Community-Aware Hard Subgraph Mining for Out-of-Distribution Generalization

Jiaxing Li, Jiashuo Liu, Weihuang Zheng, Shaohua Fan, Youyong Kong

Keywords: Graph neural networks, out-of-distribution, community, node classification

TL;DR: In this work, we propose Community-Aware Hard Subgraph Mining, a novel framework for OOD generalization on graphs that explicitly leverages latent community heterogeneity.

Abstract:

Graph Neural Networks (GNNs) are widely used for node classification tasks but often struggle to generalize when training and test nodes follow different distributions, limiting their real-world applicability. Recent approaches based on invariant learning attempt to address this issue but rely on impractical predefined environment labels or low-quality synthetic environments, and their strict invariance assumptions often fail under complex community-level variations.
In this work, we propose \textbf{CHASM} (\emph{\textbf{C}ommunity-Aware \textbf{Ha}rd \textbf{S}ubgraph \textbf{M}ining}), a novel framework for OOD generalization on graphs that explicitly leverages latent community heterogeneity. CHASM adversarially mines the hardest subgraphs via a learnable mask model, imposes community-aware regularization to enforce structural coherence, and applies adaptive subgraph augmentation to enhance robustness. A stability-driven learner is then optimized against these hardest cases, yielding a principled and effective solution to community-level shifts.
Extensive experiments under covariate and concept shifts demonstrate that CHASM consistently outperforms state-of-the-art baselines, while theoretical analysis provides further justification of its robustness.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 8224