On the Expressive Power of Sparse Geometric MPNNs

Part of International Conference on Representation Learning 2025 (ICLR 2025) Conference

Bibtex Paper Supplemental

Authors

Yonatan Sverdlov, Nadav Dym

Abstract

Motivated by applications in chemistry and other sciences, we study the expressivepower of message-passing neural networks for geometric graphs, whose nodefeatures correspond to 3-dimensional positions. Recent work has shown that suchmodels can separate generic pairs of non-isomorphic geometric graphs, though theymay fail to separate some rare and complicated instances. However, these resultsassume a fully connected graph, where each node possesses complete knowledgeof all other nodes. In contrast, often, in application, every node only possessesknowledge of a small number of nearest neighbors.This paper shows that generic pairs of non-isomorphic geometric graphs canbe separated by message-passing networks with rotation equivariant features aslong as the underlying graph is connected. When only invariant intermediatefeatures are allowed, generic separation is guaranteed for generically globallyrigid graphs. We introduce a simple architecture, EGENNET, which achieves ourtheoretical guarantees and compares favorably with alternative architecture onsynthetic and chemical benchmarks