Reproduce “Collaborative Filtering Meets Spectrum Shift: Connecting User-Item Interaction with Graph-Structured Side Information” 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

Yunhang He, Cong Xu, Jun Wang, Wei Zhang

Spectrum Shift Correction (SSC) adapts graph collaborative filtering models to user-user and item-item side graphs by correcting the eigenvalue displacement caused when these edges are added to a bipartite interaction graph. Applied to LightGCN and JGCF without explicit eigendecomposition, SSC improves multimodal Recall@20 by up to 22.69%, raises Ciao performance by 11.80%, and shows greater robustness to noisy side information than several specialized models.

Key insights

  • The weakness of direct graph augmentation is attributed to frequency mismatch: filters designed for a bipartite graph may assign inappropriate weights after side-information edges alter the available spectrum.
  • A single matrix-level correction can support different side-information types and different spectral backbones, rather than requiring a specialized architecture for each graph.
  • Exact spectral restoration is not always optimal, so estimated correction parameters are useful initializations but still benefit from validation-based tuning.

Summary by alphaXiv.

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