PATTERN MATCHING-BASED OUT-OF-DISTRIBUTION DETECTION FOR MULTI-LABEL NODE CLASSIFICATION
Keywords: Out-of-distribution
Abstract: Graph neural networks (GNNs) have achieved dominant performance in various prediction tasks on graphs. When deploying GNNs in the real world, estimating the possibility of out-of-distribution (OOD) testing samples becomes a crucial safety concern. Although some research has investigated the graph OOD detection problem, most have concentrated on single-label classification scenarios, aspecific case of the more general multi-label classification, which has broader applications, such as in social networks where nodes can represent users with multiple interests or attributes. In this paper, we first introduce and define the multi-label graph OOD detection problem and propose a simple yet effective pattern matching-based OOD detection method to address it. In particular, our method utilizes feature pattern matching and label pattern matching to obtain two matching scores. By incorporating topological structure adjustment, we ultimately derive confidence scores, serving as indicators of the likelihood that a test sample is an OOD instances. We conduct extensive comparisons with existing OOD detection methods in the context of multi-label graphs. The results show that our method achieves an impressive 7.61% reduction in FPR95 compared to the leading baselines, setting a new state-of-the-art. Furthermore, our approach can servas a benchmark for OOD detection on multi-label graphs.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2304
Loading