Constrained Variational Generation for Generalizable Graph Learning

Song Wang, Jundong Li

Keywords: Graph Neural Networks, Information Bottleneck, Variational Inference

Abstract:

Out-of-distribution (OOD) generalization aims at dealing with scenarios where the test data distribution can largely differ from training data distributions.
Existing works for OOD generalization on graphs generally propose to extract invariant subgraphs that can provide crucial classification information even under unseen test domains. However, such a strategy is suboptimal due to two challenges: (1) \textit{intra-graph correlations}, i.e., correlated structures that are partial invariant, and (2) \textit{inter-graph distinctions}, i.e., significant distribution shifts among graphs. To achieve better generalizability of learned graph representation, we innovatively propose a \textbf{\underline{C}}onstrained \textbf{\underline{V}}ariational \textbf{\underline{G}}eneration (CVG) framework to generate generalizable graphs. Our framework is implemented based on the Variation Graph Auto-Encoder (VGAE) structure and optimized under the guidance of the Graph Information Bottleneck (GIB) principle, with its effectiveness validated by our theoretical analysis.
We conduct extensive experiments on real-world datasets and demonstrate the superiority of our framework over 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/2024/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: 5894