On dimensionality of feature vectors in MPNNs
César Bravo, Alexander Kozachinskiy, Cristóbal Rojas
TL;DR
This work proves that MPNNs with one-dimensional feature vectors, coupled with any non-polynomial analytic activation, are uniformly equivalent to the WL test across graphs of any size. By constructing a 1-D MPNN with a single parameter $oldsymbol{\gamma} ext{ in }(0,1)$ and a specific update rule, the authors show that WL labels can be reconstructed and preserved through iterations, aided by a key linear-independence argument over $ obrace ext{ } obrace ext{Q}$. They provide a simple proof and substantial experimental validation, demonstrating that random choices of $oldsymbol{\gamma}$ achieve perfect WL simulation on large graph sets and that the required precision scales approximately as $O( obreak ext{log} n)$. The paper further discusses extensions to higher-order WL tests and analyzes implications for generalization and practical network design. Overall, the results indicate that low-dimensional feature spaces need not compromise the expressive power of MPNNs relative to the WL hierarchy.
Abstract
We revisit the classical result of Morris et al.~(AAAI'19) that message-passing graphs neural networks (MPNNs) are equal in their distinguishing power to the Weisfeiler--Leman (WL) isomorphism test. Morris et al.~show their simulation result with ReLU activation function and $O(n)$-dimensional feature vectors, where $n$ is the number of nodes of the graph. By introducing randomness into the architecture, Aamand et al.~(NeurIPS'22) were able to improve this bound to $O(\log n)$-dimensional feature vectors, again for ReLU activation, although at the expense of guaranteeing perfect simulation only with high probability. Recently, Amir et al.~(NeurIPS'23) have shown that for any non-polynomial analytic activation function, it is enough to use just 1-dimensional feature vectors. In this paper, we give a simple proof of the result of Amit et al.~and provide an independent experimental validation of it.
