Reproduce “Spatio-Spectral Graph Neural Networks” 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

Simon Geisler, Arthur Kosmala, Daniel Herbst, Stephan Günnemann

Researchers from Technical University of Munich introduce Spatio-Spectral Graph Neural Networks (S²GNNs), a hybrid model that combines local message-passing with global spectral filtering to overcome over-squashing and limited receptive fields in traditional GNNs. The approach achieves new state-of-the-art results on long-range interaction benchmarks while also demonstrating scalability to large graphs and competitive performance on sequence modeling tasks.

Key insights

  • S²GNNs are formally proven to circumvent the over-squashing phenomenon by enabling uniformly lower-bounded Jacobian sensitivity independent of node distance.
  • The hybrid approach yields strictly tighter error bounds for approximating complex or discontinuous graph filters compared to purely polynomial spatial filters.
  • The designed positional encodings are strictly more expressive than the 1-Weisfeiler-Lehman test, capable of distinguishing graphs that simpler architectures cannot.

Summary by alphaXiv.

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