KARM: Knowledge-Aware Temporal Gating with Multi-Head Attention for Course Recommendation

In recent years, the rapid development of Massive Open Online Courses (MOOCs) has broken the temporal and spatial constraints of traditional education, providing learners with abundant course resources. However, the explosive growth of course resources has led to severe information overload issues for learners. Consequently, achieving accurate course recommendation has become a core research focus. Although graph diffusion-based algorithms inherently suit exploring complex course relationships, existing methods still have limitations: (1) Failing to adequately capture the fine-grained influence of multiple course attributes on learner preferences; (2) Overlooking temporal dependencies in learning behaviors, making dynamic interest evolution difficult to capture. To address these challenges, we propose KARM (Knowledge-Aware Recurrent Multi-attention Network), a course recommendation model integrating curriculum graph diffusion with spatiotemporal awareness. The model employs a relation-aware multi-head attention mechanism to explicitly quantify users' attention weights across multiple course attributes. Additionally, to tackle data sparsity and temporal dependency issues in learning behaviors, we design a bi-directional GRU temporal enhancement module that effectively distinguishes between long-term stable preferences and short-term dynamic demands. Experimental results on real-world datasets demonstrate that KARM outperforms existing baseline methods.