An Improved Hybrid Structure Multi-classification Support Vector Machine

Abstract

In order to improve the speed of multi-class support vector machine, based on One-versus-One SVM, the method of combining hierarchical classification is proposed which can reduce the number of classifiers during training and testing, and use the inter-class separation degree, the intra-class sample distance, and the intra-class sample distance standard deviation as the classification measures to divide the subset of binary classification and then form the binary tree structure. Finally, the 1-v-1 training is performed on the subclasses respectively. Experiments show that compared with the traditional 1-v-1 SVM, this method can effectively shorten the time required for classification and reduce the influence of error accumulation of H-SVMs.