Expressive Power of Invariant and Equivariant Graph Neural Networks

TL;DR

This paper develops a theoretical framework to compare the expressive power of invariant and equivariant GNNs, providing the first approximation guarantees for practical architectures and demonstrating their effectiveness on complex problems like the Quadratic Assignment Problem.

Contribution

It introduces a new theoretical framework for GNN expressiveness, proves approximation guarantees for practical GNNs, and shows FGNNs outperform existing methods on NP-hard problems.

Findings

FGNNs are the most expressive tensor-based GNNs to date.

FGNNs can effectively learn solutions to the Quadratic Assignment Problem.

Practical implementation of masked tensors enables handling variable-sized graph batches.

Abstract

Various classes of Graph Neural Networks (GNN) have been proposed and shown to be successful in a wide range of applications with graph structured data. In this paper, we propose a theoretical framework able to compare the expressive power of these GNN architectures. The current universality theorems only apply to intractable classes of GNNs. Here, we prove the first approximation guarantees for practical GNNs, paving the way for a better understanding of their generalization. Our theoretical results are proved for invariant GNNs computing a graph embedding (permutation of the nodes of the input graph does not affect the output) and equivariant GNNs computing an embedding of the nodes (permutation of the input permutes the output). We show that Folklore Graph Neural Networks (FGNN), which are tensor based GNNs augmented with matrix multiplication are the most expressive architectures proposed so far for a given tensor order. We illustrate our results on the Quadratic Assignment Problem (a NP-Hard combinatorial problem) by showing that FGNNs are able to learn how to solve the problem, leading to much better average performances than existing algorithms (based on spectral, SDP or other GNNs architectures). On a practical side, we also implement masked tensors to handle batches of graphs of varying sizes.