Bi-Directional Transfer Graph Contrastive Learning for Social Recommendation

Abstract

Graph Neural Networks (GNNs) have emerged as an effective approach for social recommender systems. GNNs excel at capturing the graph-structured semantic information within the collaborative interaction graph and social networks. Recently, some methods have introduced self-supervised learning to GNNs, aiming to enhance recommendation performance by mitigating the data sparsity issue. However, these methods treat the interaction graph and social network as separate entities, which severely limits the number of samples available for self-supervised learning. Moreover, these separated methods also exacerbate the problem of information islands in collaborative and social domains, resulting in suboptimal performance. To tackle these challenges, we propose an innovative self-supervised social recommendation method called Bi-directional Transfer Graph Contrastive Learning (BTGCL). BTGCL jointly encodes node representations within both collaborative domain and social domain, then generates node views through feature augmentation. To bridge the information gap between domains, we devise a bi-directional migration mechanism that aligns features from the collaborative and social domains of the same positive pair. Through extensive experiments conducted on three publicly available datasets, we demonstrate the effectiveness of our proposed method in enhancing social recommendation performance.