Paper shows instability in class-split anomaly detection evaluation

Arxiv CS.LG (machine learning)

The paper identifies score-direction instability in within-dataset class-split evaluation for anomaly detection, showing the protocol can become ill-posed when anomaly classes overlap with normal data. This raises concerns about the reliability of common evaluation practices.

Arxiv CS.LG (machine learning)·Alejandro AscarateLeo LebratRodrigo Santa CruzClinton FookesOlivier Salvado

8 days ago