Title: Independent component analysis for compositional data
Authors: Kamila Facevicova - Palacky University Olomouc (Czech Republic) [presenting]
Christoph Muehlmann - Technical University of Vienna (Austria)
Klaus Nordhausen - Vienna University of Technology (Austria)
Alzbeta Gardlo - Palacky University (Czech Republic)
Abstract: Compositional data represent a family of multivariate data whose (strictly positive) parts carry relative information about the respective structure primarily. Due to this specific nature of the data, its direct analysis using standard multivariate statistical methods is not appropriate since it can lead to spurious results. As a way out, either the methods need to be modified with respect to the log-ratio methodology or the compositional dataset has to be expressed in a proper system of real-valued coordinates. The focus is on the adaptation of independent component analysis to the case of a compositional dataset. Independent component analysis aims at finding statistically independent components in the data and beside the dimension reduction, it is also suitable to search for groups within the data as well as outlying observations. The performance of the proposed methodology will be demonstrated on a metabolomics dataset.