Personalized Course Recommendation Based on Attribute-Interaction Joint Encoding and Hypergraph Reconstruction

Course recommendation systems based on deep learning have demonstrated powerful feature extraction capabilities in dealing with information overload in massive open online courses (MOOCs), and have become an irreplaceable mainstream method. However, the learner–course interactions are usually scarce in reality, which limits the representation power of course recommendation. In addition, the contribution of learner and course attribute information to course recommendation has not been sufficiently explored by most existing methods. To tackle these challenges, a personalized course recommendation model based on attribute-interaction joint encoding and hypergraph reconstruction (AIHR-PCRM) is proposed in this paper. Specifically, a course hypergraph reconstruction (CHR) method is designed to construct higher-order associations for each course to explore more reliable global collaboration signals. Unlike existing hypergraph constructions that directly take learners as hyperedges, CHR explicitly couples three steps, including invalid learner elimination, high-order reachability induction, and similarity-based hyperedge filtering, to substantially raise the signal-to-noise ratio of the resulting hypergraph. Based on this, a hypergraph global collaborative learning module (HGM) can alleviate the issue of data sparsity. Then, a joint encoding module (JEM) is utilized to enhance learner behavior sequence representations by simultaneously fusing hypergraph-level global signals with attribute-level local semantics. Finally, a bidirectional self-attention module (BSM) is introduced to blend the contextual information of the learner behavior sequence, and to further provide a recommendation. Experimental results on three real-world datasets revealed that the proposed model has already achieved the best recall and ndcg scores compared to those of several existing models.