Modelo de mistura com número de componentes desconhecido: estimação via método split-merge

Abstract

We propose the split-merge MCMC and birth-split-merge MCMC algorithms to analyse mixture models with an unknown number of components. The strategy for splitting is based on data and posterior distribution. Allocation probabilities are calculated based on component parameters which are generated from the posterior distribution given the previously allocated observations. The split-merge proposals are developed to be reversible and are accepted according to Metropolis-Hastings probability. This procedure makes possible a greater change in configuration of latent variables, in a single iteration of algorithms, allow a major exploration of clusters and avoid possible local modes. As an advantage, our approach determines a quick split proposal in contrary to former split procedures which require substantial computational effort. In the birth-split-merge MCMC algorithm, the birth movement is obtained directly from the procedure to update the latent variables and occurs when an observation determine a new cluster. The performance of the method is verified using artificial data sets and two real data sets. The first real data set consist of benchmark data of velocities from distant galaxies diverging from our own while the second is Escherichia Coli bacterium gene expression data.