Dual variational graph contrastive learning for social recommendation

As an emerging paradigm that merges collaborative filtering with social networking, social recommender systems endeavor to integrate additional social relationships to mitigate data sparsity issues. However, the available social data for training often remain sparse and contain noise. To tackle this, recent studies have leveraged contrastive learning methods to derive extra self-supervised signals. Despite the potential, existing approaches are limited by the cumbersome selection of augmentations and the ambiguous definition of positive pairs. To overcome these limitations, we propose an innovative framework, dual variational graph contrastive learning (DVGCL), tailored for social recommendation. Specifically, we utilize a dual variational graph autoencoder as view generator, which captures variational distributions of user preferences to exploit more underlying collaborative information during graph reconstruction. Additionally, we implement a socially aware light graph convolution network as our backbone to obtain contextual embeddings. Finally, we develop a contrastive loss function based on diverse positive instances to refine the learning of robust representations. By integrating social friends and interacted neighbors, DVGCL provides auxiliary training signals for collaborative filtering-based recommendation tasks. Extensive evaluations across three real-world datasets demonstrate that DVGCL surpasses numerous cutting-edge recommendation methods.