Map to Optimal: Adapting Graph Out-of-Distribution in Test Time

Haoxiang Zhang, Zhuofeng Li, Qiannan Zhang, Ziyi Kou, Lianyong Qi, Juncheng Li, Shichao Pei

Keywords: Out-of-distribution Generalizarion, Test-time Adaptation, Graph Neural Network, Self-supervision

Abstract:

Based on topological proximity message passing, graph neural networks (GNNs) can quickly model data patterns on graphs. However, at test time, when the node feature and topological structure of the graph data are out-of-distribution (OOD), the performance of pre-trained GNNs will be hindered. Existing test-time methods either fine-tune the pre-trained model or overlook the discrepancy between the prior knowledge in pre-trained models and the test graph. We propose a novel self-supervised test-time adaptation paradigm GOAT (https://anonymous.4open.science/r/GOAT-5C0E), through graph augmentation-to-augmentation strategy, that enables a simple adapter can mitigate the distribution gap of training data and test-time data. GOAT reduces generalization error for node classification in various pre-trained settings through experiments on six benchmark datasets spanning three distinct real-world OOD scenarios. Remarkably, GOAT outperforms state-of-the-art test-time methods, and our empirical study further demonstrates the interpretability of the OOD representation generated from our method.

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.

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: 7480