CMStatistics 2018: Start Registration
View Submission - CMStatistics
Title: On the use of latent variables to extend Gaussian mixture models Authors:  Geoffrey McLachlan - University of Queensland (Australia) [presenting]
Abstract: Hiding behind the structure of finite mixture models are the latent variables that define the component labels in the conceptualization of a mixture distribution as applying in the case where the observed random variable is selected from one of the component distributions with (prior) probabilities specified by the mixing proportions. The inclusion of additional latent variables can further extend the flexibility of finite mixture distributions to model complex data. These additional latent variables include the latent factors in the case of mixtures of factor models and their deep versions and also the latent skewing variables in the case of mixtures of skew-symmetric component distributions. Various examples are presented to illustrate the improvement in model fit by adopting this use of latent variables.