Diversifying Spurious Subgraphs for Graph Out-of-Distribution Generalization
Keywords: OOD generalization, invariant learning, graph neural networks
TL;DR: We propose a novel learning framwork that maximally randomizes the spurious subgraphs in a reduced seach space for environment extrapolation and OOD generalization
Abstract: Environment augmentation methods have gained some success in overcoming the out-of-distribution (OOD) generalization challenge in Graph Neural Networks (GNNs). Yet, there exists a challenging trade-off in the augmentation: On one hand, it requires the generated graphs as diverse as possible to extrapolate to unseen environments. On the other hand, it requires the generated graphs to preserve the invariant substructures causally related to the targets. Existing approaches have proposed various environment augmentation strategies to enrich spurious patterns for OOD generalization. However, we argue that these methods remain limited in diversity and precision of the generated environments for two reasons: i) the deterministic nature of the graph composition strategy used for environment augmentation may limit the diversity of the generated environments, and ii) the presence of spurious correlations may lead to the exclusion of invariant subgraphs and reduce the precision of the generated environments. To address this trade-off, we propose a novel paradigm that accurately identifies spurious subgraphs, and an environment augmentation strategy called spurious subgraph diversification, which extrapolates to maximally diversified spurious subgraphs by randomizing the spurious subgraph generation, while preserving the invariant substructures. Our method is theoretically sound and demonstrates strong empirical performance on both synthetic and real-world datasets, outperforming the second-best method by up to 24.19% across 17 baseline methods, underscoring its superiority in graph OOD generalization.
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: 3037
Loading