Early Prediction of Academic Performance in a Database Course Using Moodle Interaction Logs

Abstract

Academic failure and high dropout rates remain persistent challenges in STEM education, particularly in core disciplines like Computer Science. Using Moodle interaction logs for a compulsory course on Database Systems (183,917 log records from 131 students), we engineered 16 behavioral features that capture engagement, regularity, resource diversity, and temporal activity patterns. Four classification models were trained and evaluated using 5-fold cross-validation and an independent test set: Support Vector Machine, Random Forest, k-Nearest Neighbors, and Logistic Regression. The best-performing models reached an AUC of up to 0.812 (Logistic Regression) and 0.809 (Random Forest) on the independent test set. Notably, Random Forest demonstrated a recall of 81.2% for at-risk students, suggesting its potential suitability as the basis for an Early Warning System (EWS). Feature importance analysis reveals that resource access, total engagement, and formative quiz interaction are the strongest predictors of the final outcome. Furthermore, activity within the first four weeks ranks among the top six predictors, providing empirical evidence that early behavioral signals carry significant predictive power. These results suggest that identification of at-risk students is highly feasible and actionable within the first month of instruction, offering a robust foundation for automated pedagogical interventions.