Half-against-half multi-class support vector machines

A Half-Against-Half (HAH) multi-class SVM is proposed in this paper. Unlike the commonly used One-Against-All (OVA) and One-Against-One (OVO) implementation methods, HAH is built via recursively dividing the training dataset of K classes into two subsets of classes. The structure of HAH is same as a decision tree with each node as a binary SVM classifier that tells a testing sample belongs to one group of classes or the other. The trained HAH classifier model consists of at most K binary SVMs. For each classification testing, HAH requires at most K binary SVM evaluations. Both theoretical estimation and experimental results show that HAH has advantages over OVA and OVO based methods in the evaluation speed as well as the size of the classifier model while maintaining comparable accuracy.