A novelty Bayesian method for unsupervised learning of finite mixture models

Mixture models have universal applications in probabilistic modeling for multivariate data. This paper proposes a novelty parameters estimation and model selection method based on half Bayesian. In our algorithm, the mixture coefficient is a determinate variable, and the parameters of the components are random variables. Owing to the special prior distribution of the parameters, the parameters don´t converge toward a singular estimation at the boundary of the parameter space, and the redundant components can be automatically removed. In a word, our algorithm has very good performances as follows: (1) automatically selects the number of components, (2) avoids converging the boundary of the parameter space, (3) is less sensitive to initialization, (4) can fulfil simultaneously the parameters estimation and model selection in one algorithm, therefore, the computation efficiency is higher. The experimental results show that the algorithm is effective and has above good performances.