An Improved Hierarchical Multi-class Support Vector Machine with Binary Tree Architecture

Abstract

A hierarchical binary tree multi-class support vector machine (BTMSVM) based on class similarity in feature space is improved to overcome the drawbacks such as unclassifiable region which the existent methods have. The class similarity which considers class distance and distribution sphere in feature space is used to determine the classification order of hierarchical multi-class SVM. The learning samples and corresponding SVM sub-classifier are selectively re-constructed to make sure as bigger as classification margin, as much as generalization ability. The results of simulated experiments show that the proposed method is faster in training and classifying, better in classification correctness and generalization.