SocialCU: Integrating Commonalities and Uniqueness of Users and Items for Social Recommendation

Abstract

Abstract Social recommendation (SR) based on Graph Neural Networks (GNN) presents a promising avenue to significantly improve user experience by leveraging historical behavior and social data, which benefits from capturing user preferences through higher-order relationships. Although two socially connected users will prefer certain specific items, their preferences in other items are likely to be inconsistent. We argue that current GNN-based social recommendation methods only focus on the commonalities of user preferences, but ignore the uniqueness. In addition, GNN also suffers from the data sparsity problem commonly observed in recommender system. To address these limitations, we propose the Integrating Commonalities and Uniqueness of users and items method, namely SocialCU, which combines GNN and contrastive learning to gain commonalities and uniqueness for SR. To be specific, we firstly model the original data as the user-item interaction graph and user-user social graph and use GNN to obtain the commonalities of nodes (users or items). Then, we design the adaptive data augmentation to build dual contrastive learning to refine the uniqueness of nodes and mitigate data sparsity by extracting supervised signals. We have conducted extensive experiments on three real-world datasets to demonstrate the performance advantages of SocialCU over current state-of-the-art recommendation methods and the rationality of the model design.