New Benchmarks for Learning on Non-Homophilous Graphs

TL;DR

This paper introduces new graph datasets with low homophily, a better measure for homophily, and benchmarks various methods to advance research in non-homophilous graph learning.

Contribution

It provides improved non-homophilous graph datasets, a novel homophily measure, and comprehensive benchmarking to facilitate research in low-homophily settings.

Findings

Simple methods perform competitively on new datasets

Graph neural networks show varied effectiveness in low-homophily regimes

Insights suggest directions for developing better models for non-homophilous graphs

Abstract

Much data with graph structures satisfy the principle of homophily, meaning that connected nodes tend to be similar with respect to a specific attribute. As such, ubiquitous datasets for graph machine learning tasks have generally been highly homophilous, rewarding methods that leverage homophily as an inductive bias. Recent work has pointed out this particular focus, as new non-homophilous datasets have been introduced and graph representation learning models better suited for low-homophily settings have been developed. However, these datasets are small and poorly suited to truly testing the effectiveness of new methods in non-homophilous settings. We present a series of improved graph datasets with node label relationships that do not satisfy the homophily principle. Along with this, we introduce a new measure of the presence or absence of homophily that is better suited than existing measures in different regimes. We benchmark a range of simple methods and graph neural networks across our proposed datasets, drawing new insights for further research. Data and codes can be found at https://github.com/CUAI/Non-Homophily-Benchmarks.