New Statistical Framework Tames False Alarms in Anomaly Detection Under Domain Shift

Semi-supervised anomaly detection, a technique for spotting rare and critical events in machine learning, loses accuracy when data distributions shift between domains. A new study published in the International Journal of Data Science and Analytics introduces PSI-SAD-DA, described by its authors as the first post-selection inference framework for anomaly detection performed after domain adaptation. The work, authored by Tran Tuan Kiet, Nguyen Thang Loi, Tran Xuan Dat and Vo Nguyen Le Duy, was published on 28 October 2026 in volume 22 of the journal.

The core problem the researchers tackle is selection bias. In semi-supervised anomaly detection under domain adaptation, a pipeline transfers knowledge from a related source domain to build informative representations of a target domain, and then flags anomaly candidates. Because those candidates are chosen in a data-dependent manner after adaptation and detection, classical statistical inference on them becomes invalid, producing an uncontrolled false positive rate. In practical terms, the system can report far too many anomalies that turn out to be ordinary, undermining trust in its alerts.

PSI-SAD-DA addresses this gap by constructing p-values for detected anomalies that the authors describe as provably valid, explicitly accounting for the selection bias introduced when candidates are filtered through a fixed, pretrained anomaly detection and domain adaptation pipeline. According to the abstract, this ensures rigorous control of the false positive rate at a pre-specified level, such as 0.05, meaning that only about five percent of normal cases flagged as anomalies would be expected to be false alarms under the framework’s guarantees.

A key contribution is computational. Post-selection inference is typically expensive for deep neural networks, but the team introduces an efficient dynamic programming-based line-search algorithm that reuses intermediate computations, which they report significantly reduces the cost of p-value computation. This makes the theoretically grounded approach practical for large-scale deep learning architectures rather than remaining a purely mathematical construction.

The authors evaluated the method on synthetic and real-world datasets, reporting that PSI-SAD-DA consistently achieved valid false positive rate control, higher statistical power, and computational efficiency compared to existing methods. Higher statistical power means the framework is better able to detect genuine anomalies while keeping false alarms in check, a balance that is often difficult to strike in anomaly detection systems operating under domain shift. The findings position the framework as a step toward anomaly detection pipelines whose alerts carry rigorous statistical guarantees even after the data-dependent steps of adaptation and candidate selection.

Subject of Research: Post-selection inference for semi-supervised anomaly detection after domain adaptation

Article Title: Post-selection inference for semi-supervised anomaly detection after domain adaptation

Article References: Kiet, T. T., Loi, N. T., Dat, T. X., & Duy, V. N. L. (2026). Post-selection inference for semi-supervised anomaly detection after domain adaptation. International Journal of Data Science and Analytics, 22(1), Article 364. https://doi.org/10.1007/s41060-026-01345-w

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01345-w

Keywords: anomaly detection, domain adaptation, post-selection inference, semi-supervised learning, statistical inference, false positive rate, p-values, deep neural networks, selection bias, dynamic programming, Post-selection, inference