Filling in the GAP: Achieving Robust and Adaptive GNNs through Post-Processing

Dooho Lee, Myeong Kong, Jaemin Yoo

Keywords: Graph neural networks, edge distribution shift, training-free post-processing, robustness, node classification

TL;DR: We introduce FILLER, a post-processing method to enhance the robustness of any trained GNNs against edge distribution shift by inserting Edge-shift Recovery module into each GNN layer.

Abstract:

Graph neural networks (GNNs) have shown significant success in modeling graph-structured data. However, their performance often deteriorates when faced with a change in the graph structure between training and test time, such as edge addition or removal—a common scenario considering the dynamic nature of graphs. To address this challenge, we propose FILLER (Framework for Integrating Layer-Level Edge-shift Recovery), a post-processing method which enhances the robustness of a GNN against edge sparsification while maintaining its adaptability to informative edge addition. Our key idea is to fill in the representation gap caused by edge distribution shift by injecting the Edge-shfit (ER) layer into each layer of the GNN. Our ER layer is carefully designed to allow a GNN to maintain its high performance in dynamic graph environments even without any additional training, and its effectiveness is shown both theoretically and empirically. Our experiments on ten datasets for node classification and five GNN architectures demonstrate that FILLER is broadly applicable across
diverse models and scenarios.

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: 2349