Nicolas Goix – DAMEX Algorithm

DAMEX (Detecting Anomaly among Multivariate Extremes) is an algorithm introduced in [AISTATS 2016 paper -- short version] and [submitted paper -- long version with theoretical guaranties], and selected to be presented within the `Bourse aux Technologies--Industrie du Futur--Smart Manufacturing', at the French Ministry of Industry, Paris March 2016. It builds on multivariate Extreme Value Theory to learn how to rank observations in a high dimensional space, wrt their supposed degree of abnormality. The procedure relies on an original dimension reduction technique in the extremal domain that possibly produces a sparse representation of multivariate extremes and permits to gain insight into their dependence structure, escaping the curse of dimensionality. The representation output by the unsupervised methodology can be actually combined with any Anomaly Detection technique tailored to non-extreme data. The approach is novel in the sense that Extreme Value Theory has never been used in its multivariate version in the field of Anomaly Detection.