Spectral Highways: Injecting Homophily into Heterophilic Graphs

Karan Goyal, Saankhya Samanta, Mukesh Mohania, Vikram Goyal

Keywords: Heterophily, GNN, Homophily

Abstract: The application of convolution in graphs is at the core of Graph Neural Network (GNN) algorithms that led to the emergence of Graph Representation Learning (GRL). Various algorithms have been proposed over the last few years to solve the classical GRL task of node classification in a transductive setting. It is widely assumed that standard GNNs perform better on graphs with high homophily, i.e., nodes belonging to the same class are highly connected to each other. This assumption has led to the designing of specialised algorithms in the last few years for datasets that do not contain the property of homophily, i.e., heterophilic datasets. In this work, we both challenge and leverage this assumption. We argue that it is not necessary to follow the common trend of designing new algorithms but instead focus on understanding and enriching the data. We present a new technique from the perspective of data engineering that enables better performance on heterophilic datasets by both heterophilic GNN algorithms and non-heterophilic GNN algorithms. Our proposed technique, Spectral Highways, enables better connectivity and information flow between nodes in a heterophilic graph. We also draw an analogy between the performance of Spectral Highways and a recently proposed network property, i.e., Adjusted homophily. We conduct experiments on 11 baselines and 8 heterophilic datasets and achieve significant improvements in results.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7904