HGCL: A Hypergraph Contrastive Learning Framework with Graph Structure Learning for Recommendation

Yuma Dose, Shuichiro Haruta, Yihong Zhang, Takahiro Hara

Abstract: In this paper, to address the insufficient capturing of high-order correlations and the vulnerability to noise in conventional collaborative filtering models, we propose HyperGraph Contrastive Learning with graph structure learning for recommendation (HGCL). HGCL employs a user hypergraph and an item hypergraph as contrastive views in contrastive learning, whose hyperedges connect the item and user's k-order reachable neighbors. This allows HGCL to model adjacent relationships with distant nodes and explicitly capture high-order correlations, alleviating over-smoothing. Furthermore, to address the weakness of hypergraphs against noise, we modify the hypergraph structure learning for recommendation and incorporate it into HGCL. Based on node and hyperedge representations, HGCL detects potential and noisy relationships in hypergraphs for recommendation. By adding or removing these relationships from the user hypergraph and the item hypergraph, we acquire the improved user and item hypergraphs with reduced noise effects. Finally, HGCL performs contrastive learning between representations obtained from the user-item interaction graph and the hypergraphs. This integrates the user-item interaction features from the interaction graph with the high-order correlations from the user hypergraph and the item hypergraph, resulting in more information-rich representations. Experimental results on real-world datasets demonstrate that HGCL outperforms baseline models, achieving up to 5.25% improvement in NDCG@20.

External IDs:dblp:journals/jip/DoseHZH25