Self-supervised graph transformer networks for social recommendation

Abstract

Social recommendation systems often use graph data to represent users, items, and their interactions. Graph Neural Networks (GNNs) are effective at analyzing the complex relationships among nodes. However, traditional GNN models tend to focus only on immediate neighbors during information propagation, limiting their ability to capture global information. To address this limitation, we propose a Self-Supervised Graph Transformer Network (SGTN) for social recommendation. SGTN applies the Transformer to process graph data, using multi-head attention mechanisms for global node information exchange. It also includes edge feature pipeline to fully utilize edge information in social networks, enhancing the model’s understanding of user preferences. Additionally, multi-head attention makes the learned representation multi-view. SGTN uses different user representations generated from user–item interactions and user–user relationships for contrastive learning, effectively integrating these two sources of information for more accurate user representations. Extensive experiments on two real-world datasets demonstrate the effectiveness of SGTN.