Dynamic Weights in Gaussian Mixture Models: A Bayesian Approach

In this paper we consider a Gaussian mixture model where the mixture weight behaves as an unknown function of time. To estimate the mixture weight function, we develop a Bayesian nonlinear dynamic approach for polynomial models. Two estimation methods that can be extended to other situations are considered. One of them, called here component-wise Metropolis-Hastings, is more general and can be used for any situation where the observation and state equations are nonlinearly connected. The other method tends to be faster but must be applied specifically to binary data (by using a probit link function). This kind of Gaussian mixture model is capable of successfully capturing the features of the data, as observed in numerical studies. It can be useful in studies such as clustering, change-point and process control. We apply the proposed method an array Comparative Genomic Hybridization (aCGH) dataset from glioblastoma cancer studies, where we illustrate the ability of the new method to detect chromosome aberrations.