Reproduce “XSimGCL: Towards Extremely Simple Graph Contrastive Learning for Recommendation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junliang Yu, Xin Xia, Tong Chen, Lizhen Cui, Nguyen Quoc Viet Hung, Hongzhi Yin

XSimGCL introduces a graph contrastive learning method for recommendation that simplifies existing approaches by replacing complex graph augmentations with direct noise-based embedding perturbation and a single-pass architecture. This enables state-of-the-art recommendation accuracy while significantly reducing training time and mitigating popularity bias on large, sparse datasets.

Key insights

  • The InfoNCE contrastive loss, through promoting uniformity in representation distribution, is the primary driver of performance in graph CL for recommendation, rendering complex graph augmentations less critical.
  • Contrastive learning, specifically by encouraging more uniform embedding distributions, effectively mitigates popularity bias and enhances the recommendation of long-tail items.
  • Cross-layer contrast offers an efficient and effective strategy for view generation in CL, leveraging distinct frequency components captured at different GNN layers.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.