Heterogeneous collaborative filtering contrastive learning for social recommendation

Abstract

Collaborative filtering methods based on Graph Neural Networks (GNNs) have gained increasing popularity in recommendation systems. These methods enhance the representation of users and items by leveraging the information of graph structure from interaction data, improving recommendation performance. However, they often face limitations due to the data sparsity issue that is common in recommendation systems. In the constructed user–item heterogeneous bipartite graph, sparse interaction data leads to a scarcity of neighbor nodes impeding the acquisition of sufficient collaborative signals via the message-passing mechanism among these neighbor nodes. We have observed that users and items can be grouped according to characteristic similarities. These groups’ common feature information can serve as supplementary data to aid in the embedding learning. So, we present the Heterogeneous Collaborative Filtering Contrastive Learning (HCFCL) method, which aims to extract two types of heterogeneous collaborative signals from interaction data: those based on neighbor nodes and those based on group features. Specifically, we design an embedding generative hypergraph network to extract group common feature information founded on the heterogeneous bipartite graph. The group common feature information is transferred via a meta network and personalized bridge functions according to individual characteristics. Additionally, the HCFCL model, combined with contrastive learning, captures the consistency of the heterogeneous collaborative signals to enhance representation. The experiment demonstrates the superior performance of the HCFCL model compared to other methods evaluated on three public datasets, demonstrating excellent and stable performance in mitigating the data sparsity issue.