Study of Multiple-Kernel Relevance Vector Machine Based on Kernel Alignment

Abstract

The relevance vector machine (RVM) was a Bayesian framework for learning sparse regression models and classifiers, it used single kernel function to map training data from low dimension sample space to high dimension feature space. The prediction accuracy and generalization of traditional single-kernel RVM (sRVM) were not ideal both in classification and regression, so we constructed homogeneous and heterogeneous multiple kernels function (MKF) by kernel function combination in which we testified the validity of basic kernel function (BKF) and its parameters we employed by kernel alignment (KA), then we acquired optimized multiple-kernel RVM (mRVM). Experiment results on LIBSVM datasets not only indicate that both homogeneous and heterogeneous multiple-kernel RVM we constructed possess lower error rate in classification and smaller root mean square (RMS) in regression than single-kernel RVM, but also prove the effectiveness of kernel alignment.