B1444
Title: On contaminated transformation mixture models
Authors: Yana Melnykov - The University of Alabama (United States) [presenting]
Xuwen Zhu - University of Alabama (United States)
Volodymyr Melnykov - The University of Alabama (United States)
Abstract: Gaussian mixture models have been the most popular mixtures in literature for many decades. However, the adequacy of the fit provided by Gaussian components is often questioned due to skewness or heavy tails. Various distributions capable of modelling these features have recently been considered in the mixture modelling context. A contaminated transformation mixture model is introduced that is constructed based on the idea of transformation to symmetry. The proposed mixture can effectively account for skewness and heavy tails and automatically detect scatter by assigning such data points to secondary mixture components. The performance and promise of the proposed model are illustrated on synthetic data in various settings as well as popular classification data sets.
