A denoising social recommendation method by fusing global-local node information of heterogeneous graphs

Presently, the integration of social networks into recommendation systems has emerged as a focal point in research and has found widespread applications in understanding user preferences. However, existing studies primarily focus on the influence of local information on recommendation effects, while the consideration of global information is relatively insufficient, which potentially further influences the topological structure of social communities. To address this challenge, we propose a Denoising Social Recommendation method by fusing Global-Local node information (GLDSR). First, user social relations and item association relations are combined to form a multi-assisted information graph. Simultaneously, user-item interaction relationships are integrated into the same heterogeneous graph, facilitating the propagation of multiple pieces of information within a unified framework. Subsequently, a denoising pre-training strategy is employed to filter out irrelevant features of nodes in the graph structure, and distinct encoders are devised for nodes based on heterogeneous relationships and global nodes within the graph. Finally, a relation aggregation network is deployed to capture relationship features. The effectiveness of the proposed model is evaluated on three real-world datasets, namely Ciao, Epinions, and Yelp. Experimental results demonstrate notable enhancements in hit rate (HR) and normalized discounted cumulative gain (NDCG) metrics.