Zero-Shot Generalization of GNNs over Distinct Attribute Domains
Keywords: GNN, zero-shot, graph foundation models
TL;DR: We propose the first method that can zero-shot generalize to test attributed graphs with different features than those seen in train.
Abstract: Inductive GNNs are able to generalize across graphs with the same set of node attributes. However, zero-shot generalization across attributed graphs with disparate node attribute domains remains a fundamental challenge in graph machine learning. Existing methods are unable to effectively make use of node attributes when transferring to unseen attribute domains, frequently performing no better than models that ignore attributes entirely. This limitation stems from the fact that models trained on one set of attributes (e.g., biographical data in social networks) fail to capture relational dependencies that extend to new attributes in unseen test graphs (e.g., TV and movies preferences). Here, we introduce STAGE, a method that learns representations of _statistical dependencies_ between attributes rather than the attribute values themselves, which can then be applied to completely unseen test-time attributes, generalizing by identifying analogous dependencies between features in test. STAGE leverages the theoretical link between maximal invariants and measures of statistical dependencies, enabling it to provably generalize to unseen feature domains for a family of domain shifts. Our empirical results show that when STAGE is pretrained on multiple graph datasets with unrelated feature spaces (distinct feature types and dimensions) and evaluated zero-shot on graphs with yet new feature types and dimensions, it achieves a relative improvement in Hits@1 between 40% to 103% for link prediction, and an 10% improvement in node classification against state-of-the-art baselines.
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: 12237
Loading