Simultaneous non-gaussian data clustering, feature selection and outliers rejection

Abstract

A method for simultaneous non-Gaussian data clustering, feature selection and outliers rejection is proposed in this paper. The proposed approach is based on finite generalized Dirichlet mixture models learned within a framework including expectation-maximization updates for model parameters estimation and minimum message length criterion for model selection. Through a challenging application involving texture images discrimination, it is demonstrated that the developed procedure performs effectively in avoiding outliers and selecting relevant features. © 2011 Springer-Verlag Berlin Heidelberg.