Contrastive learning on high-order noisy graphs for collaborative recommendation

The graph-based collaborative filtering method has shown significant application value in recommendation systems, as it models user-item preferences by constructing a user-item interaction graph. However, existing methods face challenges related to data sparsity in practical applications. Although some studies have enhanced the performance of graph-based collaborative filtering by introducing contrastive learning mechanisms, current solutions still face two main limitations: (1) does not effectively capture higher-order or indirect user-item associations, which are critical for recommendations in sparse scenarios, and (2) does not robustly handle user feedback or noise in the graph structure, which may degrade model performance. To address this gap, we propose RHO-GCL, a novel framework that explicitly models higher-order graph structures to capture richer user-item relations, and integrates noise-enhanced contrastive learning to improve robustness against noisy interactions. Unlike previous methods, which rely on shallow neighborhood aggregation or basic contrastive setups, RHO-GCL introduces a hierarchical perspective and noise perturbation mechanisms that better align with the intrinsic properties of recommendation graphs. In the experimental validation phase, we conducted systematic tests on four benchmark datasets, including MovieLens and Yelp. Compared to existing graph-based contrastive learning collaborative filtering models, RHO-GCL showed significant performance improvements and demonstrated stronger resistance to noise interference. These results indicate that the proposed structural feature enhancement strategy and noise-balancing mechanism effectively enhance the recommendation system's ability to understand complex graph data, providing a reliable technical solution for real-world application scenarios.