Cost-effective anomaly detection with active learning

Anomaly detection is essential for ensuring quality and reliability in many domains, yet collecting labeled anomalies is costly and often impractical. Traditional unsupervised anomaly detection (UAD) methods assume access to purely normal training data. However, in real-world settings unlabeled datasets are typically contaminated with anomalies, leading to degraded performance. To address this challenge, we propose Anomaly Detection with Active Learning (ADAL), a cost-effective framework that reformulates the anomaly detection task as a classification problem within an active learning paradigm. ADAL begins with a UAD model that assigns anomaly scores without initial labeling cost, from which reliable pseudo-labels are derived to train an initial classifier and mitigate the cold-start problem. The classifier is then iteratively improved by combining pseudo-labeled data with selectively acquired true labels through query selection. Experiments on benchmark tabular and image datasets demonstrate that ADAL consistently outperforms conventional UAD and active learning baselines, achieving performance competitive with fully supervised approaches while requiring substantially fewer labeled samples. These results highlight the effectiveness of ADAL as a practical strategy for cost-efficient anomaly detection under limited labeling budgets.