Reproduce “Geometry-Aware Edge Pooling for 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
Katharina Limbeck, Lydia Mezrag, Guy Wolf, Bastian Rieck
Researchers at Helmholtz Munich and Université de Montréal developed new edge-contraction pooling layers, MagEdgePool and SpreadEdgePool, for Graph Neural Networks that explicitly preserve graph structural diversity using diffusion distances and geometric invariants. These methods achieve leading performance in graph classification and regression, demonstrating improved structural preservation and computational efficiency.
Key insights
- Magnitude and Spread, derived from diffusion geometry, serve as effective and isometry-invariant measures for quantifying and preserving structural diversity in graphs during pooling.
- Spread offers a computationally more efficient ($O(|X|^2)$ vs. $O(|X|^3)$) and numerically stable alternative to magnitude, demonstrating high correlation and comparable performance.
- Edge contraction, guided by these geometry-aware invariants, effectively preserves graph connectivity and spectral properties, leading to more geometrically faithful pooled graph representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.