Multivariate outlier detection with compositional data

Abstract

Multivariate outlier detection is usually based on Mahalanobis distances, by plugging in robust estimates of location and covariance. For compositional data, carrying only relative information, a special transformation needs to be consulted in order to be able to work in the appropriate geometry. The e ect of the trans- formation is discussed in this contribution. Furthermore, di erent possibilities for the interpretation of the identi ed multivariate outliers are presented.