Knowledge Graph-Enhanced Student Risk Prediction in Educational Data Mining

Abstract

Early identification of academically at-risk students in higher education enables targeted pedagogical interventions and attrition mitigation. While traditional Educational Data Mining (EDM) techniques leverage temporal clickstream data and demographic features from Learning Management Systems (LMS), interaction logs are commonly treated as isolated numerical counts without accounting for the topological organization of course materials. This study presents KG-Predict, a multimodal predictive framework for early student risk detection. KG-Predict constructs a Course Knowledge Graph connecting course modules, assessment entities, and Virtual Learning Environment (VLE) resource types, extracting topological descriptors that are combined with early clickstream records and assessment submission trajectories from the Open University Learning Analytics Dataset (OULAD; N = 32,593). Empirical evaluations indicate that augmenting behavioral logs with structural graph descriptors and initial assessment scores improves predictive performance, achieving an AUC-ROC of 0.8472 and an accuracy of 76.67% at Week 6 of an academic term, compared to 0.7902 AUC-ROC and 71.25% accuracy for the baseline machine learning model. Module-level evaluation indicates consistent performance across diverse curricular disciplines, achieving AUC-ROC scores between 0.6886 and 0.8790. This framework provides an observable, structured basis for early risk tracking in academic advisory dashboards.