B1402
Title: Conditional mixture modeling
Authors: Volodymyr Melnykov - The University of Alabama (United States) [presenting]
Abstract: Modern mixture models are complex and often under risk of overparameterization. To address this concern, one popular approach is to consider various parsimonious models obtained by introducing constraints on covariance matrix structures. We propose an alternative approach to parsimonious mixture modeling that is based on modeling location rather than dispersion parameters. The developed model proves to be flexible, especially in the presence of non-compact clusters. Fast parameter estimation is possible due to the availability of analytical expressions. Conducted simulation studies, as well as applications to well-known classification data sets, demonstrate the promise and competitiveness of the proposed model.
