Stochastic Rate Mixture Of Erlang Model And Its Application

Finite mixture models provide a powerful,flexible and well principled statistical ap-proach to model complex datasets in many applications.In this paper,we change the rate parameter of mixed Erlang model to be a random variable drawn from a Gamma distribution with two parameters,and the new model,Stochastic Rate Mixture of Er-lang(SRMER)Model,is obtained.The SRMER can model complex datasets especially the thick tailed ones.In this pa-per,we propose a Bayesian approach for SRMER learning using the Markov Chain Monte Carlo(MCMC)technique which simultaneously allows model selection,cluster assign-ments and parameters estimation.The Dirichlet prior on the weights empties superfluous components during MCMC in an overfitting model settled before.And the CMM algo-rithm has been improved to make it more effective to calculate the initial value of the multi-parameter mixed model.In addition,scanning mixed Gibbs algorithm reduces the influence of high autocorrelation of parameters in the iteration process,which improves the mixture and convergence of the algorithm.Simulations show that the proposed approach of SRMER model works well in mod-el selection and parameter estimation.Finally,we apply the model to two real data sets,which support that the SRMER model performs better than the traditional mixed models in fitting and classifying heavy tailed datasets.