How Powerful Are Graph Neural Networks?

By Keyulu Xu et al

Published on Feb. 22, 2019

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Preliminaries
3. Theoretical Framework: Overview
4. Building Powerful Graph Neural Networks
4.1 Graph Isomorphism Network (GIN)

Summary

Graph Neural Networks (GNNs) are effective for representation learning of graphs, with a focus on neighborhood aggregation. This paper presents a theoretical framework analyzing the expressive power of GNNs, comparing them to the Weisfeiler-Lehman (WL) graph isomorphism test. It introduces Graph Isomorphism Network (GIN) as a powerful GNN model that can distinguish different graph structures efficiently. The study explores the conditions for GNNs to be as powerful as the WL test, emphasizing the importance of injective aggregation functions. GIN utilizes deep multisets and universal functions for neighbor aggregation, achieving high discriminative power among GNN architectures.