Refining Graph OOD Detection in Heterophilous Contexts through Neural Collapse

Out-of-distribution (OOD) detection in graph neural networks plays a crucial role in relevant security deployments. Recent research on graph OOD detection focuses on the score of individual instances, utilizing aggregation operations to update scores. These approaches generally assume that neighboring instances are of the same distribution, an assumption known as homophily. However, in heteroophilic scenarios, these methods have a flaw in that the aggregation operation reduces the difference between the In-distribution (ID) data and the OOD data. To address this issue, we propose a novel inference-based approach, Graph Neural Collapse OOD (GNCO), which leverages the geometric properties of Neural Collapse and optimizes the aggregation operation, achieving generalization in heterophilous settings. In GNCO, we consider a unique geometric property of neural collapse, i.e., ID and OOD features tend to become orthogonal during the terminal phase of training. Based on this insight, the norm of the test instance feature within the ID subspace can be devised to derive the OOD score. We conducted comprehensive experiments to validate the efficacy of GNCO. Our results confirm that this method outperforms all benchmarks across a variety of real-world datasets.