Tribe: Tri-Component Information Decomposition for Graph Out-of-Distribution Detection

Danny Wang, Ruihong Qiu, Zi Huang

Keywords: graph out-of-distribution detection

Abstract: Graph neural networks are widely used for node classification, but they remain vulnerable to out-of-distribution (OOD) shifts in node features and graph structure. Existing methods trained with standard supervised learning (SL) objectives tend to capture spurious signals from either features and/or structure, leaving the model fragile under distributional changes. To address this, we propose Tribe, a novel and effective Tri-Component Information Decomposition framework that explicitly decomposes information into feature-specific, structure-specific and joint components. Tribe aims to preserve only the label-relevant component of the joint information while filtering out spurious feature- and structure-specific information, thereby enhancing the separation between in-distribution (ID) and OOD data. Technically, we develop a novel optimisation pipeline that integrates a graph Information Bottleneck (IB) objective with carefully designed regularisations. Beyond the framework, we provide theoretical and empirical analysis showing the superiority of IB in OOD detection, with higher ID confidence and a larger entropy gap between ID and OOD data compared to the typical SL objective. Extensive experiments across seven datasets confirm the efficacy of Tribe, achieving up to 34% improvement in FPR95 over strong baselines while maintaining competitive ID accuracy. Code will be released upon acceptance.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 6667