Unsupervised anomaly detection algorithms on real-world data: how many do we need?  

29 November  2023 | Online | 14:00 | Roel Bouman (Radboud University in Nijmegen) 

Abstract

Unsupervised anomaly detection is a paradigm which is applied at an increasingly rapid pace across a variety of different application domains. Knowing which methods to use in practice is however still a challenge.

Without the possibility for optimization common in supervised learning, the first attempt is even more important. To separate the wheat from the chaff, we've performed the largest comparison of unsupervised anomaly detection algorithms on real-world tabular data, with 32 algorithms applied on 52 datasets. We find that we actually don't need that many algorithms in practice, and a subset suffices for most practical purposes. 

Bio
Roel Bouman is a last-year PhD-candidate in Machine Learning at the Radboud University in Nijmegen, the Netherlands. He studied Chemistry and Molecular Life Sciences, before acquiring a double master degree in Chemistry and Computing Science, with specialisations in chemometrics and data science. Roel's research focuses on fundamental ML and AI research, mostly within the subject of Anomaly Detection, but with industrial applications in for example predictive maintenance.