Relationship-aware contrastive learning for social recommendations

The popularity bias is an outstanding challenge in recommendation systems. Prevalent work based on contrastive learning (CL) alleviates this issue but neglects the relationship among data, which limits the ability of CL and leads to a loss of personalized features of users/items, and thus degrades the performance of the recommendation systems. In this paper, we propose a modelagnostic contrastive learning framework, called Relationship-aware Contrastive Learning (ReACL), to make recommendations to users. ReACL framework utilizes the relationship homophily among data to achieve a uniform distribution for node and align the relevant features to preserve personalized features and avoids the problem of popularity bias. We first combine a graph of user-item interaction with that of a social network to mine views on the relationships of users on social networks as well as those of item commonality. We then apply these two views to construct contrastive learning pairs on the user and item sides, respectively. We design two sample selectors and an augmented contrastive loss function to preserve personalized features. Furthermore, we jointly optimize the tasks of supervised learning and contrastive learning. Finally, we conduct extensive experiments on four public datasets, and the results show that the proposed ReACL framework outperforms state-of-the-art algorithms.