A probabilistic multi-class strategy of one-vs.-rest support vector machines for cancer classification

Abstract

Support vector machines (SVMs), originally designed for binary classification, have been applied for multi-class classification with effective decomposition and reconstruction schemes. Decomposition schemes such as one-vs.-rest (OVR) and pair-wise partition a dataset into several subsets of two classes so as to produce multiple outputs that should be combined. Majority voting or winner-takes-all is a representative reconstruction scheme to combine those outputs, but it often causes some problems to consider tie-breaks and tune the weights of individual classifiers. In this paper, we propose a novel method in which SVMs are generated with the OVR scheme and probabilistically ordered by using the naïve Bayes classifiers (NBs). This method is able to break ties that frequently occur when working with multi-class classification systems with OVR SVMs. More specifically, we use the Pearson correlation to select informative genes and reduce the dimensionality of gene expression profiles when constructing the NBs. The proposed method has been validated on several popular multi-class cancer datasets and produced higher accuracy than conventional methods.