KDG-Rec: Enhanced Dual-GNN Programming Exercise Recommendation via LLM-Powered Knowledge Annotation and Preference-Decoupling

The rapid expansion of online programming exercise platforms has brought abundant learning resources for programming education, but also presents challenges for personalized exercise recommendation due to missing or imprecise knowledge annotations and the sparsity of learner-exercise interaction records, which together lead to reduced accuracy and a lack of explainability in the recommendation results. Existing natural language processing and large language models (LLM)-based annotation methods struggle to capture implicit knowledge and often generate redundant or inconsistent results. Moreover, mainstream recommendation systems are ineffective at handling the severe sparsity of interaction sequences in programming exercise datasets, and typically model student preferences in a single dimension—overlooking key educational factors such as knowledge gaps, difficulty tolerance, and preferred learning rhythm. To address these challenges, we propose KDG-Rec, a novel framework that first introduces AgentCo-KAS, a multiagent LLM-based collaborative annotation method with role-specific fine-tuning and cross-agent verification for high-precision, fine-grained knowledge annotation. Building on these enriched annotations, we develop a disentangled graph neural network model that constructs dual exercise-interaction graphs to effectively capture learning patterns at multiple granularities in interaction sequences and explicitly decouples student preferences into four interpretable dimensions for adaptive fusion. Extensive experiments on real-world datasets demonstrate that KDG-Rec outperforms nine state-of-the-art methods across multiple metrics, significantly advancing personalized programming exercise recommendation.