SoLAR: Surrogate Label Aware GNN Rewiring

ICLR 2025 Conference Submission4359 Authors

Keywords: graph neural networks, graph rewiring, homophily, heterophily, knowledge distillation

TL;DR: SoLAR rewires edges based on predicted node labels, which improves homophily and overall GNN performance.

Abstract: Rewiring the input graph of graph neural networks (GNNs) has been proposed as a pre-processing step to address issues like over-squashing and over-smoothing. However, most existing techniques rely solely on topology-based modifications, neglecting performance-critical node label information. To fill this gap, we propose SoLAR (Surrogate Label Aware Rewiring), a method that rewires the graph based on predicted node labels from a surrogate model. We prove its effectiveness in a theoretically tractable setting highlighting two key mechanisms that enable its success. The first is a denoising effect, while the second is a novel knowledge distillation-inspired process, where information from a surrogate model is encoded into the graph structure. Extensive experiments demonstrate consistent improvements of SoLAR across various datasets. Notably, the best surrogate models arise from iterative SoLAR, and reusing the same model class is a competitive strategy.

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/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 4359