Abstract

Support vector machines (SVMs) have been the predominate approach to kernel-based classification. While SVMs have demonstrated excellent performance in many application domains, they are known to be sensitive to noise in their training dataset. Motivated by the equalizing effect of bagging classifiers, we present a novel approach to kernel-based classification that we call microbagging. This method bags all possible maximal-margin estimators between pairs of training points to create a novel linear kernel classifier with weights defined directly as functions of the pairwise distance matrix induced by the kernel function. We derive relationships between linear and distance-based classifiers and empirically compare microbagging to the SVMs and robust SVMs on several datasets.