LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation

• DOI: 10.48550/arxiv.2302.08191
• Publication Date: 2023-01-01
• Authors: Xuheng Cai; Chao Huang; Lianghao Xia; Xubin Ren
• Keywords: Generality; Robustness

Graph neural network (GNN) is a powerful learning approach for graph-based recommender systems. Recently, GNNs integrated with contrastive have shown superior performance in recommendation their data augmentation schemes, aiming at dealing highly sparse data. Despite success, most existing graph methods either perform stochastic (e.g., node/edge perturbation) on the user-item interaction graph, or rely heuristic-based techniques user clustering) generating views. We argue that these cannot well preserve intrinsic semantic structures and are easily biased by noise perturbation. In this paper, we propose simple yet effective paradigm LightGCL mitigates issues impairing generality robustness of CL-based recommenders. Our model exclusively utilizes singular value decomposition augmentation, which enables unconstrained structural refinement global collaborative relation modeling. Experiments conducted several benchmark datasets demonstrate significant improvement our over state-of-the-arts. Further analyses superiority LightGCL's against sparsity popularity bias. The source code available https://github.com/HKUDS/LightGCL.