PREDICTING STUDENT ACADEMIC RISK IN MOODLE LMS: A COMPARATIVE MACHINE LEARNING APPROACH TO EDUCATIONAL DATA MINING

Abstract

While Moodle-integrated monitoring dashboards can present instructors with descriptive indicators of student performance, they remain fundamentally retrospective: they report what has already happened rather than forecasting what is about to happen. This study addresses that limitation by developing and comparing five supervised machine learning classifiers — Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and Extreme Gradient Boosting (XGBoost) — for the early prediction of academic risk from Moodle log and gradebook data. A dataset of 1,248 student-course records spanning eight courses over three academic semesters at a technical university in Uzbekistan was extracted via the Moodle Web Services API and engineered into 22 behavioral and performance features, including login regularity, resource-access latency, quiz-attempt patterns, and assignment-submission timeliness. Models were trained and validated using stratified 10-fold cross-validation and evaluated with accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). XGBoost achieved the strongest overall performance (accuracy = 90.6%, F1 = 0.887, AUC = 0.934), outperforming Random Forest (AUC = 0.921) and substantially outperforming Logistic Regression (AUC = 0.842). Critically, when restricted to data available during only the first four weeks of the semester, the ensemble models still achieved an AUC above 0.85, indicating that meaningful risk prediction is possible well before traditional grading milestones. SHAP-based feature-importance analysis identified early quiz performance, resource-access regularity, and assignment-submission delay as the three most influential predictors across models. These findings demonstrate that ensemble tree-based methods applied to routinely collected Moodle log data can provide a computationally inexpensive yet accurate early-warning mechanism, complementing rule-based monitoring tools with genuine predictive capability.