Reproduce “UltraGCN: Ultra Simplification of Graph Convolutional Networks 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

Kelong Mao, Jieming Zhu, Xi Xiao, Biao Lu, Zhaowei Wang, Xiuqiang He

UltraGCN simplifies Graph Convolutional Networks for recommendation by replacing explicit message passing with a constraint-based learning approach that approximates infinite-layer graph convolutions. This method achieves superior accuracy while dramatically improving training efficiency, demonstrating approximately a 14x speedup over LightGCN and reaching training speeds comparable to Matrix Factorization models.

Key insights

  • Explicit multi-layer message passing in GCNs is not strictly necessary; its effect can be approximated through a direct convergence constraint.
  • Careful, learned weighting of different relationship types (user-item, item-item) is more effective than implicit, fixed weighting from traditional message passing.
  • Negative sampling effectively mitigates over-smoothing while enabling the capture of higher-order graph information without requiring deep, explicit GCN layers.

Summary by alphaXiv.

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