Graph Out of Distribution Detection

Graph out-of-distribution (OOD) detection focuses on identifying graph data that differs significantly from the training distribution, a crucial task for building robust and reliable graph machine learning systems. Current research emphasizes unsupervised methods, often employing contrastive learning or graph augmentation techniques to learn representations that distinguish in-distribution from OOD graphs, sometimes incorporating substructure analysis for improved performance. This field is vital for ensuring the safety and trustworthiness of graph-based AI applications across various domains, including those with limited labeled data.