Incremental Subclass Support Vector Machine

Abstract

Support Vector Machine (SVM) is a very competitive linear classifier based on convex optimization problem, were support vectors fully describe decision boundary. Hence, SVM is sensitive to data spread and does not take into account the existence of class subclasses, nor minimizes data dispersion for classification performance improvement. Thus, Kernel subclass SVM (KSSVM) was proposed to handle multimodal data and to minimize data dispersion. Nevertheless, KSSVM has difficulties in classifying sequentially obtained data and handling large scale datasets, since it is based on batch learning. For this reason, we propose a novel incremental KSSVM (iKSSVM) which handles dynamic and large data in a proper manner. The iKSSVM is still based on convex optimization problem and minimizes data dispersion within and between data subclasses incrementally, in order to improve discriminative power and classification performance. An extensive comparative evaluation of the iKSSVM to batch KSSVM, as well as, other contemporary incremental classifiers, on real world datasets, has shown clearly its superiority in terms of classification accuracy.