Time Series Regression Based on Relevance Vector Learning Mechanism

This paper presents a relevance vector learning mechanism for time series regression analysis. The relevance vector machine (RVM) has a probabilistic Bayesian learning framework and has good generalization capability. The RVM consists of the sum of product of weight and kernel function which projects input space into high dimensional feature space. Although having the same function form as support vector machine (SVM), the RVM not only has sparser solutions, but also has not restrictions on the selection of kernel functions. An improved EM-based learning algorithm of RVM is also proposed to deal with the problem of computing the inverse matrix in the classical RVM when the model is too sparse. As a case study, the EM-base RVM is used for functional regression with noises. The simulation results illustrate effectiveness of the presented improved RVM. Both the classical and EM-based RVM are superior to SVM, and EM-base RVM is more robust faced with different kernel width.