A self-supervised graph-learning method for reliable-relation identification in social recommendation

As social commerce continues to develop, personalized recommender systems increasingly leverage social networks to understand consumer interests and aid in consumption decisions. Despite the widespread use of graph neural networks in existing social recommendation methods, a significant challenge persists: social relationships do not always represent similar interests between users. This reduces the efficacy of social recommendation methods. To address this challenge, we propose a novel self-supervised graph-learning method for reliable-relation identification in social recommendation (SGRI). In SGRI, the primary social recommendation task uses graph neural networks to learn social influence and collaborative interests from the social and interaction graphs, respectively. An auxiliary self-supervised learning task aims to identify reliable relations in these graphs, thereby enhancing the primary task's performance. This auxiliary task employs an adaptive data-augmentation strategy based on user-item-friend triadic relations to generate diverse graph views, providing users and items with credible neighborhoods. Subsequently, a local-local contrastive pretext method is used for the node self-discrimination across different graph views, and a local-context contrastive pretext method ensures interest similarity between users and their social circles. Experimental results show that our proposed SGRI method consistently outperforms the state-of-the-art methods on three real-world datasets.