Unsupervised Anomaly Detection through Mass Repulsing Optimal Transport

An anomaly, or an outlier, is a data point that is significantly different from the remaining data [Aggarwal, 2017], to such an extent that it was likely generated by a different mechanism [Hawkins, 1980]. From the perspective of machine learning, Anomaly Detection (AD) wants to determine, from a set of examples, which ones are likely anomalies, typically through a score. This problem finds applications in many different fields, such as medicine Salem et al. [2013], cyber-security Siddiqui et al. [2019], and system monitoring Isermann [2006], to name a few. As reviewed in Han et al. [2022], existing techniques for AD are usually divided into unsupervised, semi-supervised and supervised approaches, with an increasing need for labeled data. In this paper, we focus on unsupervised AD, which does not need further labeling effort in constituting datasets. As discussed in Livernoche et al. [2024], the growing number of applications involving high-dimensional and complex data begs the need for non-parametric algorithms.