A new method for multi-class support vector machines by training least number of classifiers

How to well apply Support Vector Machine (SVM) technique to multi-class classification problem is an important topic in the area of machine learning. In this paper, we propose a novel method which is different from all the existing ones. By constructing the least number of classifiers, it makes better use of the feature space partition, and can fully eliminate the unclassifiable region. The method is specially designed for 2^k-class problems first and could be possibly extended further. We compare the proposed method with several existing ones as one-against-rest (OAR), one-against-one (OAO), decision directed acyclic graph (DDAG), and decision tree (DT) based architecture. Experimental results exhibit good feasibility of the proposed model in term of generalization capability, training time and testing time.