Differential Weighting and Flexible Residual GCN-Based Contrastive Learning for Recommendation

The recommendation system based on graphs aims to infer the symmetrical relationship between unconnected users and items nodes. Graph convolutional neural networks (GCNs) are powerful deep learning models widely used in recommender systems, showcasing outstanding performance. However, existing GCN-based recommendation models still suffer from the well-known issue of over-smoothing, which remains a significant obstacle to improve the recommendation performance. Additionally, traditional neighborhood aggregation methods of GCN-based recommendation models do not differentiate the nodes’ importance and also exert a certain negative impact on the recommendation effect of the model. To address these problems, we first propose a simple yet efficient GCN-based recommendation model, named WR-GCN, with a node-based dynamic weighting method and a flexible residual strategy. WR-GCN can effectively alleviate the over-smoothing issue and utilize the interaction information among the graph nodes, enhancing the recommendation performance. Furthermore, building upon the outstanding performance of contrastive learning (CL) in recommendation systems and its robust capability to address data sparsity issues, we integrate the proposed WR-GCN into a simple CL framework to form a more potent recommendation model, WR-GCL, which incorporates an initial embedding controlling method to strike a balanced state of high-frequency information. We have conducted extensive experiments on the proposed WR-GCN and WR-GCL models on multiple datasets. The experimental results show that WR-GCN and WR-GCL outperform several state-of-the-art baselines.