Heterogeneous Normal Classes Pose a Challenge for Anomaly Detection
Keywords: Anomaly Detection, Heterogeneous Normal Class, Benchmarking
TL;DR: We evaluate SOTA Anomaly Detection methods under heterogeneous normality, highlighting a surprising performance deterioration in this setting.
Abstract: Anomaly detection is crucial for developing reliable and robust Machine Learning methods. Commonly, anomaly detection methods assume access to only normal samples during training, while at test time, the objective is to discriminate between normal and anomalous samples. Recently, the field has seen a surge in new methods, reporting impressive performances on various benchmarks. The default evaluation procedure for many of these methods, however, implicitly assumes a homogeneous normal class. In this paper, we investigate how recent methods perform under varying degrees of heterogeneity of the normal class. We find that even state-of-the-art methods struggle under non-homogeneous normality, exhibiting deteriorating performance as the heterogeneity of the normal class increases, even when increasing the amount of training data. Our results highlight the importance of evaluating anomaly detection techniques on a broader set of normal classes, encouraging future research to address this crucial aspect.
Primary Subject Area: Data collection and benchmarking techniques
Paper Type: Extended abstracts: up to 2 pages
Participation Mode: In-person
Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.
Submission Number: 93
Loading