Vol. 8, No.

1, February 2026

Student Performance Classification Using Academic,

Socioeconomic, and Digital Behavior Features: A Comparative
Study
Muhammad Arifin1, Fajar Nugraha2, Diana Laily Fithri3
123
Department of Information System, Faculty of Engineering, Muria Kudus University, Kudus, Indonesia

Received: Abstract. Accurate prediction of student academic performance is

November 21, 2025 essential for universities seeking to improve learning outcomes and
Revised:
deliver timely, data-driven support. Prior work commonly uses
January 20, 2025
regression to estimate Grade Point Average (GPA), yet numeric
Accepted:
predictions can be difficult for administrators to translate into
January 30, 2026
actionable risk levels. This study reframes the task as binary
Published:
February 13, 2026 classification, categorizing students as good (GPA ≥ 3.00) or poor
(GPA < 3.00) performers. Using 2,423 records from multiple
programs at an Indonesian university, we combine academic
indicators from the learning management system (login frequency,
Corresponding Author: assignment submission, and forum activity) with socio-economic
Author Name*:
and digital behavioral variables (parental income, extracurricular
Muhammad Arifin
participation, study-group involvement, and social media use). Seven
Email*:
machine learning models—Naïve Bayes, Generalized Linear Model,
arifin.m@[Link]
Logistic Regression, Deep Learning, Decision Tree, Random Forest,
and Gradient Boosted Trees (GBT)—are benchmarked under a
consistent evaluation design. Results indicate that integrating
DOI: academic, socio-economic, and digital behavioral features improves
10.63158/journalisi.v8i1.1460 classification performance, and ensemble methods outperform
single, traditional models. GBT yields the best accuracy of 0.75,
offering a practical basis for early-warning dashboards and
targeted interventions. The study provides comparative evidence
© 2026 Journal of
from Indonesian higher education and highlights the value of
Information Systems and
incorporating digital engagement signals alongside conventional
Informatics. This open
access article is distributed academic data for more effective student support services.
under a (CC-BY License)
Keywords: Performance prediction; Machine learning; LMS
analytics; Digital behavior; Indonesian higher education

223

Journal of Information Systems and Informatics | ISSN: 2656-5935 | e-ISSN: 2656-4882 | pp. 223-239
Published by Asosiasi Doktor Sistem Informasi Indonesia
 Vol. 8, No. 1, February 2026

1. INTRODUCTION

Student academic performance prediction has emerged as a central topic in educational

data mining (EDM) and learning analytics over the past decade [1]–[4]. Higher education
institutions increasingly rely on data-driven insights to identify students who may be at
risk of underachievement and to design targeted academic interventions [5]–[8].
Understanding the complex relationships between academic, socio-economic, and
behavioral factors is therefore essential for enhancing learning outcomes, improving
retention rates, and supporting institutional decision-making [9]–[11].

Earlier approaches to student performance prediction primarily employed regression-

based models, focusing on estimating continuous academic outcomes such as Grade
Point Average (GPA) [1], [12], [13]. Although regression offers numerical precision, its
outputs are often less interpretable and less actionable for academic administrators, who
typically require clear indicators of student risk levels rather than exact GPA values [14],
[15]. As a result, classification-based approaches that group students into discrete
performance categories have gained increasing attention.

A growing body of research has applied machine learning classification algorithms,

including Decision Tree, Random Forest, Support Vector Machine, and Logistic Regression,
to identify patterns that distinguish high-performing from low-performing students.
However, most existing works focus mainly on academic records and basic demographic
variables, while behavioral and digital engagement factors, including online learning
activities and social media usage, remain underexplored, particularly in rapidly digitalizing
learning environments Furthermore, many classification studies employ a limited number
of algorithms or single-model approaches, restricting the evaluation of model robustness.
Systematic comparisons across multiple machine learning techniques using the same
dataset remain relatively scarce, despite their importance for ensuring reliable and
methodologically sound predictive results [16]–[18]. Another significant gap concerns
contextual diversity. The majority of prior studies are conducted in Western or large-
scale educational settings, with limited empirical evidence from Indonesian higher
education [19]–[21]. Indonesia’s socio-economic diversity and distinctive digital
engagement patterns may influence student performance differently, highlighting the
need for context-specific investigations [22], [23]. Exploring these factors can therefore

Muhammad Arifin, Fajar Nugraha, et al | 224

 Vol. 8, No. 1, February 2026

offer new perspectives on how academic and non-academic variables jointly shape
student success [24]–[26].

Accordingly, this study aims to develop and evaluate a comprehensive machine learning
based classification framework for predicting student academic performance. The
proposed framework integrates academic variables derived from Learning Management
System (LMS) activity, socio-economic and demographic indicators, and digital behavioral
features that reflect broader patterns of student engagement.

The study addresses the following research questions (RQs): RQ1: How effectively can
machine learning classification models predict student academic performance when
integrating academic, socio-economic, and digital behavioral features?, RQ2: Which
machine learning algorithm provides the most accurate and reliable classification of
students into good (GPA ≥ 3.00) and poor (GPA < 3.00) performance categories?. Seven
machine learning algorithms Naïve Bayes, Generalized Linear Model, Logistic Regression,
Deep Learning, Decision Tree, Random Forest, and Gradient Boosted Trees are evaluated
using a dataset of 2,423 students from multiple study programs within a single university.
This comparative design enables a consistent and robust assessment of model
performance.

The scope of this study is limited to one Indonesian university; therefore, the findings
may not fully generalize to other institutional contexts. Nevertheless, the results provide
valuable empirical insights into student performance prediction in Indonesian higher
education and contribute to the broader learning analytics literature.

2. METHODS

This study employed a quantitative research approach to predict student academic

performance by integrating academic, socio-economic, and digital behavioral data within
a machine learning–based classification framework. The methodological design
emphasizes transparency and replicability through clearly defined stages, including data
collection, preprocessing, model development, and performance evaluation.

225 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

2.1. Research Design

This research adopts a supervised classification approach to predict student academic
performance. It extends prior regression-based analysis by transforming continuous GPA
prediction into categorical classification, thereby improving interpretability and
institutional relevance. The classification framework is intended to support academic
decision-making by clearly distinguishing students at risk of underperformance. As
illustrated in Figure 1, the research workflow consists of data collection, data
preprocessing, feature analysis, model training, and comparative evaluation.

Figure 1. Research Methodology

2.2. Dataset Description

The dataset was collected from a single Indonesian higher education institution and
consists of 2,423 student records drawn from multiple study programs and academic
years. Each record integrates academic activity data obtained from the Learning
Management System (LMS) with survey-based non-academic and digital behavioral
attributes. Student academic performance was defined as the target variable and
categorized according to institutional academic regulations: good performance (GPA ≥
3.00) and poor performance (GPA < 3.00). The dataset exhibits a moderately imbalanced

Muhammad Arifin, Fajar Nugraha, et al | 226

 Vol. 8, No. 1, February 2026

class distribution, with the good-performance category slightly more prevalent than the
poor-performance category.

The dataset consists of 2,423 student records categorized into two performance classes
based on GPA. A total of 1,512 students (62.4%) were classified as good performers (GPA
≥ 3.00), while 911 students (37.6%) were categorized as poor performers (GPA < 3.00). This
distribution indicates a moderate class imbalance, which was addressed during model
training using SMOTE applied exclusively to the training data. A summary of feature
descriptions, and data types is provided in Table 2.

Table 2. Feature descriptions, and data types

Feature Category Feature Name Description Data Type

Total number of LMS login

LMS Logins Numerical
activities
Academic Assignments Number of assignments
Numerical
Variables Submitted submitted via LMS
Number of discussion posts
Forum Participation Numerical
or replies in LMS forums
Average monthly family
Parental Income Categorical
income (categorized ranges)
Participation in student
Organizational
organizations or Binary
Non-Academic Involvement
extracurriculars
Variables
Student gender
Gender Binary
(male/female)
Distance from Distance between residence
Numerical
Campus and campus (in kilometers)
Average daily hours spent
Social Media Usage Numerical
Digital Behavioral on social media
Vars Study Group Regular participation in
Binary
Participation collaborative study groups

227 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

2.3. Feature Definition

Predictor variables were grouped into three categories to capture complementary
dimensions of student learning behavior. Academic variables represent formal learning
engagement, including LMS login frequency, assignment submissions, and discussion
forum participation. Non-academic variables describe students’ socio-economic and
demographic backgrounds, such as parental income, organizational involvement, gender,
and distance from campus. Digital behavioral variables capture broader engagement
patterns beyond formal coursework, including daily social media usage and participation
in collaborative study groups. This integrated feature structure reflects the multifaceted
nature of student learning in digitally mediated higher education environments.

2.4. Data Preprocessing

Several preprocessing steps were applied to ensure data consistency and reliability prior
to model training. Records containing incomplete or inconsistent information were
removed during data cleaning. Continuous variables were normalized using Min–Max
scaling, while categorical variables were transformed using one-hot encoding. To address
class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied
exclusively to the training data. Specifically, SMOTE was implemented only within the
training folds during cross-validation to avoid information leakage into the test data,
with the number of nearest neighbors set to k = 5. After preprocessing, the dataset was
divided into 80% training data and 20% testing data for final evaluation.

2.5. Machine Learning Algorithms

Seven machine learning classifiers were evaluated in this study: Naïve Bayes, Generalized
Linear Model, Logistic Regression, Deep Learning, Decision Tree, Random Forest, and
Gradient Boosted Trees. These algorithms represent probabilistic, linear, tree-based, and
ensemble learning paradigms commonly used in educational data mining. All models were
implemented using standard configurations provided by the Scikit-learn and [Link]
libraries. No hyperparameter optimization or tuning procedures were applied, allowing a
fair and consistent comparison of baseline model performance across classifiers.

2.6. Deep Learning Configuration

The Deep Learning model was implemented using a multilayer perceptron (MLP)
architecture. The network consisted of an input layer matching the number of features,

Muhammad Arifin, Fajar Nugraha, et al | 228

 Vol. 8, No. 1, February 2026

two hidden layers with 64 and 32 neurons, and a single-node output layer for binary
classification. ReLU activation functions were used in the hidden layers, while a sigmoid
activation function was applied in the output layer. The model was trained using the
Adam optimizer for a maximum of 100 epochs, with early stopping based on validation
loss to reduce the risk of overfitting.

2.7. Evaluation Metrics and Experimental Setup

Model performance was evaluated using accuracy, precision, recall, F1-score, and area
under the ROC curve (AUC). These metrics provide a balanced evaluation of classification
effectiveness, particularly in the presence of class imbalance. All experiments were
conducted in a Python 3.10 environment using an Intel Core i7 workstation with 16 GB
RAM. Ten-fold cross-validation was applied during model training to ensure stable and
reliable performance estimates, while final results were reported on the held-out test
set.

3. RESULTS AND DISCUSSION

This section presents the results of the classification models developed in this study and
discusses their effectiveness in predicting student academic performance. The models
were evaluated using multiple performance metrics to identify the most suitable
approach for early academic risk identification. In line with the study objective, the
classification framework focuses on identifying students at risk of poor academic
performance rather than predicting exact GPA values.

3.1. Model Performance Comparison

Seven machine learning algorithms were employed to classify student academic
performance into two categories: good (GPA ≥ 3.00) and poor (GPA < 3.00). In this study,
the “poor performance” category was treated as the positive class, as correctly identifying
academically at-risk students is critical for early warning and intervention systems.
Accordingly, recall values reported in Table 3 reflect the models’ ability to correctly detect
poor-performing students. Table 3 summarizes the comparative performance of all
models in terms of AUC, accuracy, precision, and recall. All reported metrics represent
the mean values obtained from 10-fold cross-validation, with standard deviations ranging

229 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

between ±0.01 and ±0.03 across models, indicating relatively stable performance across
folds.

The results show that Gradient Boosted Trees (GBT) achieved the highest overall accuracy
(0.75), followed closely by Random Forest and Logistic Regression (0.74). Although Naïve
Bayes produced a slightly lower accuracy (0.73), its high recall value (0.89) indicates strong
sensitivity in identifying students at risk of poor academic performance, which is
particularly desirable in early warning contexts.

Table 3. Performance comparison of classification models

Model AUC Accuracy Precision Recall
Naïve Bayes 0.65 0.73 0.76 0.89
Generalized Linear Model 0.63 0.74 0.76 0.92
Logistic Regression 0.63 0.74 0.76 0.92
Deep Learning 0.61 0.73 0.76 0.90
Decision Tree 0.62 0.74 0.74 1.00
Random Forest 0.62 0.74 0.75 0.95
Gradient Boosted Trees 0.63 0.75 0.75 0.95

The Decision Tree model achieved a perfect recall score (1.00), indicating that it
successfully identified all poor-performing students. However, this result was
accompanied by lower AUC and precision values, suggesting a tendency toward
overfitting and reduced generalization capability. In contrast, ensemble-based methods
such as Random Forest and GBT exhibited more balanced performance across metrics,
highlighting their robustness in handling heterogeneous academic, socio-economic, and
behavioral features. Although AUC values across models are modest (0.61–0.65), this
outcome is expected in educational datasets characterized by overlapping feature
distributions and moderate class imbalance. In early warning applications, recall is often
prioritized over AUC, as failing to identify at-risk students (false negatives) carries greater
institutional consequences than issuing additional alerts. To further examine error
characteristics, a confusion matrix was generated for the best-performing model (GBT).
The results indicate that false negatives were relatively limited compared to false
positives, supporting the suitability of the classification framework for early intervention
scenarios where sensitivity to academic risk is essential.

Muhammad Arifin, Fajar Nugraha, et al | 230

 Vol. 8, No. 1, February 2026

3.2. Feature Importance Analysis

Feature importance analysis was conducted using the Gradient Boosted Trees model to
identify variables that contributed most strongly to classification outcomes. The ranking
of features is presented in Table 4. Social media usage emerged as the most influential
feature (weight = 0.2931), followed by total LMS logins (0.1629) and gender (0.1218). This
finding suggests that digital behavioral indicators are strongly associated with academic
performance categories; however, the results reflect predictive association rather than
causal relationships.

Among academic indicators, LMS login frequency and assignment submissions

contributed notably, reinforcing the importance of sustained engagement in digital
learning environments. Non-academic factors such as domicile distance and
organizational involvement showed moderate influence, while parental income and study
group participation exhibited relatively low importance within this dataset.

Table 4. Feature importance ranking

Attribute Weight
Social Media Usage (N_medsos_acces) 0.2931
Total LMS Logins (Total_login) 0.1629
Gender 0.1218
Assignments Submitted (N_assignments_submitted) 0.0857
Domicile Distance 0.0511
Organizational Involvement 0.0214
LMS Forum Answers (N_answered_questions) 0.0073
Study Group Participation 0.0047
Parental Income (Economy) 0.0027

3.2. Correlation Analysis

Pearson correlation analysis was conducted to examine relationships among predictors
and assess potential multicollinearity. As illustrated in Figure 2, correlations among
variables were generally weak to moderate. Social media usage showed a weak negative
correlation with academic performance classification (r = –0.293), while LMS login
frequency exhibited a weak positive correlation (r = 0.163). The low inter-variable

231 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

correlations indicate minimal multicollinearity, supporting the suitability of the selected

predictors for machine learning–based classification.

Figure 2. Correlation Features

3.3. Discussion
The findings indicate that reframing student performance prediction from regression-
based GPA estimation to a binary classification task produces outputs that are more
interpretable and operationally useful for academic decision-making. Instead of reporting
an exact GPA value—which may be difficult to translate into policy actions—the
classification framework directly identifies whether a student is likely to fall into the
poor (GPA < 3.00) or good (GPA ≥ 3.00) performance category. Importantly, defining poor
performance as the positive class aligns model evaluation with the practical goal of early
risk detection. In this context, recall becomes a priority metric because false negatives
(at-risk students incorrectly classified as not at risk) can delay intervention and
potentially worsen academic outcomes.

Across the seven evaluated algorithms, ensemble-based approaches demonstrated the

most consistent and balanced performance, with Gradient Boosted Trees (GBT) achieving
the highest accuracy (0.75) and strong recall (0.95), followed closely by Random Forest
(accuracy = 0.74; recall = 0.95). These results support prior evidence that ensemble
methods are well suited for heterogeneous educational datasets that combine academic,
socio-economic, and behavioral indicators [27]–[31]. While Logistic Regression and the
Generalized Linear Model also performed competitively (accuracy = 0.74; recall = 0.92),

Muhammad Arifin, Fajar Nugraha, et al | 232

 Vol. 8, No. 1, February 2026

the ensemble models provided a better balance between sensitivity and overall predictive
stability under the same experimental design (10-fold cross-validation with low variability
across folds). By contrast, the single Decision Tree achieved perfect recall (1.00), meaning
it identified all poor-performing students; however, this came with lower precision (0.74)
and modest AUC (0.62), suggesting reduced generalization and potential overfitting. For
early warning systems, such a model may generate more false alarms, which can burden
academic support units and reduce stakeholder trust in the system.

Although the AUC values across models are relatively modest (0.61–0.65), this pattern is
common in educational prediction tasks where feature distributions overlap and
outcomes are influenced by many unobserved factors. In such settings, AUC alone may
understate practical utility, especially when the institutional objective is to maximize
detection of at-risk students. The high recall achieved by most models—including Naïve
Bayes (0.89), Deep Learning (0.90), and particularly the ensemble methods (0.95)—
demonstrates strong sensitivity to academic risk. This supports the suitability of the
proposed framework for early intervention scenarios, where a manageable increase in
false positives is often preferable to missing students who genuinely require support.
Moreover, the use of SMOTE exclusively within training folds helps mitigate the moderate
class imbalance (62.4% good vs 37.6% poor) while reducing the risk of information
leakage, strengthening confidence that the reported performance reflects genuine
predictive signal rather than inflated results.

The feature importance analysis using GBT provides additional insight into which
variables most strongly differentiate performance categories in this dataset. Social media
usage emerged as the most influential predictor (weight = 0.2931), followed by total LMS
logins (0.1629) and gender (0.1218). Together with the correlation results (social media
usage showing a weak negative association with performance, r = –0.293; LMS logins
showing a weak positive association, r = 0.163), these findings suggest that digital
engagement signals can meaningfully complement traditional academic indicators. A
plausible interpretation is that high social media exposure may reflect time displacement,
reduced concentration, or fragmented study routines, whereas sustained LMS activity
may signal consistent academic engagement. However, these patterns should be
interpreted as predictive associations rather than causal relationships. Social media use,
for example, may also be a proxy for other unmeasured factors (stress, motivation, or

233 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

learning habits). Likewise, gender’s relatively high importance warrants careful handling:
it may capture structural or behavioral differences in learning engagement, but it should
not be used to justify biased decision-making. Institutional implementation should
prioritize supportive interventions and avoid stigmatization or differential treatment
based on demographic attributes.

Among the academic variables, assignment submissions (0.0857) and forum participation
(0.0073) contributed less than login frequency, suggesting that broad engagement
intensity (regular access and presence in LMS) may be more informative than specific
activity counts within this dataset. Non-academic factors showed mixed influence:
domicile distance (0.0511) had moderate importance—possibly reflecting commuting
constraints or time availability—while organizational involvement (0.0214) contributed
modestly. Parental income (0.0027) and study group participation (0.0047) appeared least
influential, which may reflect limited measurement granularity (e.g., income recorded in
broad ranges), context-specific characteristics of the sampled institution, or the
possibility that academic behaviors captured through LMS overshadow these variables in
predictive power. Notably, the Pearson correlation matrix indicates generally weak inter-
variable correlations, suggesting minimal multicollinearity and supporting the
appropriateness of combining these predictors within machine learning models without
severe redundancy.

From an institutional perspective, the results reinforce the value of integrating

behavioral and digital engagement features with conventional academic variables to
enhance early warning systems. Rather than treating social media usage or other digital
behaviors as standalone risk markers, institutions should interpret them as
complementary signals that help refine risk screening when combined with learning
engagement indicators (e.g., LMS logins and submissions). Operationally, models such as
GBT and Random Forest are attractive because they offer strong recall with
comparatively balanced precision, which can reduce the likelihood of overwhelming
academic advisors with excessive alerts while still prioritizing at-risk detection.

Ethical and privacy considerations are critical when incorporating digital behavior data
into educational analytics. Institutions should ensure transparency about what data are
collected and why, minimize the use of personally sensitive attributes, and rely on

Muhammad Arifin, Fajar Nugraha, et al | 234

 Vol. 8, No. 1, February 2026

consent-based and appropriately aggregated indicators. Predictive outputs should be

used strictly for supportive academic interventions (e.g., outreach, tutoring, counseling
referrals) rather than punitive actions. In addition, governance mechanisms should be
established to monitor potential bias—especially when demographic variables contribute
meaningfully to predictions—and to ensure fair access to support services for all
students.

With respect to the research questions, RQ1 is addressed by the observed improvement
in predictive capability when combining academic (LMS), socio-economic, and digital
behavioral features, as reflected in consistently high recall values across models and
competitive accuracy levels, particularly for ensemble methods. RQ2 is answered by the
comparative evaluation showing that Gradient Boosted Trees provides the most accurate
and reliable performance overall (accuracy = 0.75; recall = 0.95), closely followed by
Random Forest (accuracy = 0.74; recall = 0.95). Finally, the study’s single-institution scope
remains a limitation for generalizability; future work should validate these findings across
multiple Indonesian universities, consider hyperparameter tuning, and incorporate
interpretability analyses (e.g., local explanations) to better support responsible
deployment in real academic settings.

4. CONCLUSION

This study compared seven machine learning algorithms to classify student academic
performance into good and poor categories, adopting a classification framework to
enhance interpretability beyond regression-based GPA prediction. The results indicate
that ensemble-based models, particularly Gradient Boosted Trees and Random Forest,
achieve the most reliable performance, with the highest accuracy reaching 0.75. RQ1 is
addressed by showing that integrating academic, socio-economic, and digital behavioral
features enables effective prediction of student academic performance, while RQ2 is
answered by identifying Gradient Boosted Trees as the best-performing classifier.

Feature importance analysis highlights digital engagement indicators especially social

media usage and LMS activity as influential predictors, complementing traditional
academic variables, though these relationships reflect predictive associations rather than
causal effects. From an institutional perspective, the proposed classification models can

235 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

support early warning systems by enabling timely identification of students at academic

risk and facilitating targeted academic interventions.

This study is limited by the use of data from a single institution and by moderate AUC
values, which reflect overlapping student performance characteristics. Future research
should explore multi-institutional or longitudinal datasets and richer behavioral
indicators to improve generalizability and predictive robustness. Overall, the findings
demonstrate that classification-oriented modeling provides actionable insights that
effectively support data-driven decision-making in higher education.

ACKNOWLEDGMENT

The authors would like to thank Universitas Muria Kudus for funding this research
through the Internal Research Scheme (PFR). We also extend our gratitude to all parties
who have contributed to the completion of this study.

REFERENCES

[1] A. Alshanqiti and A. Namoun, “Predicting Student Performance and Its Influential
Factors Using Hybrid Regression and Multi-Label Classification,” vol. 8, pp. 203827–
203844, 2020, doi: 10.1109/ACCESS.2020.3036572.
[2] S. M. Dol and P. M. Jawandhiya, “Systematic Review and Analysis of EDM for
Predicting the Academic Performance of Students,” J. Inst. Eng. Ser. B, vol. 105, no.
4, pp. 1021–1071, 2024, doi: 10.1007/s40031-024-00998-0.
[3] Khan, A., and S. K. Ghosh, “Student performance analysis and prediction in classroom
learning: A review of educational data mining studies,” Educ. Inf. Technol., vol. 26,
no. 1, pp. 205–240, 2021.
[4] I. Papadogiannis and M. Wallace, “Educational Data Mining : A Foundational
Overview,” Encyclopedia, vol. 4, no. 4, pp. 1644–1664, 2024.
[5] P. E. O. David Iyanuoluwa Ajiga, Oladimeji Hamza, Adeoluwa Eweje, Eseoghene
Kokogho, “Data-Driven Strategies for Enhancing Student Success in,” vol. 11, no. 1, pp.
411–424, 2025, doi: 10.56201/[Link].411.424.

Muhammad Arifin, Fajar Nugraha, et al | 236

 Vol. 8, No. 1, February 2026

[6] N. P.-M. Esomonu, “Utilizing AI and Big Data for Predictive Insights on Institutional
Performance and Student Success: A Data-Driven Approach to Quality Assurance,”
AI Ethics, Acad. Integr. Futur. Qual. Assur. High. Educ., vol. 29, 2025.
[7] Gonugunta, K. C., and K. Leo, “Role of data-driven decision making in enhancing
higher education performance: A comprehensive analysis of analytics in
institutional management,” Int. J. Acta Informat., vol. 3, no. 1, pp. 149–159, 2024.
[8] Wakeel, S., D. Sher, A. Kauser, and K. B. A. H. K. Niazi, “Investigating how predictive
analytics and student data modeling influence interventions, curriculum design, and
educational policy,” Crit. Rev. Soc. Sci. Stud., vol. 3, no. 2, pp. 1506–1521, 2025.
[9] H. Luo, “Prediction of Student Decision-Making Behaviour based on Machine
Learning Algorithms,” Pakistan J. Life Soc. Sci., vol. 22, no. 2, pp. 16382–16390, 2024.
[10] A. Mahmoud et al., “Policy Reviews in Higher Education Understanding the drivers
of student loan decision-making and its impact on graduation rates in Ghanaian
public universities,” Policy Rev. High. Educ., vol. 8, no. 1, pp. 85–101, 2024, doi:
10.1080/23322969.2024.2358008.
[11] Thelma, C. C., “Student retention in higher learning institutions of Zambia,” Int. J. Res.
Publ. Rev., vol. 5, no. 6, pp. 433–441, 2024.
[12] Gul, M. N., W. Abbasi, M. Z. Babar, A. Aljohani, and M. Arif, “Data driven decisions in
education using a comprehensive machine learning framework for student
performance prediction,” Discover Comput., vol. 28, no. 1, Art. no. 153, 2025.
[13] K. Patil, K. Yesugade, and K. B. Naikwadi, “A Study on Regression Based Machine
Learning Models to Predict the Student Performance,” J. Eng. Educ. Transform., pp.
177–186, 2024, doi: 10.16920/jeet/2024/v38i2/24200.
[14] H. Almaghrabi, B. Soh, and A. Li, “SoK : The Impact of Educational Data Mining on
Organisational Administration,” Information, vol. 15, no. 11, p. 738, 2024.
[15] H. Sahlaoui, E. L. Arbi, A. Alaoui, and M. M. Jaber, “Predicting and Interpreting Student
Performance Using Ensemble Models and Shapley Additive Explanations,” IEEE
Access, vol. 9, pp. 152688–152703, 2021, doi: 10.1109/ACCESS.2021.3124270.
[16] P. Boozary, S. Sheykhan, H. Ghorbantanhaei, and C. Magazzino, “International Journal
of Information Enhancing customer retention with machine learning : A
comparative analysis of ensemble models for accurate churn prediction,” Int. J. Inf.
Manag. Data Insights, vol. 5, no. 1, p. 100331, 2025, doi: 10.1016/[Link].2025.100331.

237 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..
 Vol. 8, No. 1, February 2026

[17] D. J. Lemay, C. Baek, and T. Doleck, “Computers and Education : Arti fi cial Intelligence
Comparison of learning analytics and educational data mining : A topic modeling
approach,” Comput. Educ. Artif. Intell., vol. 2, no. March, p. 100016, 2021, doi:
10.1016/[Link].2021.100016.
[18] M. Motevalli, “Comparative analysis of systematic, scoping, umbrella, and narrative
reviews in clinical research: critical considerations and future directions,” Int. J. Clin.
Pract., vol. 2025, no. 1, Art. no. 9929300, 2025.
[19] R. Awashreh, “Bridging Cultural Gaps : Enhancing Student Motivation and Academic
Integrity in Oman ’ s Universities,” Forum Linguist. Stud., vol. 07, no. 02, pp. 265–279,
2025.
[20] P. Banerjee, “Connecting the dots: A systematic review of explanatory factors
linking contextual indicators, institutional culture and degree awarding gaps,” High.
Educ. Eval. Dev., vol. 18, no. 1, pp. 31–52, 2024, doi: 10.1108/HEED-07-2023-0020.
[21] T. Getaneh, T. Zaw, and K. Jozsa, “Bridging theoretical gaps to improve students’
academic success in higher education in the digital era: A systematic literature
review,” Int. J. Educ. Res. Open, vol. 9, no. April, p. 100510, 2025, doi:
10.1016/[Link].2025.100510.
[22] E. N. Syamiya, S. Lestari, D. Wulandari, and D. Ekawati, “Digital literacy analysis on
online learning outcomes for macroeconomics with gender-mediated and family
socio-economics as moderating variables,” J. Kependidikan, vol. 8, no. 2, pp. 450–459,
2022.
[23] B. Maunah, “Social and cultural capital and learners’ cognitive ability: Issues and
prospects for educational relevance, access and equity towards digital
communication in Indonesia,” J. Soc. Stud. Educ. Res., vol. 11, no. 1, pp. 163–191, 2020.
[24] B. A. Al-sheeb, A. M. Hamouda, and G. M. Abdella, “Modeling of student academic
achievement in engineering education using cognitive and non-cognitive factors,”
vol. 11, no. 2, pp. 178–198, 2025, doi: 10.1108/JARHE-10-2017-0120.
[25] M. Mohzana, “The impact of the new student orientation program on the adaptation
process and academic performance,” Int. J. Educ. Narrat., vol. 2, no. 2, pp. 169–178,
2024.
[26] G. J. Palardy, “School peer non-academic skills and academic performance in high
school,” Front. Educ., vol. 4, p. 57, 2019, doi: 10.3389/feduc.2019.00057.

Muhammad Arifin, Fajar Nugraha, et al | 238

 Vol. 8, No. 1, February 2026

[27] M. Vaarma and H. Li, “Technology in Society Predicting student dropouts with
machine learning : An empirical study in Finnish higher education,” Technol. Soc., vol.
76, no. February, p. 102474, 2024, doi: 10.1016/[Link].2024.102474.
[28] S. Zhao, D. Zhou, H. Wang, and D. Chen, “Enhancing Student Academic Success
Prediction Through Ensemble Learning and Image-Based Behavioral Data
Transformation,” Appl. Sci., vol. 15, no. 3, p. 1231, 2025.
[29] T. K. Shoukath and M. Chakkaravarthy, “Predictive analytics in education: machine
learning approaches and performance metrics for student success—a systematic
literature review,” Data Metadata, vol. 4, Art. no. 730, 2025, doi: 10.56294/dm2025730.
[30] M. J. Shayegan and R. Akhtari, “A Stacking Machine Learning Model for Student
Performance Prediction Based on Class Activities in E-Learning,” Comput. Syst. Sci.
Eng., vol. 48, no. 5, 2024, doi: 10.32604/csse.2024.052587.
[31] T. T. Id and Z. Li, “Predicting learning achievement using ensemble learning with
result explanation,” PLoS One, vol. 20, no. 1, p. e0312124, 2025, doi:
10.1371/[Link].0312124.

239 | Student Performance Classification Using Academic, Socioeconomic, and Digital …..