Audit Questions Test Labels in Graph Anomaly Detection Benchmark

Graph-level anomaly detection aims to identify abnormal graphs using only normal training data. However, a new arXiv paper argues that common benchmarks may inadvertently use test labels during model selection, which could inflate performance and misrepresent label-free detection.

The paper specifically audits UB-GOLD, a benchmark whose results have been reused by subsequent studies. The audit examines whether those results genuinely support label-free detection or whether they depend on access to test labels.

The findings raise questions about the validity of comparisons in this area. The authors suggest that benchmark practices need to be re-examined to ensure that reported performance reflects the intended deployment scenario.