Reproduce “Weisfeiler and Lehman Go Cellular: CW 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
Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yu Guang Wang, Pietro Liò, Guido Montúfar, Michael Bronstein
CW Networks leverage regular cell complexes and a cellular Weisfeiler-Lehman test to enhance the expressive power of geometric deep learning models for graph-structured data. This approach enables principled modeling of higher-order interactions, leading to state-of-the-art results on molecular graph prediction tasks and distinguishing graphs beyond the capabilities of 3-WL tests.
Key insights
- Regular cell complexes provide a more flexible and powerful topological framework than simplicial complexes, allowing for diverse and expressive graph 'lifting' transformations.
- The Cellular Weisfeiler-Lehman (CWL) test, particularly with skeleton-preserving liftings like k-cliques or induced cycles, is provably strictly more powerful than the 1-WL test and can even surpass the 3-WL test in discriminative power.
- Higher-dimensional cells in CWNs act as computational 'shortcuts,' enabling efficient information propagation across distant parts of the graph in fewer layers, thereby mitigating over-smoothing and over-squashing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.