Enhancing Collaborative Filtering via Structural Denoised and Scalable Graph

Collaborative filtering (CF) based on graph neural networks (GNN) has demonstrated excellent performance in recommender systems and has been widely adopted. However, real-world interaction graphs often contain noise, which significantly affects the effectiveness and robustness of GNN models. While existing studies have addressed data denoising, they either employ complex strategies or only focus on denoising semantic neighbors, failing to directly address the noise issue. To address this problem, this paper proposes SDGCF, a novel robustness-enhanced structural denoised CF model. SDGCF consists of two key modules: a graph denoising module and an enhanced graph fusion module. The graph denoising module removes noisy interactions by assessing structural similarity between users and items, creating two denoised enhanced graphs. These graphs, retaining high structural similarity interactions, improve message passing and model robustness. The enhanced graph fusion module facilitates information sharing during message passing, leveraging graph diversity for better generalization. By utilizing mutual information maximization (MIM) and the InfoNCE loss function, SDGCF ensures consistency across the denoised graphs, further reducing noise impact. Our extensive experiments on three real-world datasets demonstrate that SDGCF consistently outperforms existing baseline methods, confirming its superior performance in handling noisy interactions and providing accurate recommendations.