ResTran: A GNN Alternative To Learn Graph With Features

Shota Saito, Takanori Maehara, Mark Herbster

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.

Keywords: GNN, Spectral Clustering

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

TL;DR: We propose a vector representation for "graph with features," which can be an alternative to GNNs.

Abstract:

This paper considers a vertex classification task where we are given a graph and associated vector features.
The modern approach to this task is graph neural networks (GNNs).
However, due to the nature of GNN architectures, GNNs are known to be biased to primarily learn homophilous information.
To overcome this bias in GNN architectures, we take a simpler alternative approach to GNNs.
Our approach is to obtain a vector representation capturing both features and the graph topology.
We then apply standard vector-based learning methods to this vector representation.
For this approach, we propose a simple transformation of features, which we call \textit{Resistance Transformation} (abbreviated as \textit{ResTran}).
We provide theoretical justifications for ResTran from the effective resistance, $k$-means, and spectral clustering points of view.
We empirically demonstrate that ResTran is more robust to the homophilous bias than established GNN methods.

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: 2051