Dual-view Enhanced Knowledge Contrastive Learning for Recommendation

Weijun Xu, Fan Gao, Jinsong Su, Qingqiang Wu, Meihong Wang

Abstract: To alleviate the problem of data sparsity in traditional recommender systems, knowledge graphs (KGs) serve as side information to provide rich semantic information. However, existing KG-based recommendations have the following two issues. Knowledge noise: KG contains many recommendation-irrelevant relations, which would be amplified during representation learning. Knowledge sparsity: the number of triplets corresponding to some entities is extremely sparse, leading to detrimental learning of semantic information. Inspired by this, we propose a novel Dual-view Enhanced Knowledge Contrastive Learning (DKCL) to address the two problems. Specifically, we first devise a relation-aware attention mechanism to learn the relation scores and discard a certain percentage of relations with lower scores, which effectively suppresses the knowledge noise introduced by task-irrelevant relations. Secondly, we utilize clustering algorithm to generate semantic entities and replace the original entities with them. In this way, each semantic entity corresponds to a larger number of triplets than the original entities. Finally, we develop a cross-view contrastive learning pattern to bridge the knowledge semantic signals with the collaborative signals, which enhances the item representations in the user-item graph with the denoised semantic information. Extensive experiments on three real-world datasets demonstrate the superiority of our proposed model compared to state-of-the-art methods.