A fast algorithm for relevance vector machine

Abstract

This paper presents a fast algorithm for training relevance vector machine classifiers for dealing with large data set. The core principle is to remove dependent data points before training a relevance vector machine classifier. The removal of dependent data points is implemented by the Gram-Schmidt algorithm. The verification using one group of toy data sets and three benchmark data sets shows that the proposed fast relevance vector machine is able to speed up the training time significantly while maintaining the model performance including testing accuracy, model robustness and model sparseness. © Springer-Verlag Berlin Heidelberg 2006.