Self-Supervised Signed Graph Attention Network for Social Recommendation

In recent times, social recommendation has become a popular technique in recommender systems due to its ability to enhance the accuracy of recommendations by leveraging the social relationships among users. Despite its widespread use, the prevalent social recommendation methods are often marred by sparsity and noise issues that negatively impact their practicality. Additionally, these methods fail to consider complex user interactions, which could potentially provide additional information. To address these limitations, this paper proposes a novel self-supervised signed graph attention network (SSAN) that incorporates user attitudes in the construction of higher-order user relations. This approach integrates signed networks in the formation of reasonable higher-order local neighborhood relations and aggregates user interests in different social relations through graph convolution and balance theory. Furthermore, two self-supervised signals, derived from social theory, are designed and incorporated into the recommendation framework to better exploit the rich structural and semantic information in social relationship graphs. Empirical results on three publicly available datasets demonstrate that SSAN outperforms existing state-of-the-art social recommendation methods.