Reproduce “Rethinking Over-Smoothing in Graph Neural Networks: A Perspective from Anderson Localization” 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

Kaichen Ouyang

“Rethinking Over-Smoothing in Graph Neural Networks” frames deep message passing through the lens of Anderson localization, relating graph spectral modes, participation degree, and structural disorder to the loss of node-specific information. The work develops a conceptual framework and a degree-aware aggregation hypothesis, but provides no benchmark experiments or empirical validation.

Key insights

  • Over-smoothing is interpreted as a spectral propagation problem in which low-frequency components dominate while high-frequency, node-distinguishing information is attenuated or spatially restricted.
  • The framework suggests participation degree and localization length as possible layer-wise diagnostics for studying how graph signals spread or become confined.
  • The Anderson-localization analogy is suggestive but not yet fully resolved: globally smooth low-frequency modes and localized eigenstates are not equivalent, and degree variance alone may not capture graph disorder.

Summary by alphaXiv.

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