Subgraph Collaborative Graph Contrastive Learning for Recommendation

Graph Collaborative Filtering discovers potential connections between users and items using graph neural networks. However, graph neural networks may aggregate the noise in the user-item interaction data when updating node features. In addition, increasing the number of network layers may lead to the over-smoothing problem, whereby the representations of distinct nodes become excessively similar. We propose a new Subgraph Collaborative Graph Contrastive Learning framework (SGCL) to mitigate these issues. The SGCL model mainly consists of a self-alignment module and a subgraph module. The self-alignment module uses two encoders to create contrasting views, optimizing alignment and uniformity to improve user and item representations. The subgraph module consists of three parts: Initially, we select neighbors that most accurately represent each node for subgraph sampling, aiming to lessen the impact of high neighbors. This method helps alleviate the effect of over-smoothing. Furthermore, we filter out noisy nodes in the frequency domain to achieve the subgraph denoising. Lastly, we apply subgraph collaborative representation to enhance the node representations. The SGCL mitigates noisy nodes and over-smoothing issues. Our experiments across three public datasets show the SGCL performance improvement of about 7.8% .