A new hybrid discriminative/generative model using the full-covariance multivariate generalized Gaussian mixture models

Discriminative models have been shown to be more advantageous for pattern recognition problem in machine learning. For this study, the main focus is developing a new hybrid model that combines the advantages of a discriminative technique namely the support vector machines (SVM) with the full efficiency offered through covariance multivariate generalized Gaussian mixture models (MGGMM). This new hybrid MGGMM applies the Fisher and Kullback–Leibler kernels derived from MGGMM to improve the kernel function of SVM. This approach is based on two different learning techniques explicitly: the Fisher scoring algorithm and the Bayes inference technique based on Markov Chain Monte Carlo and Metropolis–Hastings algorithm. These learning methods work with two model selection approaches (minimum message length and marginal likelihood) to determine the number of clusters. The effectiveness of the framework is demonstrated through extensive experiments including synthetic datasets, facial expression recognition and human activity recognition.