Robust Graph Contrastive Learning for recommender systems: Addressing data sparsity and noise

Graph Contrastive Learning (GCL) enhances recommender systems by leveraging Graph Neural Networks (GNNs) and self-supervised learning (SSL). However, existing methods struggle with data sparsity and noise. We propose Robust Graph Contrastive Learning (RoGCL), a novel framework that generates high-quality contrastive views through dual-perspective generators. The local generator employs Variational Graph Autoencoders (VGAE) to capture micro-level collaborative patterns by sampling from user–item interaction distributions. The global generator utilizes Singular Value Decomposition (SVD) to reconstruct macro-level structures while filtering noise through low-rank approximation. By incorporating Information Bottleneck (InfoBN) to minimize redundancy between views, RoGCL learns robust representations combining local and global collaborative signals. Extensive experiments on Last.FM, Yelp, and BeerAdvocate datasets demonstrate that RoGCL significantly outperforms state-of-the-art methods including Self-supervised Graph Learning (SGL), Neural Collaborative Learning (NCL), and Adaptive Graph Contrastive Learning (AdaGCL). Results show improved Recall@20 by up to 8.7% and NDCG@20 by 5.8% compared to best baselines. Notably, RoGCL exhibits exceptional robustness, maintaining over 90% relative performance with 25% noise injection and showing 37.7% improvement for sparse user groups, making it particularly suitable for real-world applications with imperfect data.