GEO: An Approach for Out-of-Distribution Detection in Graph Neural Networks Using Energy Scoring and One-Class Learning

Abstract

Graph neural networks (GNN) have emerged as a powerful approach for learning from data that is structured as a graph. However, a remaining challenge is developing reliable and robust GNN models that can accurately detect inputs that are outside the distribution of training data. In this study, we propose a novel method for detecting Out-of-distribution (OOD) inputs that combines GNN with energy-based learning and one-class classification objectives. We name this approach as GEO (Graph Energy-based Out-of-distribution detection). GEO assigns anomaly scores using an energy function derived from GNN logits, correlating with sample density for reliable uncertainty estimates. Our approach, which combines classification, energy scoring, and one-class objectives, excels in OOD detection, outperforming existing methods on sparse graph datasets. GEO achieves top AUROC scores, including 95.51% on Cora and 99.99% on Amazon, enhancing GNN reliability and offering a theoretically sound solution for detecting distributional shifts.