Heterogeneous graph structural and semantics contrastive recommendation learning

Graph Neural Networks (GNN) have emerged as a potent tool for modeling the structural information of graph in recommendation systems. However, in real-world recommendation scenarios, there are a large number of heterogeneous relationships (e.g., situational information, user social relationships, product attributes, etc.), which harbor rich information to help us model users. In this paper, we delve into the challenge of enhancing relational learning for recommendation within the realm of heterogeneous graphs. In light of the success of contrastive self-supervised learning, we propose a Heterogeneous Graph Structural and Semantics Contrastive Recommendation Learning (HSSCL), which seamlessly incorporates heterogeneous relational semantics into the modeling of user-item interactions, utilizing contrastive learning to enhance the model's ability to discern between various structural and semantic features. This facilitates effective knowledge transfer across different perspectives. The introduction of meta-networks enhances the learning process, enabling personalized knowledge transfer and improving the model's adaptability. Our extensive experiments on mainstream recommendation datasets (Yelp, Epinions, Ciao) reveal the substantial performance improvement achieved by the HSSCL framework compared to other robust baselines. Through an ablation analysis, we validate the effectiveness of three key components within HSSCL that contribute to the enhancement of recommendation performance