A New Boosting Multi-class SVM Algorithm

Abstract

Support Vector Machines (SVM) have originally designed for binary classification problems. However, Multi-class SVMs (MCSVM) are implemented by combining several binary SVMs. This paper presents a new boosting Multi-class SVMs (BmSVM) to overcome computational complexity of existing construction MCSVM methods. The other two objectives of the paper are: first, to show the robustness of BmSVM against different constructing Multi-class SVM methods such as One-Against-All, One-Against-One; Second, to compare the performance and complexity of BmSVM against SMO, AdaBoost, Decision Tree, and MCSVM. The simulation results demonstrate that the BmSVM on hypothyroid dataset with polynomial kernel is superior to the others.

Keywords: Boosting Multi-class SVM, Multi-class SVM , Boosting Algorithm, Boosting Multi-class SVM (BmSVM), SVM