A novel multi-class SVM classifier based on DDAG

Abstract

Presents a method of constructing a multi-class SVM classifier, which is based on the structure of a decision directed acyclic graph (DDAG) and using an active constraint for each SVM classifier. For a k-class problem, it combines k(k-1)/2 two-class SVM classifiers, one for each pair of classes. In order to speed up the training and decision process of the classifier, we make three changes to the standard two-class classifiers, ie. large margin, 2-norm squared for the error for the soft margin and active constraint While in the testing phase, we use a rooted binary directed acyclic graph which has k(k-1)/2 internal nodes and k leaves. A computational experiment indicates that this is a simple and fast approach to generating multi-class SVM classifiers.