Hypergraph Enhanced Contrastive Learning for News Recommendation

Abstract

With the explosion of news information, user interest modeling plays an important role in personalized news recommendation. Many existing methods usually learn user representations from historically clicked news articles to represent their overall interest. However, they neglect the diverse user intents when interacting with items, which can model accurate user interest. Moreover, GNN methods based on bipartite graph cause the over-smoothing effect when considering high-order connectivity, which declines the news recommendation quality. To tackle the above issue, we propose a novel Hypergraph Enhanced Contrastive Learning model, named HGCL, to incorporate the intent representation and the hypergraph representation with a cross-view contrastive learning architecture. Specifically, we design an intent interaction learning module, which explores user intents of each user-item interaction at a fine-grained topic level and encodes useful information into the representations of users and items. Meanwhile, the designed hypergraph structure learning module enhances the discrimination ability and enriches the complex high-order dependencies, which improves the presentation quality of the recommendation system based on hypergraph enhanced contrastive learning. Extensive experiments on two real-world datasets demonstrate the superiority of our model over various state-of-the-art news recommendation methods.