Towards a Trustworthy Anomaly Detection for Critical Applications through Approximated Partial AUC Loss

Arnaud Bougaham, Benoît Frénay · 2025

Read Towards a Trustworthy Anomaly Detection for Critical Applications through Approximated Partial AUC Loss by Arnaud Bougaham, Benoît Frénay (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11570.txt.

Description

Arnaud Bougaham and Benoît Frénay train anomaly detectors that refuse to miss a real fault. Their 2025 paper targets critical fields like industry, medicine, and cybersecurity, where a missed case costs far more than a false alarm. Their loss function reshapes training toward that priority.

The method optimizes one slice of the ROC curve: the part below a perfect catch rate. Training pushes that slice toward fewer false alarms while holding catches high. The best threshold with zero misses is then locked for testing. Across six datasets, the detector catches about ninety-two and a half percent of faults at a twenty percent false alarm rate. That beats prior methods on catches by over four points. Code is public.

Catalog record

CREATORArnaud Bougaham, Benoît Frénay
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-11570

Browse the full library · About the Tower of Babel