Comparative Evaluation of Machine Learning Models for Early Prediction of Student Academic Performance Using the Open University Learning Analytics Dataset

Assel Omarbekova, Aizhan Nazyrova, Gulmira Bekmanova, Zhanar Lamasheva, Rakhila Turebayeva, Zhainagul Abzal

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3991/ijim.v20i17.62574
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Early identification of students at risk of academic failure is a central problem in learning analytics, yet the multi-table, behaviorally heterogeneous, and class-imbalanced nature of educational data complicates reliable prediction. This study develops and comparatively evaluates four supervised machine learning models—logistic regression, support vector machine (SVM) with a radial basis function kernel, random forest, and extreme gradient boosting (XGBoost)—for multi-class prediction of final student outcome on the Open University Learning Analytics Dataset (OULAD), comprising 32,593 anonymized student records and more than ten million virtual-learning-environment interaction logs. A reproducible, leakage-controlled feature-engineering pipeline is proposed: final examination records are excluded, and virtual-learning-environment activity is restricted to the first 120 days of each presentation, yielding 85 demographics, assessment, and behavioral features. All models are trained on an identical stratified 80/20 split and assessed within a unified evaluation framework using accuracy, weighted and macro F1-score, balanced accuracy, per-class recall, and stratified cross-validation. XGBoost achieved the strongest overall performance (accuracy 0.7302, macro F1 0.6658, balanced accuracy 0.6555, and cross-validated accuracy 0.7314), followed closely by random forest and logistic regression, while SVM, despite the weakest aggregate scores, produced the highest recall for the practically critical Fail class (0.4026). The analysis demonstrates that aggregate accuracy substantially overstates the usefulness of these models for risk detection and that the fail class remains the principal bottleneck. A deployed Streamlit prototype illustrates operationalization. The contribution lies in the leakage-aware early-window pipeline and the imbalance-sensitive, algorithm-aware comparison that together expose where predictive value is gained and lost.