongly associated with dropout.
Figure 4 presents the distribution of the articles selected in this RSL concerning these objectives.
Notice that only six studies (8%) are restricted to investigating the patterns and relationships be-
tween the variables (attributes) analyzed and student dropout. The vast majority, 65 (92%), focus
on developing predictive models, among which a large proportion, 46 (65% of the total number
of articles), also address the investigation of patterns or attributes related to dropout.

Figure 4: Distribution of the selected articles according to their objectives.

Considering the first category, in common, the six studies demonstrate a strong relationship
between student dropout and students’ low academic performance. In more detail, Iam-On and
Boongoen (2017a) used the clustering approach to build descriptive models related to dropout,
considering demographic/social and academic (school and university) information from students
at Mae Fah Luang University (Thailand). Groups were generated to evaluate newly admitted stu-
dents and those who completed the first year of their course/degree separately. From the analysis
of the representative profile of each group, the authors identified that students with high academic
performance in school continue to perform well in university and that students with low academic
performance tend to drop out after the first year.
Yoo et al. (2017) and de Assis et al. (2022) used academic and demographic, and only
academic data in association rule mining to identify patterns related to the dropout of Computer
Science students from a public university in the USA and of Production Engineering students at the
Federal Center for Technological Education of Rio de Janeiro (Brazil), respectively. By mining
sequential patterns, which also involves analyzing the temporality of the data, on the records
of subjects/courses taken, Yoo et al. (2017) found that students tend to drop out after taking –
229
Colpo et al. RBIE v.32 – 2024

usually more than once due to failure – the subjects/courses of introductory programming and
mathematics. de Assis et al. (2022) performed association rules mining on student performance
data augmented by social network metrics, in which the degree of propagation of Grade Point
Average (GPA) was used as a proxy for the existing bond among students. As a result, the authors
identified that lower school performance and minor participation in social networks lead to delayed
graduation and dropout.
Using academic and demographic data from engineering students at six universities in the
European Union, Prada et al. (2020) developed a web-based software tool for tutoring support.
In addition to using dimensionality reduction, clustering, and visualization techniques to enable
exploratory data analysis, the tool builds/uses classification models of academic performance and
dropout with a descriptive objective. That is, to provide global hints about student behavior.
Among the patterns found, the authors pointed out that the age of entry hurts graduation. At the
same time, the admission score and the student’s performance in the first semester have a positive
impact on graduation.
Also using students’ academic and demographic data, Yang et al. (2021) employed ex-
ploratory statistical analysis, correlation, and data mining/classification techniques to investigate
how individual courses and course sequences influence student dropout/graduation on Computer
Science majors at San Francisco State University. About the use of EDM, dropout prediction
models based on the Random Forest (RF), Logistic Regression (LR), and Support Vector Machine
(SVM) classifiers were developed using grade data from courses that approximately correspond
to the completion of 2, 3, 4, and 5 terms, respectively. Among the results, analyzing the impor-
tance rankings of the models, the authors identified that performance in mathematics and physics
courses has a more significant predictive impact at the initial stage of the degree.
Finally, Perchinunno et al. (2021) developed LR and Decision Tree (DT) models from de-
mographic and academic data of students who completed their first year of undergraduate or
single-cycle master’s degrees (combination of bachelor and master) at the University of Bari Aldo
Moro (Italy). The models were used descriptively to analyze the attributes and patterns related to
dropouts between the first and second year at university. Among the results, the authors identified
that the risk of dropping out is higher for male students or those who completed fewer than twelve
credits in their first year.
In general, researchers developing predictive models are more concerned with improving
results and exploring and evaluating different techniques in data pre-processing and generating
classification models. Since assessing the quality of predictive models involves simulating their
performance with varying evaluation metrics on a test dataset, perhaps there is greater clarity
about the need for improvement and, consequently, a more effective search for techniques that
help increase the performance and reliability of these models.
For example, Gamao and Gerardo (2019) and Queiroga et al. (2020) used evolutionary al-
gorithms to optimize their classification results and evaluate different classification algorithms
in developing their predictive models. In Gamao and Gerardo (2019), they proposed a swarm
optimization algorithm to select an optimal subset of attributes and maximize the accuracy of pre-
dictive models. To do so, the authors considered the demographic, academic, and economic data
of first-year students at Davao del Norte State College (Philippines), in addition to Naive Bayes
(NB) and DT classifiers showing better predictive results. Queiroga et al. (2020) proposed the use

230
Colpo et al. RBIE v.32 – 2024

of a genetic algorithm to optimize the hyperparameters of dropout prediction models, considering

the data count of students’ interactions with a Virtual Learning Environment (VLE) in a distance
technical degree at Sul-rio-grandense Federal Institute (Brazil). In the proposed approach, models
based on the DT, LR, RF, Multilayer Perceptron (MLP), and Adaptive Boosting (AdaBoost) al-
gorithms, with different hyperparameter configurations, competed against each other. Thus, they
used the best classifier and its best combination of hyperparameters at the end of the evolutionary
process. Compared to the Grid Search optimization method and the default-configured models,
the authors demonstrated better predictive results of their approach.
Barros et al. (2019) also considered technical degrees, but using academic, demographic,
and economic data from face-to-face students of integrated degrees7 , at Federal Institute of Rio
Grande do Norte (Brazil). In addition to optimizing the hyperparameters and evaluating the DT,
MLP, and Balanced Bagging (BB) classification algorithms, different balancing techniques and
evaluation metrics were tested. As a result, the BB classifier without additional balancing tech-
niques outperformed the other models associated with varying balancing techniques. The authors
also showed that it is unreliable to consider only the target class’s precision, recall, and f1-score
in the context of unbalanced data. It is essential to look at metrics more sensitive to performance
discrepancies between classes, such as Unweighted Average Recall (UAR) and Geometric Mean
(G-mean).
Using demographic and academic data from a public dataset of school students in India,
Masood and Begum (2022) also evaluated different metrics and several resampling techniques for
handling imbalanced data in developing LR and SVM classifiers. The authors found that SVM
models constructed with random undersampling and SMOTE (Synthetic Minority Over-sampling
TEchnique) in pre-processing yielded superior results. Furthermore, Masood and Begum (2022)
pointed to the greater effectiveness of the AUC-ROC (Area Under the Receiver Operating Curve)
for measuring predictive performance in the minority class.
As a last example, Solis et al. (2018) used academic, demographic, and economic data of
undergraduate students at the Instituto Tecnológico de Costa Rica. The authors evaluated the RF,
MLP, LR, and SVM classifiers, considering different perspectives on student data representation.
These perspectives aimed to assess (i) whether the learning of the predictive models would benefit
from including active/enrolled students among the examples of the non-dropout class and (ii)
whether the dropout behavior would be better represented from the entire academic trajectory
(i.e., a data record for each semester attended by the student) or from the last semester of the
student. As a result, in addition to pointing out the better performance of RF, the authors identified
that the best strategy was to use data from all semesters attended and only consider records of
graduated/concluded students as negative examples of dropout.
Among studies that develop predictive models, those that also investigate patterns or at-
tributes most strongly related to dropout usually do so by applying techniques that provide the
importance of each feature or by analyzing the model itself when it is interpretable. Costa et al.
(2021) and Naseem et al. (2022), for example, have built predictive dropout models based on data
from undergraduate Computer Science students at the Federal University of Pelotas (Brazil) and
the University of the South Pacific (Fiji), respectively. While Costa et al. (2021) used academic
and socioeconomic data from the student’s first three semesters in the development of DT, RF, and
7 Brazilian integrated degrees combine secondary and technical education.

231
Colpo et al. RBIE v.32 – 2024

LR models; Naseem et al. (2022) used academic, socioeconomic, and interactional data (from in-
teraction with a VLE) from the student’s first year to build DT, RF, NB, LR, and KNN (K-Nearest
Neighbor) models, considering three different prediction moments (enrollment and end of the first
and second semesters). The importance rankings established by the models and the attribute se-
lection results from the Boruta algorithm were analyzed by the first and second studies to identify
the attributes with a more significant influence on dropout. In both studies, the characteristics
associated with student academic performance were more important. While Costa et al. (2021)
highlighted the better performance of the RF model, Naseem et al. (2022) pointed out that the NB
excelled in predicting at the enrollment stage and the LR in predicting at the end of the first and
second semesters.
Berka and Marek (2021) built different dropout prediction models considering academic
and demographic data from face-to-face or distance bachelor’s degree students at a university in
the Czech Republic. The authors used data available at admission (student’s first enrollment) and
the end of the four initial semesters. Models were generated for each prediction moment and
different attribute combinations using the DT, LR, and RF algorithms. In addition, as dropout
examples, the models considered: (i) only students who the university dismissed for not meeting
any academic rule/requirement; or (ii) also including students who dropped out of the degree on
their own. As a result, the authors identified that the first strategy slightly improves the learning
of the positive dropout class and that the LR classifiers showed better predictive performance.
Furthermore, the authors interpreted the DT models to verify the influence of dropout-related
attributes and patterns. They performed a dependency analysis, using different association rule
algorithms and only data available at admission, to identify differences between students who did
or did not follow the degree after the first semester. Among the results, the percentage of credits
lost in the last semester was identified as the most essential attribute for the prediction models. At
the same time, dependency analysis pointed out that the time between the end of high school and
entry into higher education is the highest risk factor for early dropout.
Although less common because it is a more recent area of research, eXplainable Artificial
Intelligence (XAI) techniques, particularly Shapley Additive explanations (SHAP), also appear
as a resource for interpreting black box models. Baranyi et al. (2020) proposed applying a Fully
Connected Deep Neural Network (FCNN) in dropout prediction. Two other deep neural network
models and two ensembles of DTs were also trained and evaluated using academic and demo-
graphic data from undergraduate student admissions at the Budapest University of Technology
and Economics (Hungary). The authors observed a slightly higher predictive performance of the
FCNN. To better understand the decisions of their model, they evaluated the importance of at-
tributes based on permutations and the SHAP approach. Among the results, Baranyi et al. (2020)
found that the time between high school completion and higher education admission, the general
admission score, and the mathematics score on the entrance exam showed greater predictive im-
portance. Similarly, Bassetti et al. (2022) used SHAP values to measure the impact of each feature
on the predictions of a Gradient Boosting Trees (GBT) model. More specifically, the authors pre-
sented ISIDE, the prototype of a student dropout alert system integrated into the online student
portal of the Sapienza University of Rome. ISIDE works asynchronously, making predictions on
data collected every week and returning a list with the probability of students dropping out. Be-
fore opting for the best performance of the GBT in the predictive task, the authors also evaluated
DT, RF, LR, SVM, and MLP models, all built from demographic, economic, and academic data
from the university’s undergraduate and graduate students. Looking at the SHAP values, Bassetti
232
Colpo et al. RBIE v.32 – 2024

et al. (2022) identified that students with more extended time since last enrollment, fewer credits
earned, or lower grade averages are likelier to drop out.

4.1.2 Education Levels, Modalities and Systems

Also related to context, but answering the second research question, Figures 5(a) and 5(b) show
the number of selected articles8 by educational levels and modalities or systems, respectively. An-
alyzing Figure 5(a), it is possible to observe little exploration at School and Technical education
levels, with the vast majority of works linked to Undergraduate. This quantity is understandable
since the studies are usually developed within universities using their proprietary data. In addition,
higher education institutions use academic management systems and VLEs more widely, gener-
ating greater data availability at this level of education. Regarding the modality, notice that the
selected works deal mainly with face-to-face teaching9 . However, this result may be influenced
by the research interest of this review, which restricts dropout to the institutional and academic
degree scopes, disregarding studies at the level of subjects/courses, which seems to be a trend
in the works focused on distance modality as they explore the availability and constant updating
of interactive student data in VLE. Considering Figure 5(b), it is also evident that most studies
address dropout in public institutions. Most of the works presented in Section 4.1.1 follow these
results, focusing on higher, public, and face-to-face education. However, it is also essential to
describe the few initiatives that consider the other levels, modalities, and educational systems to
exemplify and encourage the expansion of research in these contexts.

Figure 5: Number of selected articles by educational levels and modalities (a) or systems (b), respectively.

Among the eleven studies aimed at school education, Masood and Begum (2022) has already
been presented. Sorensen (2019) and Orooji and Chen (2019) used academic, socioeconomic, and
behavioral data obtained from public school administrative databases in the USA states of North
Carolina and Louisiana, respectively. In the first study, which used data from the third through
eighth grades of elementary school to predict the dropout of students enrolled in high school,
SVM and boosting of DTs models showed, in that order, higher accuracy when compared to sim-
pler prediction models. From the interpretation of DT and LR models and the feature importance
8 Notice that the same article can be counted in more than one chart category according to the used data. This
observation is valid for all graphs presented in this work.
9 Although many works have been categorized as “Undefined” since they do not mention or give evidence of this

information, these studies probably deal with the face-to-face modality, given its predominance in formal education.

233
Colpo et al. RBIE v.32 – 2024

of bagging DTs, the authors also identified that behavioral and academic attributes showed greater
predictive power than demographic ones. Orooji and Chen (2019) evaluated different classifica-
tion algorithms and balancing techniques and found that balancing techniques benefited recall.
In Chung and Lee (2019) and Lee and Chung (2019), the authors used behavioral and aca-
demic data of high school students from South Korea’s National Education Information System.
While in the first work, the authors balanced the data and developed an RF model, in the second
one, a comparative analysis of oversampling was presented, considering the RF classifier and a
boosting of DTs. As a result, in addition to validating the model’s predictive performance, the
authors found in the first study that attributes related to unauthorized absences and delays had
the highest predictive importance, respectively. Furthermore, in the second research, the authors
identified that boosting DTs without applying the balancing technique had the best results, demon-
strating the need to evaluate techniques of interest before including them in a solution.
Mduma and Machuve (2021) andMduma et al. (2019b) used public datasets with academic,
demographic, and economic data from students in Tanzania, Kenya, and Uganda, and only Tan-
zania, respectively, to develop models for predicting school dropout. Both studies employed data
balancing techniques and permutation of feature importance to identify the features that contribute
most to dropout. As a result, Mduma and Machuve (2021) pointed to the better performance of
the LR model when compared to the MLP and RF ones, as well as the greater importance of the
attributes of gender, age, income, parent check child’s books, and meals per day. Although they
don’t present comparative or predictive performance data, Mduma et al. (2019b) signal that they
have recognized an ensemble developed by soft combining tuned LR and MLP models as the
best model. In addition to the attributes identified in Mduma and Machuve (2021), Mduma et al.
(2019b) also pointed out as essential attributes that indicate whether the student reads books with
their parent and whether their parent discusses the child’s progress with the teacher. Mduma et al.
(2019b) selected these essential features and considered them as input for a prototype web-based
system. This system used the ensemble model and was developed to predict whether a student
will drop out of school based on the information entered and to display a visualization of schools
with a high risk of dropping out.
Considering the threat to the usefulness of predictive models posed by concept drifts during
the COVID-19 pandemic, Xu and Wilson (2021) used imputation-based simulations to analyze
the impact of data quality and availability on the performance of the dropout prediction model
implemented in Rhode Island’s Early Warning System (EWS). EWS is RF-based and trained
on academic, demographic, and economic data of Rhode Island’s public high school students,
considering techniques to deal with imbalance. By building and evaluating models based on
different imputation strategies and concept drift simulations, the authors found that in certain
circumstances, some predictive models, while imperfect, can still be helpful in assisting decision-
making.
Crespo (2020) simulated and compared the effectiveness of an income-proxy means test
(PMT) and mechanisms based on dropout prediction in reaching the poor and future school
dropouts for targeting social cash transfer policies. To do this, the author used academic, demo-
graphic, and economic data from Chilean governmental databases, covering students from seventh
grade onwards and present in the Chilean Social Protection File. GBT, SVM, RF, elastic net, lasso,
and generalized additive models were evaluated in the dropout prediction task, with elastic nets
performing best. Based on the feature ranking of the models, Crespo (2020) identified greater

234
Colpo et al. RBIE v.32 – 2024

importance in attributes related to the student’s age, grades, attendance, and scholar grade/year, in
addition to the previous average dropout rate in the school. The author also observed that using
predictive results in conjunction with the PMT increased the targeting effectiveness, except when
the social valuation of the poor and future dropouts is very different.
Queiroga et al. (2022) describe a nationwide learning analytics project in Uruguay that aims
to mitigate school dropout and retention in secondary education. Using academic (covering the
first grade of primary school to the second grade of secondary school) and socioeconomic data
of students, the authors built eight RF models to predict dropout at different times of the first
two grades (before the start of the school year and after the initial assessment meeting for each
grade) of Uruguayan regular and technical basic secondary education. In addition to considering
the use of balancing, attribute selection, and optimization techniques in the construction of the
models, Queiroga et al. (2022) conducted a bias analysis of the models considering three protected
attributes, which resulted in the approval of seven of them. The project has also developed a web
API that allows the results/predictions of the approved models to be used and made available, both
synchronously and asynchronously.
Like Queiroga et al. (2022), Barros et al. (2019) also considers technical and school edu-
cation, taking into account the context of integrated education in the Brazilian system. Thus, in
addition to being linked to the school level, they are also counted at the technical level, along
with the work of Queiroga et al. (2020). Barros et al. (2019) and Queiroga et al. (2020) have
already been presented in Section 4.1.1, as well as those by Perchinunno et al. (2021) and Bassetti
et al. (2022), which addressed dropouts in undergraduate and graduate degrees, the same situation
as Oreshin et al. (2020), Del Bonifro et al. (2020), and Fernández-García et al. (2021).
In Oreshin et al. (2020), the authors used academic, demographic, and economic data, static
(available on admission) and dynamic (added semiannually), from students at the University of
Information Technologies, Mechanics and Optics (Russia). In addition, interaction and senti-
ment analysis data from the students’ publications in a social network were considered among the
dynamic information. Thus, with dropout as the target variable, in addition to the early predic-
tion model, designed to provide predictions immediately after student admission, eight prediction
models were developed, considering data extracted from one to eight consecutive semesters of the
degree and the GBT algorithm (chosen empirically). As a result, the authors highlighted good pre-
dictive results pointing to high school certification type, hometown, and academic center/student’s
area of knowledge as the top three attributes in the early model. However, for the dynamic mod-
els, the characteristics of academic performance in previous semesters showed more significant
predictive potential.
Del Bonifro et al. (2020) used social and academic (school and university) data of under-
graduate and single-cycle master’s students from an unidentified university to build early dropout
prediction models. The authors trained the models by considering only data available at admis-
sion or including data before the end of the first year. Considering random undersampling to deal
with data imbalance and different combinations of hyperparameters, the authors developed mod-
els based on SVM, RF, and Linear Discriminant Analysis (LDA) algorithms for the different sets
of attributes. As a result, the authors chose SVM and RF models for the scenarios with and with-
out first-year-related features, respectively. Moreover, although both models are necessary and
valuable for risk anticipation, as expected, a considerable improvement in predictive performance
was observed when introducing data related to students’ first year.

235
Colpo et al. RBIE v.32 – 2024

Given the difficulty of accessing personal data and privacy issues, Fernández-García et al.
(2021) sought to develop the best possible dropout prediction models based solely on socioeco-
nomic and academic data of undergraduate and master students from an engineering school of a
Spanish public university. To do this, the authors used several feature and instance engineering
techniques in the pre-processing, including balancing/resampling. Considering hyperparameters
optimization, Fernández-García et al. (2021) developed GBT, RF, and SVM classifiers, along with
a heterogeneous ensemble, for five different stages, specifically before enrolment and at the end
of each of the first four semesters. As a differential, except for the models intended for the first
stage, the others consider the prediction result of the immediately preceding stage model among
their attributes. As a result, Fernández-García et al. (2021) pointed to a better performance of the
GBT at the time of enrollment and of the heterogeneous ensemble in the predictions at the end of
the first and second semesters.
On the other hand, Alturki et al. (2022) and Shilbayeh and Abonamah (2021) only con-
sider data from graduate-level students, more specifically master’s degrees. Alturki et al. (2022)
used demographic and academic performance data of master’s students in Business Informatics
at the University of Mannheim (Germany). Considering SMOTE for oversampling, the authors
developed LR, RF, KNN, NB, SVM, and ANN (Artificial Neural Networks) models for the tasks
of predicting student status (degree completion/non-completion) and academic grade (average,
above average, and below average) at the end of the first and second semesters. Among the
results, Alturki et al. (2022) identified better performance of RF models trained after oversam-
pling and that predictions made after the second semester are more accurate. In addition, using
RF permutation importance, the authors identified the semester grades and the distance from the
student’s accommodation to the university as the most important features for predicting the stu-
dent’s achievements. Shilbayeh and Abonamah (2021) used socioeconomic and admission and
previous academic data of master students from Abu Dhabi School of Management (United Arab
Emirates). A boosted regression tree ensemble was developed to predict the number of enrolled
students in subsequent academic years. In addition, the authors applied the Apriori algorithm to
extract association rules and discovery patterns about master’s students and who is most likely to
drop out. As a result, Shilbayeh and Abonamah (2021) highlighted the superior performance of
the ensemble when compared to a single boosted regression tree, and characteristics related to the
30-40 age range and the 2.5-3 undergraduate GPA range showed recurrence in the rules associated
with dropping out.
Queiroga et al. (2020), Berka and Marek (2021), Kang and Wang (2018), Yu et al. (2021),
and Ortigosa et al. (2019) are examples of articles that consider the context of distance learning.
Kang and Wang (2018) and Yu et al. (2021) used academic and demographic data of undergrad-
uate students at USA public universities to develop models for predicting dropouts during winter
break and the first year of the degree, respectively. Kang and Wang (2018) developed logistic pre-
diction models using or not using data random undersampling. As each model favored accuracy
or recall, authors used both to predict the dropout risk of active students, generating two listings
of at-risk students for educational managers. In Yu et al. (2021), the authors produced separate
models for students in distance and face-to-face degrees based on the data available at the end of
the first period/term of the degree. The models were generated from the LR and GBT algorithms,
considering or not the inclusion of sensitive attributes, such as gender and color/race, to assess
whether these data produce discriminatory models. As a result, after evaluating the overall per-
formance and fairness of the predictions (variation of predictive performance on minority groups)
236
Colpo et al. RBIE v.32 – 2024

of each model, the authors identified that the GBT models performed better and that sensitive
attributes had little impact on the algorithmic fairness of the predictions, being preferable to use
them.
In addition to describing the development and evaluation of prediction models, Ortigosa
et al. (2019) presented the challenges and technical changes imposed by implementing an online
dropout risk prediction system for undergraduate students at Universidad a Distancia de Madrid
(UDIMA). The system provides (i) static early predictions, which consider data available on stu-
dent enrollments in their respective academic years/cycles, and (ii) dynamic predictions, which
include data from student interaction in the VLE, collected and updated periodically. These pre-
dictions are based on demographic, economic, academic, and interactional data. By dividing the
academic year into ten monthly periods, the authors developed 22 specialized models (one static
and ten dynamic for each type of student – new or recurrent) based on the C5.0 DT algorithm.
In addition, the authors identified that age and university access type were considered necessary
attributes in the prediction models of new students but ignored by those of veterans, who be-
gan to consider attributes related to student performance in previous years/cycles. Following this
behavior, the greater the data the models consider, the better the results presented.
As well as Ortigosa et al. (2019), another 13 works consider data from the private system.
Perez et al. (2018) and Hannaford et al. (2021), for example, devoted their research to specific un-
dergraduate degrees. They developed predictive models for the dropout of Systems Engineering
students and the risk of not graduating Nursing students, considering private universities in Colom-
bia and the USA, respectively. Perez et al. (2018) developed DT, LR, NB, and RF models utilizing
academic and demographic data of the students. Hannaford et al. (2021), in turn, developed 126
predictive models, considering different combinations between demographic and academic data
(from high school and university). Models were built for five prediction moments, considering
the beginning of additional academic years, based on eight classifiers and one ensemble based
on the weighted average of the results of these classifiers. As a result, Perez et al. (2018) iden-
tified that the RF model showed superior performance, while the ensemble with eight classifiers
scored better in Hannaford et al. (2021). The latter work also pointed out that the more advanced
the prediction moment, i.e., the more academic data is considered, the higher the model’s perfor-
mance. In common, both studies identified that data related to university academic performance
have greater predictive importance. The most essential attributes included cumulative GPA and
GPA in the degree-specific subjects/courses.
Hutagaol and Suharjito (2019) also proposed to combine the results of three prediction mod-
els (KNN, NB, and DT) into an ensemble; however, using a GBT meta-classifier to predict the
dropout of students from the Faculty of Social and Political Sciences of a private Indonesian uni-
versity. Starting from demographic, economic, and academic data, the authors performed data
balancing and attribute selection in pre-processing by considering SMOTE and Learning Vector
Quantization (LVQ) techniques, respectively. As a result, in addition to the ensemble outper-
forming the individual classifiers, LVQ highlighted the greater predictive importance of attributes
related to students’ attendance and academic performance.
da Silva et al. (2019) used demographic, academic, and economic data of students from on-
site degrees at public and private universities provided by the Brazilian Census and Higher Educa-
tion Flow Indicators. The authors built and evaluated three ensembles of homogeneous regressors,
each corresponding to the bagging of linear, robust, or ridge regressors, in addition to reproducing

237
Colpo et al. RBIE v.32 – 2024

a heterogeneous ensemble proposed by another work, with seven base algorithms combined by
a Ridge meta-regression. The authors trained these four regression ensembles from two distinct
subsets of attributes, selected by Stepwise and Pearson correlation methods, respectively. As a
result, the homogeneous ensembles outperformed the heterogeneous ones, with emphasis on the
bagging of linear regressors trained on the attributes selected by the Stepwise method, which had
lower error rates. In addition, the degree retention and completion rates, the number of students
remaining, and the study shift were identified as the most significant predictive importance for
dropout models.
In Palacios et al. (2021), the authors developed models to predict dropout at the global level
(i.e., independent of the period) and specialized in the first, second, or third year of the degree.
Using academic (school and university), demographic, and economic data of undergraduate stu-
dents at Universidad Católica del Maule (Chile), in addition to the SMOTE balancing technique,
the authors trained DT, KNN, LR, NB, RF, and SVM classifiers for the four prediction objectives.
However, only the models designed to predict dropouts in the second and third years considered
student performance data at the university. As a result, the authors highlighted the importance of
data balancing and good predictive performances, especially of RF models. Moreover, through
information gain analysis, the authors identified that the attributes related to the student’s school
performance and the socioeconomic background of the place of origin are the most important for
the first three scenarios. However, university performance attributes also proved relevant in the
third one. On the other hand, attributes related to students’ university performance proved to be
more critical for predicting dropping out in the third year of the degree (fourth scenario).
As a last example, Shiau (2020) addressed the dropout problem of full-time undergraduate
students on Taiwan’s private and technological university campus. The author used a decision
table and a DT model in the descriptive analysis of dropout determinant rules to help establish
counseling strategies and identify at-risk students. Both techniques related the dropout rules, for
the most part, to students’ low academic performance. Considering the DT model, the author
identified, for example, that first and second-year students with no tuition discounts and average
academic performance lower than 45.6 in the last semester tend to drop out. In addition, to provide
early predictions, Shiau (2020) developed a model based on the discriminant analysis method and
data on student absences. The resulting function was shown to be most influenced by absences
related to sick leave. Therefore, it has been incorporated into a form allowing academic advisors
to enter students’ up-to-date absence figures and get their dropout risk in real time.

4.2 Data

Regarding the perspective of the data used and the third research question of this RSL, Figure 6(a)
shows the distribution of the selected articles according to the types (nature) of attributes and
educational levels that they considered. It is necessary to mention that information related to
student absences was classified as academic in this categorization, although some works cited
in Section 4.1 refer to this data as behavioral. In addition, some studies used complementary
attributes extracted from questionnaires/surveys, such as daily time for leisure and homework,
which were not considered in the categorization. This fact is because, in addition to not being
a common practice among the works, data resulting from questionnaires/surveys are subject to
limitations and biases, as mentioned in Section 1.

238
Colpo et al. RBIE v.32 – 2024

Figure 6: Number of articles by attribute type (a) and dataset size (b).

Considering that we hide the labels related to a single article so as not to pollute the graphs,
notice that the Undergraduate level is the only one to have works linked to all data types, which
is understandable since it is the most researched level. Academic, social, and economic data are
widely used in predicting institutional and educational degree dropout, as opposed to interactional,
temporal, and sentiment analysis data. Although not shown in Figure 6(a), it is worth specifying
some attributes pointed out by the studies as being of greater predictive importance. Among the
academic qualities, we can highlight those related to academic performance, such as admission
exam grades, and grade point and frequency averages (in the term or accumulated), and to the
student’s position in the degree, such as the number of terms/semesters taken, and the total number
of credits earned or lost (in the term or accumulated). Regarding demographic data, the attributes
of age at entry, gender, and time between finishing high school and entering higher education
stand out, in addition to those related to the students’ origin, such as hometown and indicators
of socioeconomic background. Finally, among the economic information, we can mention family
income attributes and student aid or benefits receipt.
While Naseem et al. (2022), Oreshin et al. (2020), Ortigosa et al. (2019), Queiroga et al.
(2020), Deho et al. (2022), Park and Yoo (2021), and Villegas-Ch et al. (2020) correspond to ar-
ticles that used the interactional data type, the work of Oreshin et al. (2020) was the only one to
use predictive attributes resulting from sentiment analysis. However, it is related to both under-
graduate and graduate levels. Yoo et al. (2017) and Kurniawati and Maulidevi (2022), in turn,
considered in their analysis the temporality associated with students’ academic records10 .
Using academic, demographic, and economic data from undergraduate students at an In-
donesian university and taking advantage of the semester temporality/sequentiality of the edu-
cational path data, Kurniawati and Maulidevi (2022) developed recurrent neural network models,
more specifically Long-Short Term Memory (LSTM) and Gate Recurrent Units (GRU) classifiers,
to predict graduation/dropout and final GPA (below average, average, good, very good). As a dif-
ferential, the authors built three models: one model for each prediction task (dropout and final
GPA), constituting a separate method; and one model to predict the data with dropout and final
GPA cross-labels, constituting a combined method. Evaluating the predictive performance of the
models over different semesters, Kurniawati and Maulidevi (2022) found that GRU and LSTM
10 Other studies have used temporal data, such as the semester in which the student is enrolled or the interval
between the end of high school and entry into higher education. Still, they have not extracted information based on
the sequentiality of these data, nor have they considered their time series.

239
Colpo et al. RBIE v.32 – 2024

outperformed in graduation/dropout and GPA prediction, respectively. In addition, they pointed

out that the models showed better results in the separate method, with satisfactory predictive per-
formance from the first and second semesters for the graduation and GPA models, respectively.
Also considering data characteristics, Figure 6(b) shows the relationship between the se-
lected articles and the sizes of the datasets (number of students/records) used, considering each
education level separately. Although the works are based on datasets of quite varied dimensions, it
is clear that studies related to the school education level mostly concentrate on the broadest range
of data. This fact is understandable as these studies usually use data from all schools that make up
the public system of a state or country, as seen in Section 4.1.2. Similarly, the other works in this
range, such as da Silva et al. (2019), Yu et al. (2021) and Oreshin et al. (2020)11 , generally using
undergraduate student data from all degrees or several departments of an institution.
As exceptions, Beaulac and Rosenthal (2019) and Viloria et al. (2019) considered only data
from the Faculty of Arts and Science at the University of Toronto and the Departments of Engi-
neering at the University of Mumbai, respectively. In more detail, Beaulac and Rosenthal (2019)
used first-year academic performance data to build models from the RF and LR (baseline) algo-
rithms to predict whether students would get their degrees. As a result, the authors confirmed
the better performance of the RF model, pointing out the predictive importance of attributes as-
sociated with credits and grades in a seminar subject/course and other subjects/courses related to
specific departments, especially mathematics. Viloria et al. (2019), in turn, used academic, demo-
graphic, and economic data to build predictive models of dropout based on Bayesian Networks
(BN), ANN, and DT algorithms. In addition to considering the results of all models as satisfac-
tory, the authors confirmed, through the interpretation of the models or their attribute rankings,
the hypothesis that academic and socioeconomic conditions (such as scores on the entrance exam
and student benefits, respectively) are decisive for the student permanence or dropout.
On the other hand, works that consider datasets of more restricted size (range of 500 to
1000 students/records) generally address dropout prediction in a single degree, such as Costa et al.
(2021), Hannaford et al. (2021), Naseem et al. (2022), Perez et al. (2018), Queiroga et al. (2020),
and Yoo et al. (2017). However, there are exceptions, such as the Iam-On and Boongoen, 2017b
(2017a, 2017b) studies, in which the datasets, although small, cover academic and demographic
data of students from 26 academic departments at Mae Fah Luang University—as in the first
work, described in Section 4.1, Iam-On and Boongoen (2017b) considered separately the contexts
of students who had just entered and completed the first year of the degree. However, in this
second research, the authors developed predictive models of dropout to validate a proposed data
transformation approach based on the Link-Based Cluster Ensemble. As a result, despite requiring
more computational effort, the authors demonstrated that their approach improved predictions and
outperformed other dimensionality reduction techniques, considering models developed for both
contexts from the DT (C4.5), KNN, NB, and ANN algorithms.
The relationship between the articles and the nature/type of the attribute used, now consid-
ering the perspective of the education modality, is presented in Figure 7(a). Notice that adopting
data types for face-to-face teaching generally follows the same trend as for all selected studies.
However, interactional data show the same representativeness as economic data in distance learn-
ing, being considered by 57% (4 out of 7) of the works in this modality, which confirms the greater
11 Notice that the graph counts the Oreshin et al. (2020) study at both undergraduate and graduate levels.

240
Colpo et al. RBIE v.32 – 2024

exploration of student interaction with VLEs in this context. Additionally, in Figure 7(b), we can
see the mapping about the amount of data types considered in the articles. Notice that most studies
use three types of data, followed by those adopting two or one type.

Figure 7: Total articles by data types and modalities (a); and by combinations of types (b).

We identified the use of four or five types of data only in Ortigosa et al. (2019), Naseem
et al. (2022), Park and Yoo (2021), Villegas-Ch et al. (2020), Kurniawati and Maulidevi (2022),
and Oreshin et al. (2020). Most of these studies consider academic, demographic/social, eco-
nomic, and interactional information (A, D, E, I), with the exception of Kurniawati and Maulidevi
(2022), which considers temporal rather than interactional data (A, D, E, T), and Oreshin et al.
(2020), which adds sentiment analysis data to the first four types (A, D, E, I, S). The works linked
to three types generally used academic, demographic/social, and economic data (A, D, E), as
in Barros et al. (2019), Costa et al. (2021), Gamao and Gerardo (2019), and Solis et al. (2018).
Yoo et al. (2017) was an exception and used academic, demographic, and temporal (A, D, T).
In works related to two types, as in Berka and Marek (2021), Iam-On and Boongoen, 2017b
(2017a, 2017b), and Perchinunno et al. (2021), the authors generally used a combination of aca-
demic and demographic data (A, D). The exceptions correspond to Deho et al. (2022) and Freitas
et al. (2020), who adopted demographic and interactional (D, I) or economic (D, E) data, respec-
tively. It is worth mentioning that Freitas et al. (2020) developed an IoT (Internet of Things)
system for predicting early dropout of engineering students at Instituto Federal do Ceará, Campus
Fortaleza (Brazil). Using only seven socioeconomic attributes, the authors built LR, DT, SVM,
KNN, DNN, and MLP models. An interface has been provided to users so that they can, among
other facilities, select the model to be used in the prediction, submit student records to be pre-
dicted or consult the risk of students with stored data. Regarding the predictive results, the authors
pointed out better performances for the DT and DNN models.
Finally, when the works used only one type, the academic (A) stood out, as in Beaulac and
Rosenthal (2019), Chung and Lee (2019), Lee and Chung (2019), Rovira et al. (2017), de Assis
et al. (2022), and Kuzilek et al. (2021). The only exception was the Queiroga et al. (2020) study,
which considered just interactional data (I), as described in Section 4.1.1.

241
Colpo et al. RBIE v.32 – 2024

4.3 Tools and Techniques

Answering the first item of the third research question, the articles selected in this RSL addressed
dropout prediction mainly through the Classification task, either to identify at-risk students or
to determine attributes and patterns related to dropout. The exceptions were the Iam-On and
Boongoen (2017a), Yoo et al. (2017), da Silva et al. (2019), de Assis et al. (2022), and Shilbayeh
and Abonamah (2021) studies, which used only Clustering, Association, or Regression techniques.
We categorized the algorithms used in the works according to the methods/techniques they
represent. Figure 8(a) shows the frequencies of the categories that showed recurrence. We can
see that the authors adopted ensemble methods in most studies, specifically in 48. Ensemble algo-
rithms have gained attention in recent years because they combine multiple classifiers and usually
increase the quality of predictions due to the greater generalizability of this set of models (Lee &
Chung, 2019). Although not shown in the graph, we should mention that most occurrences of the
ensembles (43) considered DTs as base classifiers, with emphasis on the high frequency of the
RF algorithm used, for example, in Chung and Lee (2019), Solis et al. (2018), and Queiroga et al.
(2022). Also, we can see that DTs are not only applied as base classifiers. Their applications, as
in Freitas et al. (2020), Perez et al. (2018), and Ortigosa et al. (2019), correspond to the second
most recurrent category of an algorithm. The reason may be associated not only with their good
predictive and computational performance but also with the interpretability of their results. This
is because the DTs are constituted by rules describing the discovered patterns (Han et al., 2012).
A similar motivation may account for the fact that linear models, used in Kang and Wang (2018),
Mduma and Machuve (2021), Perchinunno et al. (2021), and Shiau (2020), for example, appear
as the third most frequent category, with emphasis on LR models.

Figure 8: Articles by categories of algorithms (a) and tools/languages (b) used.

Still, regarding Figure 8(a), it is evident that categories related to the SVM, neural network
(simple or deep), and probabilistic (based on Bayes’ theorem) algorithms also had significant
frequencies. The category of instance-based learning, represented mainly by the KNN algorithm,
had a more discrete recurrence. Freitas et al. (2020), Hannaford et al. (2021), and Palacios et
al. (2021) are examples of works that used KNN, among other algorithms, in their comparative
analyses. Although it is also a traditional algorithm, this lower adoption of KNN may be associated
with the computational cost it requires in classifying new instances, which in the worst case may
increase linearly with the size of the training dataset (Raschka & Mirjalili, 2017). Finally, the
other categories were slightly used, having been cited in less than five studies.
242
Colpo et al. RBIE v.32 – 2024

G. Santos et al. (2020) considered academic and demographic data of undergraduate stu-
dents from a Brazilian federal institution in developing a dropout prediction model named Evolve-
DTree.To improve the learning of the DT classifier in the face of the unbalanced dataset, the au-
thors employed the K-Means clustering algorithm in pre-processing. Thus, the data was stratified
into ten groups of records, which were used in training by cross-validation. Furthermore, to avoid
overfitting the DT model, an attribute selection method based on an evolutionary (genetic) algo-
rithm was also used, resulting in eight selected attributes: an indicator of participation in social
programs, degree closure semester, writing grade in the entrance exam, year, and semester of ad-
mission, GPA, race/ethnicity, and the number of years in the degree. In addition to validating the
results of each step of their approach’s results, the authors demonstrated their model’s superior
performance over KNN, AdaBoost, MLP, RF, NB, SVM, and Quadratic Discriminant Analysis
(QDA) models. Finally, the visualization of the model also showed that EvolveDTree considered
potential dropouts, mainly students with lower academic performance or who had not yet passed
the second year of the degree.
Chen et al. (2018) proposed using survival analysis, based on Aalen’s Additive and Cox’s
Proportional Hazard models, to predict the occurrence and timing of undergraduate student dropout.
The research used demographic and academic (school and university) data of students at George
Mason University (USA) who were linked to Science, Technology, Engineering, and Mathematics
(STEM) fields. The models were trained incrementally, considering the information available in
different periods, from the data available at the end of the second semester to the data available at
the end of the fifth semester. The authors compared their survival analysis models to traditional
machine learning models (LR, DT, RF, NB, and AdaBoost). As a result, in addition to observing
that the predictive results improved as new semester data were considered, they identified that
survival analysis models performed better when a few semester data were available. Addition-
ally, the attributes of semester-wise GPA and the number of semesters attended/enrolled showed
greater predictive capacity.
Figure 8(b) shows the tools and programming languages used by the selected articles in
the data mining process. Although some works do not mention the tools adopted, analyzing the
graph makes it possible to see the wide use of Python and R languages. It is necessary to say that
the use of Python appears in general associated with the machine learning library Scikit-Learn,
such as Barros et al. (2019), Queiroga et al. (2020), and Rovira et al. (2017). In the Rovira et
al. (2017) work, models were developed to predict, among other targets, the risk of undergraduate
students at the University of Barcelona not re-enrolling in the second and third year of their degree.
The authors used the academic performance data of students in the first year of Law, Computer
Science, and Mathematics degrees separately. Five prediction models were developed for each
degree, considering the LR, SVM, RF, NB, AdaBoost classifiers, as well as the SMOTE balancing
technique. As a result, the RF and AdaBoost models performed best in the Law degree, while the
LR and NB models excelled in the other two. Considering the sizes of the datasets used for each
degree, the authors pointed out that non-parametric classifiers may be better suited to the context
of low data availability. In contrast, parametric models may be better adjusted to the opposite
situation.
Similarly, the Caret package is frequently used by studies adopting the R language, such
as Perchinunno et al. (2021), Lee and Chung (2019), and Solis et al. (2018). Among the other
tools, only WEKA, RapidMiner, and H2O.ai were referenced more than once by the selected stud-

243
Colpo et al. RBIE v.32 – 2024

ies. Palacios et al. (2021), Viloria et al. (2019), and Vega et al. (2022) are examples of research
using WEKA. The only two references to RapidMiner occurred in Berka and Marek (2021) and
Nagy and Molontay (2018). Although Berka and Marek (2021) used RapidMiner in the classifica-
tion, the authors also applied the LISp-Miner tool in the association mining. Similarly, Nagy and
Molontay (2018) has also used implementations of the H2O.ai platform, also adopted by Deho
et al. (2022), in some of its models.
Based on academic and socioeconomic data of undergraduate students from the Faculty of
Systems Engineering and Informatics of the Universidad Nacional Mayor de San Marcos (Peru),
Vega et al. (2022) developed a DT model and a (not web) system for dropout prediction. The
authors used six feature selection methods available in the WEKA tool to select the features fre-
quently considered necessary, and applied SMOTE to deal with the data unbalanced problem. As
a result, Vega et al. (2022) presented a diagram of use cases and interface prototypes related to the
system developed, highlighting the importance of the cumulative and last cycle weighted average
grades and the number of credits earned features.
Nagy and Molontay (2018) developed early dropout prediction models from academic and
demographic data available when undergraduate students were admitted at the Budapest Univer-
sity of Technology and Economics. In addition to using different attribute selection strategies
(evolutionary, by correlation, and based on feature importance metrics dependent on GBT and
DNN models), the authors considered several classifiers, specifically those of linear regression,
DT, RF, GBT, LR, NB, KNN, DNN, AdaBoost. Among the attributes selected as important, the
authors identified the performance in humanities subjects/courses in high school. This selection
suggests that a good and broad education is essential for university success, even for engineering
and exact sciences students. Regarding the models, GBT and DNN performed better on the se-
lected and complete sets of attributes, respectively. As a final result, Nagy and Molontay (2018)
developed a web application to provide predictions from the GBT model, trained from the re-
duced/selected attribute set. Thus, by entering the attribute data of a specific student, the user can
obtain the prediction of their risk of dropping out right at the time of enrollment.
In addition to the technical characteristics already presented, it is crucial to position the
studies concerning some additional aspects. For example, Figures 9(a) and 9(b) show the propor-
tions of works concerning the adoption of balancing and attribute selection techniques, respec-
tively. These analyses are necessary because unbalanced data tends to be a reality in the context
of dropouts, and this inequality can affect the learning of the underrepresented class of students.
In addition, attribute selection techniques aim to remove irrelevant or redundant dimensions from
the data, reducing the processing cost and preventing these variables from negatively influencing
the models Han et al. (2012).
Notice that only 41% (29) of the works cited the application or evaluation of balancing tech-
niques. Although some databases are naturally balanced, this low percentage shows that many
studies have neglected this critical issue. Among the works that balanced their training data, the
use of the SMOTE technique stands out, as in Hutagaol and Suharjito (2019), Lee and Chung
(2019), and Rovira et al. (2017). Attribute selection techniques were considered in 35% (25) of
the studies. This result is understandable since many works already start from a reduced num-
ber of attributes, either due to a lack of access to more information or because they base the
choice of attributes on previous experiences or research. Among the selection approaches used
in the works already described, we can mention: evolutionary methods, such as Gamao and Ger-

244
Colpo et al. RBIE v.32 – 2024

Figure 9: Proportions of articles concerning technical characteristics.

ardo (2019) and G. Santos et al., 2020; by correlation and stepwise, such as Fernández-García
et al., 2021 and da Silva et al., 2019; LVQ in Hutagaol and Suharjito (2019); selection based on
model-dependent importance metrics, in Nagy and Molontay (2018); and Boruta algorithm, which
operates as an RF wrapper, in Naseem et al. (2022).
Another example of a study applying attribute selection is Tsai et al. (2020). Using de-
mographic, economic, and academic performance data from first-year students at a Taiwanese
university, the authors developed models to predict dropout over the next three years. The authors
performed a statistical analysis based on logistic regression to attribute selection. This analysis
pointed to four main attributes: academic performance (class ranking percentage), student loan
application indicator, number of absences, and number of subjects/courses in which the student
received academic performance alerts. Considering these four attributes for each student’s first
two semesters, MLP and LR classifiers were trained and showed better specificity and sensitivity,
respectively. As a final result, the authors used the most sensitive model in the prediction task,
providing the university with lists of at-risk students, thus enabling the execution of preventive
strategies.
Figure 9(c) shows the proportion of papers that automatically predicted the risk of dropout
for enrolled (active) students. Although many studies (80%) applied and validated prediction
models on historical data to identify patterns related to dropout or define the best prediction algo-
rithm, only 20% (14) of the works applied the models built on current data to predict and detect
future dropout automatically. Specifically, Kang and Wang (2018), Nagy and Molontay (2018),
Ortigosa et al. (2019), Mduma et al. (2019b), Freitas et al. (2020), Shiau (2020), Tsai et al. (2020),
Shilbayeh and Abonamah (2021), Xu and Wilson (2021), Bassetti et al. (2022), Queiroga et al.
(2022), Vega et al. (2022), Demeter et al. (2022), and Urbina-Najera and Mendez-Ortega (2022)
correspond to the latter case.
Interested in the predictive utility of financial aid records, Demeter et al. (2022) used aca-
demic, demographic, and economic data from undergraduate students who applied for free fed-
eral aid in their first two consecutive years of enrollment at a public university in the southeastern
USA. In addition to selecting attributes from a theoretically oriented process, the authors used
supporting analytical and data-driven techniques, such as stepwise logistic regression, and the
model’s feature importance. As a differential, Demeter et al. (2022) used a two-level hierarchical
classification process to make predictions based on records available at the end of the second to
sixth semester after matriculation. This process comprises two RF models, one to initially predict

245
Colpo et al. RBIE v.32 – 2024

whether students would drop out or graduate, and another to be applied to expected graduates to
predict whether they would complete their degree late or on time (4 years). In addition to validat-
ing the predictive performance of their models, the authors applied them to actual student data.
They identified a list of potential dropouts that was shared with academic advisors for their analy-
sis. Demeter et al. (2022) also highlighted the predictive importance of attributes related to credit
earned, college and high school GPA, estimated family financial contribution, and enrollment and
grades in required gateway courses within a student’s major.
Urbina-Najera and Mendez-Ortega (2022) used academic, demographic, and economic data
from undergraduate students in 25 engineering, social science, and administrative science pro-
grams at an unidentified university. By oversampling the minority class (dropouts) and undersam-
pling the majority class (non-dropouts), the authors evaluated different feature selection methods
(RF wrapper, correlation, and consistency-based) in the development of RF and ANN models for
dropout prediction. Among the results, they pointed out that resampling benefited the predictive
ability of the models and that the RF with features selected by the RF wrapper showed better
performance. Urbina-Najera and Mendez-Ortega (2022) also reports that the (best) RF model has
been implemented in the institutional system and has allowed predicting an approximate number
of possible dropouts per period, contributing to the preventing actions.
Another factor to highlight is that not all the studies that predict future dropout made their
predictions available to academic managers through interfaces, as indicated in Figure 9(d). That is,
although 92% of the works developed predictive models for dropout (Figure 4), only 15% ensured
intuitive and continuous access to the actual predictions, facilitating monitoring and development
of preventive strategies for at-risk students. In addition, it is essential to mention that among
the 11 works that provided systems/interfaces, only the ones by Ortigosa et al. (2019), Xu and
Wilson (2021), and Bassetti et al. (2022) involve projects that integrate the prediction systems
with the databases of administrative/institutional systems. The others require the user to inform
the students’ data by filling out a form or submitting a .csv file to have their status predicted.

5 Trends, Opportunities, and Challenges

The analysis of the studies covered by this SLR allows us to observe some trends, opportunities,
and challenges related to the application of EDM and machine learning in predicting student
dropout, considering the degree and institutional levels.
From the contextual perspective, we can highlight the lack and consequent opportunity for
research at school (primary and secondary) and technical education levels and in countries with
the lowest HDI, which are precisely the contexts with the greatest need. This is because basic
education levels cover the most significant number of students and are extremely important for
the socio-economic and educational development of the population. Furthermore, it is necessary
to consider the challenges that may be at the root of this lack and that need to be addressed to
drive progress in this research context. Although they concentrate most of the students, public
schools have limited resources. They may not have dedicated teams specializing in information
technology, making it difficult to conduct this type of research/study locally (within the schools).
Therefore, expanding the use of EDM and machine learning in the school dropout context de-
pends on government bodies investing in the development of this research internally or facilitating

246
Colpo et al. RBIE v.32 – 2024

it externally by making public datasets available, preferably covering data on different aspects of
students12 . Mduma and Machuve (2021) also pointed out the difficulty of finding publicly avail-
able datasets that address the problem of student dropout for the development of studies involving
the application of machine learning techniques. Also in line with the remarks on the school con-
text, in this SLR, two of the three studies related to technical education were carried out in Brazil,
using data from institutions of the Federal Network of Professional and Technological Education,
which also have undergraduate and graduate degrees, providing specialized technical staff and a
more conducive environment for research.
Regarding the data used, we observed that many studies utilize a combination of demo-
graphic, economic, and academic data despite the widespread use of VLEs in universities. This
shows a research opportunity, as analyzing student interaction data in VLEs can aid in identifying
behavioral patterns. Furthermore, they can be precious in early prediction models since they are
accessible throughout the entire semester, even before the student has completed any academic
terms and obtained GPA information (Park & Yoo, 2021). Another opportunity behind the ex-
traction and use of interactional data is that the texts of students’ posts and their participation in
forums can be processed and serve as input for a sentiment mining process, the result of which
can also contribute to the prediction of dropouts. Similarly, further exploration of the temporality
and sequentiality of academic trajectory data is needed. Studies comparing the predictive results
of traditional classification models with those of time series would be of great practical use. They
would shed light on whether the temporality of the data truly offers a predictive advantage in the
dropout context.
In general, the studies on dropout prediction also show little adaptation of techniques to the
specific domain. Although the research questions are focused on the dropout problem, the stud-
ies generally use traditional data mining and machine learning techniques without adapting them
or their applications to the domain’s particularities. Fernández-García et al. (2021) and Demeter
et al. (2022) are examples of studies that propose different classification structures and can serve
as inspiration for other research in this direction. It is also common to develop specialized models
for degrees or prediction periods. Although exploring more specific factors, such as student per-
formance in particular subjects/courses, may benefit such models, it is crucial to notice that this
approach may provide less comprehensive results institutionally. Therefore, when possible, de-
veloping models that focus on the generic prediction of dropout (i.e., prediction models intended
to predict dropouts in different degrees and periods/grades of the academic trajectory) may have
greater returns for educational practice.
Still considering the technical perspective, although we observed the concern to evaluate
different machine learning algorithms (this is indeed important) for the development of new stud-
ies, we should highlight the importance of validating the technical decisions throughout the entire
model development process, considering the institutional reality and the objectives of each re-
search. In developing a generic prediction model, for example, is it best to consider absolute
attributes (such as the number of completed credits)? Would not using relative attributes (such
as the percentage of completed credits) provide more informative and generalizable data for dif-
ferent curricular structures? Studying the choice and prioritization of evaluation metrics is also
extremely necessary. Otherwise, all technical decisions may be based on biased results, generat-
12 This is because much public data, such as that from educational censuses, is limited to socioeconomic information

on students, which limits the research potential and, consequently, the interest of researchers.

247
Colpo et al. RBIE v.32 – 2024

ing falsely viable models. Although many studies point to the good performance of their models,
these conclusions are sometimes not adequately substantiated. For example, classification mod-
els should not be evaluated solely based on their overall hit rates (accuracies), especially if data
imbalance has not been deal with. In addition, although DT-based models have been widely used
because they combine good prediction and execution performance with interpretability, it is worth
noting that there are several possibilities for applying more complex models, such as deep neural
networks. This is because if these models provide a more significant predictive advantage and
acceptable execution times for the reality and needs of the project/research, XAI techniques can
be used to identify the factors contributing to the prediction of dropout (Rondado de Sousa et al.,
2021).
In addition, while there is growing interest in using EDM to predict dropout, most studies
only analyze historical data. They don’t often apply the models to active/enrolled student data
to predict who is at risk of dropping out. Furthermore, not all of the few works that perform
predictions provide academic managers with interfaces/tools for intuitive and continuous access
to the results of the predictions. This would be important to support decision-making and planning
of preventive strategies. We believe that this underutilization of the efforts and potential of many
studies in educational practice is the greatest challenge. This is because, as already mentioned,
most of the research is carried out in an academic environment, often associated with degree
final works, dissertations, or theses. In this type of research, due to privacy issues, access to
data is usually completely anonymous and partial, with no continuous access to databases. As a
result, these studies generally do not allow for adequate progress in identifying at-risk students.
In our view, the solution to this problem depends on carrying out projects involving researchers
and stakeholders linked to the information technology teams and administrative managers of the
institutions that hold the data. In this way, models can be developed in the academic sphere and
integrated into institutional systems with the effort of the partner institution.
Finally, we must point out the limitations and threats to the validity of this RSL. First, due to
the large number of papers collected, we did not use peer review to select and analyze the articles.
In other words, each publication was examined by a single researcher. In the interpretive context
inherent in any secondary study, this can lead to subjective decisions. Another threat could be
the lack of scientific databases that are important for the specific contexts of different countries.
However, since this work aims to investigate the international scenario and we don’t know the
relevant databases for each country, including some specific databases could introduce bias in the
results. For this reason, we decided to consider only four major international scientific databases,
namely ACM Digital Library, IEEE Xplore, Web of Science, and Scopus.

6 Conclusion and Final Remarks

This work presents a systematic review of the application of EDM in the context of student dropout
prediction. This review considered the scopes of institutional and academic degree dropout and
identified contextual, technical, and data characteristics addressed in this research topic.
The selected articles focused more on the undergraduate level, the public education system,
and the face-to-face modality. This last trend may be related to the scope of this work, which
disregards publications associated with the prediction of dropout in subjects/courses that are re-

248
Colpo et al. RBIE v.32 – 2024

current in the distance learning modality. Regarding the information used, besides a remarkable
variation in the number of records considered, almost all studies adopt from one to three types of
data in the representation of students. Combinations of academic, social, and economic attributes
are the most frequently used. Although socioeconomic attributes have been highlighted as nec-
essary by many studies, mainly for the early prediction of dropouts, several studies have pointed
out that, when available, data on students’ academic performance overlap other types of attributes,
dominating the decisions of predictive models.
From the perspective of the techniques and tools used, most works address the problem
of dropout prediction through the classification task, using DT models, individually or in an en-
semble. The use of Python and R languages also stands out in conducting the EDM process. In
addition, although we already expected limited adoption of attribute selection techniques, consid-
ering that many studies start from a small number of variables, the low application of balancing
techniques caused a surprise. This is because dropout datasets usually have a significant imbalance
between classes, which can impair the quality of the models.
Still, on the technical characteristics, although interest in adopting EDM in dropout pre-
diction is growing, most works only analyze past data. They do not apply the models to data of
active/enrolled students to predict who is at risk of dropping out. Moreover, not all of the few
works that perform predictions provide academic managers with interfaces/tools for intuitive and
continuous access to the predictive results. This is important to assist in decision-making and
planning preventive strategies. So, these results indicate the underutilization of the efforts and
potential of many studies in the educational practice.
Finally, at the end of this study, we also presented and discussed research opportunities
and challenges that follow the general trends observed among the works analyzed. Despite the
limitation of not conducting a peer review in carrying out the SLR, we hope that the results and
findings of this work will contribute to the development of new studies related to the application
of the EDM in predicting student dropout. In addition to instigating the expansion of this area
of research in underdeveloped countries and greater caution regarding technical decisions, we
hope to encourage the more practical application of predictive models to enrolled student data
and the availability of these predictions to academic managers. This will undoubtedly add to the
quality and usefulness of the work in educational practice. These observations will also guide the
continuity of our research.

Acknowledgements

This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível
Superior – Brazil (CAPES) – Finance Code 001. The authors thank Coordenação de Aperfeiçoa-
mento de Pessoal de Nível Superior, Universidade Federal de Pelotas, and Instituto Federal de
Educação, Ciência e Tecnologia Farroupilha.

249
Colpo et al. RBIE v.32 – 2024

References
Agrusti, F., Bonavolontà, G., & Mezzini, M. (2019). University dropout prediction through educa-
tional data mining techniques: A systematic review. Journal of e-Learning and Knowledge
Society, 15(3), 161–182. https://doi.org/10.20368/1971-8829/1135017 [GS Search].
Agrusti, F., Mezzini, M., & Bonavolonta, G. (2020). Deep learning approach for predicting univer-
sity dropout: a case study at Roma Tre University. Journal of e-Learning and Knowledge
Society, 16(1, SI), 44–54. https://doi.org/10.20368/1971-8829/1135192 [GS Search].
Aguirre, C. E., & Pérez, J. C. (2020). Predictive data analysis techniques applied to dropping out
of university studies. 2020 XLVI Latin American Computing Conference (CLEI), 512–521.
https://doi.org/10.1109/CLEI52000.2020.00066 [GS Search].
Alturki, S., Cohausz, L., & Stuckenschmidt, H. (2022). Predicting master’s students’ academic
performance: An empirical study in germany. Smart Learning Environments, 9(1). https:
//doi.org/10.1186/s40561-022-00220-y [GS Search].
Baker, R., Isotani, S., & Carvalho, A. (2011). Mineração de dados educacionais: Oportunidades
para o brasil. Revista Brasileira de Informática na Educação, 19(02), 03–13. https://doi.
org/10.5753/rbie.2011.19.02.03 [GS Search].
Baranyi, M., Nagy, M., & Molontay, R. (2020). Interpretable deep learning for university dropout
prediction. Proceedings of the 21st Annual Conference on Information Technology Educa-
tion, 13–19. https://doi.org/10.1145/3368308.3415382 [GS Search].
Barros, T. M., Souza Neto, P. A., Silva, I., & Guedes, L. A. (2019). Predictive Models for Imbal-
anced Data: A School Dropout Perspective. Education Sciences, 9(4), 275, 1–17. https:
//doi.org/10.3390/educsci9040275 [GS Search].
Bassetti, E., Conti, A., Panizzi, E., & Tolomei, G. (2022). ISIDE: Proactively Assist University
Students at Risk of Dropout. 2022 IEEE International Conference on Big Data (Big Data),
1776–1783. https://doi.org/10.1109/BigData55660.2022.10020920 [GS Search].
Beaulac, C., & Rosenthal, J. S. (2019). Predicting University Students’ Academic Success and
Major Using Random Forests. Research in Higher Education, 60(7), 1048–1064. https:
//doi.org/10.1007/s11162-019-09546-y [GS Search].
Berka, P., & Marek, L. (2021). Bachelor’s degree student dropouts: Who tend to stay and who
tend to leave? Studies in Educational Evaluation, 70, 100999. https://doi.org/10.1016/j.
stueduc.2021.100999 [GS Search].
Bitencourt, W. A., Silva, D. M., & do Carmo Xavier, G. (2022). Pode a inteligência artificial
apoiar ações contra evasão escolar universitária. Ensaio, 30(116), 669–694. https://doi.
org/10.1590/S0104-403620220003002854 [GS Search].
Böttcher, A., Thurner, V., Häfner, T., & Hertle, J. (2021). A data science-based approach for
identifying counseling needs in first-year students. 2021 IEEE Global Engineering Edu-
cation Conference (EDUCON), 420–429. https://doi.org/10.1109/EDUCON46332.2021.
9454042 [GS Search].
Brasil. (1996). Diplomação, retenção e evasão nos cursos de graduação em instituições de ensino
superior públicas (tech. rep.). Ministério da Educação, Comissão Especial de Estudos so-
bre a Evasão nas Universidades Públicas Brasileiras: ANDIFES; ABRUEM; SESu/MEC.
Brasília, DF. http://dominiopublico.mec.gov.br/pesquisa/DetalheObraForm.do?select%
5C_action=&co%5C_obra=27010 [GS Search].

250
Colpo et al. RBIE v.32 – 2024

Chen, Y., Johri, A., & Rangwala, H. (2018). Running out of stem: A comparative study across
stem majors of college students at-risk of dropping out early. Proceedings of the 8th Inter-
national Conference on Learning Analytics and Knowledge, 270–279. https://doi.org/10.
1145/3170358.3170410 [GS Search].
Chung, J. Y., & Lee, S. (2019). Dropout early warning systems for high school students using
machine learning. Children and Youth Services Review, 96, 346–353. https://doi.org/10.
1016/j.childyouth.2018.11.030 [GS Search].
Colpo, M. P., Primo, T. T., & de Aguiar, M. S. (2021). Predição da evasão estudantil: Uma análise
comparativa de diferentes representações de treino na aprendizagem de modelos genéricos.
Anais do XXXII Simpósio Brasileiro de Informática na Educação, 873–884. https://doi.
org/10.5753/sbie.2021.218517 [GS Search].
Colpo, M. P., Primo, T. T., Pernas, A. M., & Cechinel, C. (2020). Mineração de dados educacionais
na previsão de evasão: Uma RSL sob a perspectiva do congresso brasileiro de informática
na educação. Anais do XXXI Simpósio Brasileiro de Informática na Educação, 1102–1111.
https://doi.org/10.5753/cbie.sbie.2020.1102 [GS Search].
Costa, A. G., Mattos, J. C. B., Primo, T. T., Cechinel, C., & Muñoz, R. (2021). Model for predic-
tion of student dropout in a computer science course. 2021 XVI Latin American Conference
on Learning Technologies (LACLO), 137–143. https://doi.org/10.1109/LACLO54177.
2021.00020 [GS Search].
Crespo, C. (2020). Two become one: Improving the targeting of conditional cash transfers with a
predictive model of school dropout. Economia-Journal of the Latin American and Caribbean
Economic Association, 21(1), 1–45. https://doi.org/10.1353/eco.2020.0011 [GS Search].
da Silva, P. M., Lima, M. N. C. A., Soares, W. L., Silva, I. R. R., de A. Fagundes, R. A., & de
Souza, F. F. (2019). Ensemble regression models applied to dropout in higher education.
2019 8th Brazilian Conference on Intelligent Systems (BRACIS), 120–125. https://doi.org/
10.1109/BRACIS.2019.00030 [GS Search].
de Assis, B. d. S., Ogasawara, E., Barbastefano, R., & Carvalho, D. (2022). Frequent pattern
mining augmented by social network parameters for measuring graduation and dropout
time factors: A case study on a production engineering course. Spcio-Economic Planning
Sciences, 81. https://doi.org/10.1016/j.seps.2021.101200 [GS Search].
Deho, O. B., Zhan, C., Li, J., Liu, J., Liu, L., & Le, T. D. (2022). How do the existing fairness met-
rics and unfairness mitigation algorithms contribute to ethical learning analytics? British
Journal of Education Technology, 53(4), 822–843. https://doi.org/10.1111/bjet.13217 [GS
Search].
Del Bonifro, F., Gabbrielli, M., Lisanti, G., & Zingaro, S. (2020). Student dropout prediction.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intel-
ligence and Lecture Notes in Bioinformatics), 12163 LNAI, 129–140. https://doi.org/10.
1007/978-3-030-52237-7_11 [GS Search].
Demeter, E., Dorodchi, M., Al-Hossami, E., Benedict, A., Walker, L. S., & Smail, J. (2022).
Predicting first-time-in-college students’ degree completion outcomes. Higher Education,
84(3), 589–609. https://doi.org/10.1007/s10734-021-00790-9 [GS Search].
de Oliveira, C. F., Sobral, S. R., Ferreira, M. J., & Moreira, F. (2021). How does learning analytics
contribute to prevent students’ dropout in higher education: A systematic literature review.
Big Data and Cognitive Computing, 5(4). https : / / doi . org / 10 . 3390 / bdcc5040064 [GS
Search].

251
Colpo et al. RBIE v.32 – 2024

Fernández-García, A. J., Preciado, J. C., Melchor, F., Rodriguez-Echeverria, R., Conejero, J. M.,
& Sánchez-Figueroa, F. (2021). A real-life machine learning experience for predicting uni-
versity dropout at different stages using academic data. IEEE Access, 9, 133076–133090.
https://doi.org/10.1109/ACCESS.2021.3115851 [GS Search].
Flores, V., Heras, S., & Julian, V. (2022). Comparison of predictive models with balanced classes
using the smote method for the forecast of student dropout in higher education. Electronics,
11(3). https://doi.org/10.3390/electronics11030457 [GS Search].
Fontana, L., Masci, C., Ieva, F., & Paganoni, A. M. (2021). Performing learning analytics via
generalised mixed-effects trees. Data, 6(7). https : / / doi . org / 10 . 3390 / data6070074 [GS
Search].
Freitas, F. A. d. S., Vasconcelos, F. F. X., Peixoto, S. A., Hassan, M. M., Dewan, M. A. A., de
Albuquerque, V. H. C., & Reboucas Filho, P. P. (2020). IoT System for School Dropout
Prediction Using Machine Learning Techniques Based on Socioeconomic Data. Electron-
ics, 9(10), 1613, 1–14. https://doi.org/10.3390/electronics9101613 [GS Search].
Gamao, A., & Gerardo, B. (2019). Prediction-based model for student dropouts using modified
mutated firefly algorithm. International Journal of Advanced Trends in Computer Science
and Engineering, 8(6), 3461–3469. https://doi.org/10.30534/ijatcse/2019/122862019 [GS
Search].
Han, J., Kamber, M., & Pei, J. (2012). Data mining: Concepts and techniques (3rd). Morgan
Kaufmann Publishers. [GS Search].
Hannaford, L., Cheng, X., & Kunes-Connell, M. (2021). Predicting nursing baccalaureate program
graduates using machine learning models: A quantitative research study. Nurse Education
Today, 99, 104784. https://doi.org/10.1016/j.nedt.2021.104784 [GS Search].
Hoffait, A.-S., & Schyns, M. (2017). Early detection of university students with potential difficul-
ties. Decision Support Systems, 101, 1–11. https://doi.org/10.1016/j.dss.2017.05.003 [GS
Search].
Hutagaol, N., & Suharjito. (2019). Predictive modelling of student dropout using ensemble clas-
sifier method in higher education. Advances in Science, Technology and Engineering Sys-
tems, 4(4), 206–211. https://doi.org/10.25046/aj040425 [GS Search].
Iam-On, N., & Boongoen, T. (2017a). Generating descriptive model for student dropout: a review
of clustering approach. Human-Centric Computing and Information Sciences, 7, 1, 1–24.
https://doi.org/10.1186/s13673-016-0083-0 [GS Search].
Iam-On, N., & Boongoen, T. (2017b). Improved student dropout prediction in Thai University
using ensemble of mixed-type data clusterings. International Journal of Machine Learning
and Cybernetics, 8(2), 497–510. https://doi.org/10.1007/s13042-015-0341-x [GS Search].
Kang, K., & Wang, S. (2018). Analyze and predict student dropout from online programs. Pro-
ceedings of the 2nd International Conference on Compute and Data Analysis, 6–12. https:
//doi.org/10.1145/3193077.3193090 [GS Search].
Karimi-Haghighi, M., Castillo, C., & Hernandez-Leo, D. (2022). A causal inference study on
the effects of first year workload on the dropout rate of undergraduates. In M. Rodrigo,
N. Matsuda, A. Cristea, & V. Dimitrova (Eds.), Artificial intelligence in education, pt i
(pp. 15–27, Vol. 13355). https://doi.org/10.1007/978-3-031-11644-5_2 [GS Search].
Kiss, B., Nagy, M., Molontay, R., & Csabay, B. (2019). Predicting dropout using high school
and first-semester academic achievement measures. 2019 17th International Conference

252
Colpo et al. RBIE v.32 – 2024

on Emerging eLearning Technologies and Applications (ICETA), 383–389. https://doi.org/

10.1109/ICETA48886.2019.9040158 [GS Search].
Kitchenham, B. A., & Charters, S. (2007). Guidelines for performing systematic literature reviews
in software engineering (tech. rep. No. EBSE-2007-01). School of Computer Science and
Mathematics, Keele University. Keele, UK. https://www.elsevier.com/__data/promis_
misc/525444systematicreviewsguide.pdf. [GS Search].
Kurniawati, G., & Maulidevi, N. U. (2022). Multivariate sequential modelling for student per-
formance and graduation prediction. 2022 9th International Conference on Information
Technology, Computer, and Electrical Engineering (ICITACEE), 293–298. https://doi.org/
10.1109/ICITACEE55701.2022.9923971 [GS Search].
Kuzilek, J., Zdrahal, Z., & Fuglik, V. (2021). Student success prediction using student exam be-
haviour. Future Generation Computer Systems - The International Journal of Escience,
125, 661–671. https://doi.org/10.1016/j.future.2021.07.009 [GS Search].
Lee, S., & Chung, J. Y. (2019). The Machine Learning-Based Dropout Early Warning System for
Improving the Performance of Dropout Prediction. Applied Sciences-Basel, 9(15), 3093,
1–14. https://doi.org/10.3390/app9153093 [GS Search].
Lottering, R., Hans, R., & Lall, M. (2020). A model for the identification of students at risk of
dropout at a university of technology. 2020 International Conference on Artificial Intel-
ligence, Big Data, Computing and Data Communication Systems (icABCD), 1–8. https:
//doi.org/10.1109/icABCD49160.2020.9183874 [GS Search].
Marques, L. T., Castro, A. F. D., Marques, B. T., Silva, J. C. P., & Queiroz, P. G. G. (2019).
Mineração de dados auxiliando na descoberta das causas da evasão escolar: Um mapea-
mento sistemático da literatura. Novas Tecnologias na Educação, 17(3), 194–203. https:
//doi.org/10.22456/1679-1916.99470 [GS Search].
Masood, S. W., & Begum, S. A. (2022). Comparison of resampling techniques for imbalanced
datasets in student dropout prediction. 2022 IEEE Silchar Subsection Conference (SIL-
CON), 1–7. https://doi.org/10.1109/SILCON55242.2022.10028915 [GS Search].
Mduma, N., Kalegele, K., & Machuve, D. (2019a). A survey of machine learning approaches
and techniques for student dropout prediction. Data Science Journal, 18:14, 1–10. https:
//doi.org/10.5334/dsj-2019-014 [GS Search].
Mduma, N., Kalegele, K., & Machuve, D. (2019b). An ensemble predictive model based prototype
for student drop-out in secondary schools. Journal of Information Systems Engineering
and Management, 4(3). https://doi.org/10.29333/jisem/5893 [GS Search].
Mduma, N., & Machuve, D. (2021). Machine learning model for predicting student dropout: A
case of tanzania, kenya and uganda. 2021 IEEE AFRICON, 1–6. https://doi.org/10.1109/
AFRICON51333.2021.9570956 [GS Search].
Nagy, M., & Molontay, R. (2018). Predicting dropout in higher education based on secondary
school performance. 2018 IEEE 22nd International Conference on Intelligent Engineering
Systems (INES), 000389–000394. https://doi.org/https://doi.org/10.1109/INES.2018.
8523888 [GS Search].
Naseem, M., Chaudhary, K., & Sharma, B. (2022). Predicting freshmen attrition in computing
science using data mining. Education and Information Technologies, 27(7), 9587–9617.
https://doi.org/10.1007/s10639-022-11018-3 [GS Search].
Nuanmeesri, S., Poomhiran, L., Chopvitayakun, S., & Kadmateekarun, P. (2022). Improving dropout
forecasting during the covid-19 pandemic through feature selection and multilayer per-

253
Colpo et al. RBIE v.32 – 2024

ceptron neural network. International Journal of Information and Education Technology,

12(9), 851–857. https://doi.org/10.18178/ijiet.2022.12.9.1693 [GS Search].
Opazo, D., Moreno, S., Alvarez-Miranda, E., & Pereira, J. (2021). Analysis of first-year university
student dropout through machine learning models: A comparison between universities.
Mathematics, 9(20). https://doi.org/10.3390/math9202599 [GS Search].
Oreshin, S., Filchenkov, A., Petrusha, P., Krasheninnikov, E., Panfilov, A., Glukhov, I., Kaliberda,
Y., Masalskiy, D., Serdyukov, A., Kazakovtsev, V., Khlopotov, M., Podo-lenchuk, T., Smetan-
nikov, I., & Kozlova, D. Implementing a machine learning approach to predicting students
academic outcomes. In: 2020, 78–83. https : / / doi . org / 10 . 1145 / 3437802 . 3437816 [GS
Search].
Orooji, M., & Chen, J. (2019). Predicting louisiana public high school dropout through imbalanced
learning techniques. 2019 18th IEEE International Conference On Machine Learning And
Applications (ICMLA), 456–461. https : / / doi . org / 10 . 1109 / ICMLA . 2019 . 00085 [GS
Search].
Ortigosa, A., Carro, R. M., Bravo-Agapito, J., Lizcano, D., Alcolea, J. J., & Blanco, O. (2019).
From lab to production lessons learnt and real-life challenges of an early student-dropout
prevention system. IEEE Transactions on Learning Technologies, 12(2), 264–277. https:
//doi.org/https://doi.org/10.1109/TLT.2019.2911608 [GS Search].
Pachas, D. A. G., Garcia-Zanabria, G., Cuadros-Vargas, A. J., Camara-Chavez, G., Poco, J., &
Gomez-Nieto, E. (2021). A comparative study of who and when prediction approaches
for early identification of university students at dropout risk. 2021 XLVII Latin American
Computing Conference (CLEI), 1–10. https://doi.org/10.1109/CLEI53233.2021.9640119
[GS Search].
Palacios, C. A., Reyes-Suarez, J. A., Bearzotti, L. A., Leiva, V., & Marchant, C. (2021). Knowl-
edge Discovery for Higher Education Student Retention Based on Data Mining: Machine
Learning Algorithms and Case Study in Chile. Entropy, 23(4), 485, 1–23. https://doi.org/
10.3390/e23040485 [GS Search].
Park, H. S., & Yoo, S. J. (2021). Early dropout prediction in online learning of university us-
ing machine learning. International Journal on Informatics Visualization, 5(4), 347–353.
https://doi.org/10.30630/JOIV.5.4.732 [GS Search].
Perchinunno, P., Bilancia, M., & Vitale, D. (2021). A statistical analysis of factors affecting higher
education dropouts. Spcial Indicators Research, 156(2-3, SI), 341–362. https://doi.org/10.
1007/s11205-019-02249-y [GS Search].
Perez, B., Castellanos, C., & Correal, D. (2018). Predicting student drop-out rates using data
mining techniques: A case study. In A. Orjuela-Canon, J. Figueroa-Garcia, & J. Arias-
Londono (Eds.), Applications of computational intelligence, ColCACI 2018 (pp. 111–125,
Vol. 833). https://doi.org/10.1007/978-3-030-03023-0_10 [GS Search].
Pontili, R., Staduto, J., & Henrique, J. (2018). Abandono e atraso escolar e sua relação com in-
dicadores socioeconômicos: Uma análise para a região sul do brasil. Gestão & Regionali-
dade, 34(101), 4–22. https://doi.org/10.13037/gr.vol34n101.4173 [GS Search].
Prada, M. Á., Domínguez, M., Vicario, J. L., Alves, P. A. V., Barbu, M., Podpora, M., Spagnolini,
U., Pereira, M. J. V., & Vilanova, R. (2020). Educational data mining for tutoring support
in higher education: A web-based tool case study in engineering degrees. IEEE Access, 8,
212818–212836. https://doi.org/10.1109/ACCESS.2020.3040858 [GS Search].

254
Colpo et al. RBIE v.32 – 2024

Queiroga, E. M., Batista Machado, M. F., Paragarino, V. R., Primo, T. T., & Cechinel, C. (2022).
Early prediction of at-risk students in secondary education: A countrywide k-12 learning
analytics initiative in uruguay. Information, 13(9). https://doi.org/10.3390/info13090401
[GS Search].
Queiroga, E. M., Lopes, J. L., Kappel, K., Aguiar, M., Araujo, R. M., Munoz, R., Villarroel, R.,
& Cechinel, C. (2020). A Learning Analytics Approach to Identify Students at Risk of
Dropout: A Case Study with a Technical Distance Education Course. Applied Sciences-
Basel, 10(11), 3998. https://doi.org/10.3390/app10113998 [GS Search].
Raschka, S., & Mirjalili, V. (2017). Python machine learning: Machine learning and deep learning
with python, scikit-learn, and tensorflow (2nd). Packt Publishing. [GS Search].
Romero, C., & Ventura, S. (2020). Educational data mining and learning analytics: An updated
survey. WIREs Data Mining and Knowledge Discovery, 10(3), e1355. https://doi.org/10.
1002/widm.1355 [GS Search].
Rondado de Sousa, L., Oliveira de Carvalho, V., Penteado, B. E., & Affonso, F. J. A system-
atic mapping on the use of data mining for the face-to-face school dropout problem. In:
In Proceedings of the 13th international conference on computer supported education -
volume 1: Csedu. INSTICC. SciTePress, 2021, 36–47. ISBN: 978-989-758-502-9. https:
//doi.org/10.5220/0010476300360047 [GS Search].
Rovira, S., Puertas, E., & Igual, L. (2017). Data-driven system to predict academic grades and
dropout. PLOS ONE, 12(2), e0171207. https://doi.org/10.1371/journal.pone.0171207 [GS
Search].
Santos, G., Belloze, K., Tarrataca, L., Haddad, D., Bordignon, A., & Brandao, D. (2020). Evolvedtree:
Analyzing student dropout in universities. 2020 International Conference on Systems, Sig-
nals and Image Processing (IWSSIP), 173–178. https://doi.org/10.1109/IWSSIP48289.
2020.9145203 [GS Search].
Santos, J. R., & Zaboroski, E. (2020). Ensino remoto e pandemia de COVID-19: Desafios e opor-
tunidades de alunos e professores. Interacções, 16(55), 41–57. https://doi.org/10.25755/
int.20865 [GS Search].
Segura, M., Mello, J., & Hernandez, A. (2022). Machine learning prediction of university student
dropout: Does preference play a key role? Mathematics, 10(18). https://doi.org/10.3390/
math10183359 [GS Search].
Shiau, Y. (2020). University dropout prevention through the application of big data. Proceedings
of the 2020 3rd International Conference on Information Management and Management
Science, 1–7. https://doi.org/10.1145/3416028.3416029 [GS Search].
Shilbayeh, S., & Abonamah, A. (2021). Predicting student enrolments and attrition patterns in
higher educational institutions using machine learning. International Arab Journal of In-
formation Technology, 18(4), 562–567. https://doi.org/10.34028/18/4/8 [GS Search].
Silva, G. P. d. (2013). Análise de evasão no ensino superior: Uma proposta de diagnóstico de seus
determinantes. Avaliação: Revista da Avaliação da Educação Superior (Campinas), 18(2),
311–333. https://doi.org/10.1590/S1414-40772013000200005 [GS Search].
Silva Filho, R. L. L., Motejunas, P. R., Hipolito, O., & Lobo, M. B. C. M. (2007). A evasão no
ensino superior brasileiro. Cadernos de Pesquisa, 37(132), 641–659. https://doi.org/10.
1590/S0100-15742007000300007 [GS Search].
Solis, M., Moreira, T., Gonzalez, R., Fernandez, T., & Hernandez, M. (2018). Perspectives to
predict dropout in university students with machine learning. 2018 IEEE International

255
Colpo et al. RBIE v.32 – 2024

Work Conference on Bioinspired Intelligence (IWOBI), 1–6. https : / / doi . org / 10 . 1109 /
IWOBI.2018.8464191 [GS Search].
Sorensen, L. C. (2019). “Big Data” in Educational Administration: An Application for Predicting
School Dropout Risk. Educational Administration Quarterly, 55(3), 404–446. https://doi.
org/10.1177/0013161X18799439 [GS Search].
Tsai, S.-C., Chen, C.-H., Shiao, Y.-T., Ciou, J.-S., & Wu, T.-N. (2020). Precision education with
statistical learning and deep learning: a case study in Taiwan. International Journal of
Educational Technology in Higher Education, 17(1), 12. https://doi.org/10.1186/s41239-
020-00186-2 [GS Search].
Urbina-Najera, A. B., & Mendez-Ortega, L. A. (2022). Predictive model for taking decision to
prevent university dropout. International Journal of Interactive Multimedia and Artificial
Intelligence, 7(4), 205–213. https://doi.org/10.9781/ijimai.2022.01.006 [GS Search].
Vasquez Verdugo, J., Gitiaux, X., Ortega, C., & Rangwala, H. (2022). Faired: A systematic fair-
ness analysis approach applied in a higher educational context. LAK22: 12th International
Learning Analytics and Knowledge Conference, 271–281. https : / / doi . org / 10 . 1145 /
3506860.3506902 [GS Search].
Vega, H., Sanez, E., De La Cruz, P., Moquillaza, S., & Pretell, J. (2022). Intelligent system to
predict university students dropout. International journal of online and biomedical engi-
neering, 18(7), 27–43. https://doi.org/10.3991/ijoe.v18i07.30195 [GS Search].
Villegas-Ch, W., Palacios-Pacheco, X., & Lujan-Mora, S. (2020). A business intelligence frame-
work for analyzing educational data. Sustainability, 12(14). https : / / doi . org / 10 . 3390 /
su12145745 [GS Search].
Viloria, A., Garcia Padilla, J., Vargas-Mercado, C., Hernandez-Palma, H., Orellano Llinas, N.,
& Arrozola David, M. (2019). Integration of data technology for analyzing university
dropout. In E. Shakshuki, A. Yasar, & H. Malik (Eds.), 16th International Conference
on Mobile Systems and Pervasive Computing (MOBISPC 2019), 14th International Con-
ference on Future Networks and Communications (FNC-2019), 9TH International Con-
ference on Sustainable Energy Information Technology (pp. 569–574, Vol. 155). https :
//doi.org/10.1016/j.procs.2019.08.079 [GS Search].
Xu, Y., & Wilson, K. (2021). Early alert systems during a pandemic: A simulation study on the
impact of concept drift. LAK21: 11th International Learning Analytics and Knowledge
Conference, 504–510. https://doi.org/10.1145/3448139.3448190 [GS Search].
Yang, H., Olson, T. W., & Puder, A. (2021). Analyzing computer science students’ performance
data to identify impactful curricular changes. 2021 IEEE Frontiers in Education Confer-
ence (FIE), 1–9. https://doi.org/10.1109/FIE49875.2021.9637474 [GS Search].
Yoo, J. S., Woo, Y.-S., & Park, S. J. (2017). Mining course trajectories of successful and failure
students: A case study. 2017 IEEE International Conference on Big Knowledge (ICBK),
270–275. https://doi.org/10.1109/ICBK.2017.55 [GS Search].
Yu, R., Lee, H., & Kizilcec, R. F. (2021). Should college dropout prediction models include pro-
tected attributes? Proceedings of the Eighth ACM Conference on Learning @ Scale, 91–
100. https://doi.org/10.1145/3430895.3460139 [GS Search].

256