Seminário
Conjunto UFSCar/ICMC ? 09/05/2014 -
14h00 LOCAL: Sala de
seminários ? DEs-UFSCar TÍTULO: A
Split-Merge MCMC based on Mahalanobis distance for analysis of mixture models
with an unknown number of components PALESTRANTE:
Erlandson Ferreira Saraiva ? -
UFMS We present a split-merge MCMC algorithm for analysis of mixture models with an unknown number of components. In a split proposal each observation is allocated to one of two split components based on probabilities which are calculated according to Mahalanobis distance. Conditional on new proposal allocation new component parameters are generated from a candidate-generating density that is chosen according to known form of posterior distributions of the component parameters. In order to maintain detailed balance equation within each move type, the acceptance probability for split-merge proposals are calculated according to reversible-jump procedure using a ``dimension matching'' scheme where the Jacobian term is unity. The main advantage of the proposed algorithm is that it is easy to implement, even for the multivariate case, and do not requires the invention of ``good'' jumping moves to apply it to a new family of mixtures. We illustrate it on both univariate and bivariate data. |