An online approach for learning non-Gaussian mixture models with localized feature selection

This paper presents an online algorithm for mixture model-based clustering. Online algorithms allow data points to be processed sequentially, which is critical for real-time and large-scale applications. The proposed online algorithm is based on finite generalized Dirichlet (GD) mixtures together with a unsupervised localized feature selection scheme. By learning the proposed model in an online manner through a variational inference framework, all the involved model parameters, number of components and features weights are estimated simultaneously in closed forms. Additionally, the problem of overfitting is avoided due to the nature of Bayesian learning. The proposed method is validated by both synthetic data and an application concerning the online automatic image annotation.