Abstract
Student attrition remains a persistent challenge in higher education, with institutions reporting substantial dropout rates and many at-risk students failing to receive timely intervention. Existing learning analytics models predominantly rely on click frequency and interaction volume where models count interactions but do not adequately interpret the quality, timing, and diversity of engagement. This paper uses the Resource Diversity Index (RDI), a novel behavioural engagement feature that measures how broadly students distribute their interactions across different Virtual Learning Environment (VLE) activity types. Using the Open University Learning Analytics Dataset (OULAD), which contains records for 32,593 students across 22 courses, RDI was evaluated alongside Engagement, Regularity, and Anti-Procrastination features within an Extreme Gradient Boosting (XGBoost) classification framework using stratified five-fold cross-validation. Preliminary results show that adding RDI improved Accuracy from 0.8303 to 0.8357, AUC-ROC from 0.9014 to 0.9068, and F1-score from 0.7852 to 0.7918. These pilot findings indicate that the click-distribution diversity captured by RDI provides meaningful predictive information beyond interaction volume alone and offers a promising direction for more interpretable and proactive academic risk prediction.
Keywords
Academic risk prediction, learning analytics, resource diversity index, behavioural engagement, XGBoost, explainable AI.
