Contamination-robust anomaly detection under weak supervision with explainable representations

Weakly supervised anomaly detection leverages a small set of labeled anomalies together with large amounts of unlabeled data. Existing methods improve generalization by modeling diverse anomaly patterns but often assume unlabeled data to be clean, ignoring the common issue of anomaly contamination, which can bias representation learning and degrade performance. We propose CRAD, a contamination-robust framework that explicitly models contamination under weak supervision. Rather than treating unlabeled samples as normal, CRAD reconstructs them away from the anomalous distribution and incorporates decontaminated reconstructions into robust representation learning. In addition, CRAD provides feature-level anomaly explanations via an attribution mechanism that quantifies each feature's contribution, enhancing interpretability and detection reliability. Experiments on 15 real-world datasets show that CRAD consistently outperforms five state-of-the-art methods, achieving superior robustness, generalization, and reliable feature-level explanations.