Joint Feature Selection and Classification Design Based on Variational Relevance Vector Machine
一种基于变分相关向量机的特征选择和分类结合方法

The relevance vector machine (RVM) is a fully probabilistic model with an equivalent functional form as the support vector machine (SVM), which can give posterior distributions over all parameters through the variational Bayesian (VB) method. Moreover, the RVM with linear kernel can realize not only classification but also linear feature selection by imposing sparsity in feature space where data is originally represented. In this paper, a joint feature selection and classification design is proposed based on the traditional VB-RVM. In the proposed framework, the Probit model is utilized to connect the regression problem with the classification problem, and the feature dimension extension by power transformation can make full use of the samples form the nonlinear classification boundary, and can realize nonlinear feature selection as well. The experiments based on the synthetic data and measured data demonstrate the practicability and effectiveness of the proposed method.