Social Recommendation Based on Preference-Aware Denoising Graph Convolutional Networks

Abstract

The issue of sparsity in user-item interaction data is commonly encountered in the generation of collaborative filtering recommendations. To address this issue, social recommendations introduce users' social relationships. However, several social recommendation systems based on Graph Neural Network (GNN) cannot aggregate high-order neighbor information according to user preferences during the message-passing process, which leads to embedding oversmoothing and noise. To solve these problems, this study proposes a social recommendation model based on preference-aware denoising Graph Convolution Network(GCN), known as PD-GCN. This model employs unsupervised learning to allocate users with similar preferences to social and user-item interaction subgraphs. Higher-order graph convolution operations are performed within these subgraphs to mitigate the oversmoothing problems observed in existing models. Considering global and local perspectives, the model identifies and removes noisy nodes by weighing the feature similarity among nodes with the same preferences and diversity of neighborhood node preference distributions, thereby enhancing the robustness of the model against noise in user-item interaction and social relationships. Experimental results on three public datasets (LastFM, Ciao, and Yelp) show that the proposed model performs better than other mainstream models in terms of Recall and Normalized Discounted Cumulative Gain (NDCG), verifying the effectiveness of the PD-GCN model.