Self-Supervised Meta-Path-Based Heterogeneous Graph Embedding for Recommender Systems

In recent years, significant progress has been made in research on recommender systems based on Heterogeneous Information Networks (HIN). However, most existing HIN-based recommender system methods rely on meta-path-based embedding models, which do not fully exploit the inherent homogeneous and contrastive difference information present in heterogeneous networks. In this study, we propose a method for learning representations by constructing graphs using both heterogeneous graph structures and homogeneous similarity graphs. We then apply contrastive loss to obtain embeddings that capture the differences between these two types of graphs. Our proposed recommendation method combines two perspectives through data embedding of the two types of graphs to train Graph Neural Networks (GNN). Specifically, we find that using primitive paths is the most effective way to directly embed heterogeneous graphs. Integrating and supplementing information rationally through realistic logic also makes sense. Additionally, supplementing data through similarity analogies between viewing sequences and users themselves is also meaningful. Through a twoview neighborhood selection process of logical relations and established facts, experiments show that our approach can improve HIN-based recommendation models.