Graph-based Knowledge Tracing for Personalized MOOC Path Recommendation

Abstract

Massive Open Online Courses (MOOCs) expose learners to thousands of videos, exercises, and discussions, yet most platforms still rely on one-size-fits-all curricula. Two lines of learning analytics research address this gap: knowledge tracing (KT), which estimates a learner’s evolving mastery from interaction logs, and learning path recommendation, which selects the next learning activity. However, many KT models treat a course as a flat sequence of interactions and ignore heterogeneous relations among students, learning objects, and knowledge components; conversely, many path recommenders operate on coarse resource graphs without fine-grained mastery tracking.

This paper proposes GKTPR, a unified framework for personalized MOOC-style path recommendation built on graph-based knowledge tracing. We construct a heterogeneous learning graph with node types {course, unit, concept, exercise, student} and typed relations such as unit–exercise inclusion, exercise–concept tagging, and student–exercise attempts. A relational graph neural network (R-GCN) encodes global structural signals into graph-aware embeddings, while a causal Transformer performs sequence modeling for KT using graph-aware tokens. The model outputs (i) next-response probabilities for KT and (ii) a per-concept mastery vector used by a mastery-gain objective to rank candidate next activities. We conduct full offline experimental evaluations on two public large-scale online learning benchmarks commonly used for knowledge tracing: ASSISTments 2012 and KDD Cup 2010 (Bridge to Algebra). Following the protocol in Section 2, we report empirically measured results for next-response prediction (AUC, ACC, NLL), ablation and robustness analyses, and offline learning-path recommendation metrics (Hit@K, NDCG@K, and predicted mastery gain) for GKTPR and strong baselines (DKT, DKVMN, SAKT, AKT, and GKT). Across both datasets, GKTPR consistently improves both prediction and recommendation quality, indicating that fusing heterogeneous graph structure with long-range sequential attention yields more accurate mastery estimation and more effective next-activity ranking.