Research On Nonparametric Bayesian Classification Model Via Data Augmentation

In target recognition community, when dealing with large-scale and complex distributed data, it is very expensive to train a classifier using all input data and the underlying structure of the data is ignored. As a generative model, FA only focuses on the observations without utilization of any label information. To overcome these limitations, we fully utilize the Bayesian estimation of the discriminative probabilistic latent models, especially the mixture models.In this paper, we first develop the max-margin similarity preserving factor analysis(MMSPFA) model which utilizes the latent variable support vector machine(LVSVM) as the classification criterion in the latent space to learn a discriminative subspace with maxmargin constraint. MMSPFA jointly learns factor analysis(FA) model, similarity preserving(SP) term and max-margin classifier in a united Bayesian framework to improve the prediction performance.Furthermore, to deal with multimodally distributed data, we further extend MMSPFA to infinite Gaussian mixture model and develop the Dirichlet process mixture max-margin similarity preserving factor analysis(DPM-MMSPFA) model, via the consideration of Dirichlet process mixtures(DPM). In this way, DPM-MMSPFA combines the advantages of Bayesian model to capture the latent feature of data, max-margin classifiers, label information and Bayesian nonparametrics to determine the number of clusters in a united framework.Thanks to the conditionally conjugate property, the parameters in our model can be inferred via the simple and efficient Gibbs sampler. Finally, we test our methods on Benchmark datasets and measured radar HRRP data to demonstrate their efficiency and effectiveness.