Research Of Multi-class Classification Methods Based On Support Vector Machine

Support vector machine is a novel machine learning method proposed by Vapnik and his group at AT&T Bell laboratory for classification and regression questions. It is a new tool to solve problems of machine learning based on optimization technique. Support vector machine based on statistical learning theory and structural risk minimization principle has good extension and better accuracy. It characterized by the use of optimal classification hyperplane, technique of kernels and convex quadratic programming can solve some problems such as over study, dimension disaster and local minimization effectively. Support vector machine has became research hotspot for its excellent studying performance in machine learning domain, and is widely applied in much regions such as pattern recognition, regression estimation and so on. Traditional support vector machine is developed for binary classification problems. How to extend it for multi-class classification is a significant issue.In this paper, the main work researched is as follows:1. Machine learning, statistical learning theory, the development and research actuality of support vector machine are introduced. The theory and algorithm of support vector machine are expounded. And the theory of kernel function, parameters selection and other hotspot issues are discussed simultaneously.2. Feature selection of support vector machine on binary classification is investigated. A new feature selection method of improving the performance of support vector machine is proposed. The validity of the method is proved by experiments.3. Common multi-class classification methods including one versus rest, one versus one, binary tree, decision directed acyclic graph and error correcting output code are summarized. The advantage and disadvantage of these methods are compared. Aiming at the shortcoming of binary tree support vector machine, new binary trees are established to improve the decision speed and the accuracy of multi-classifier based on the effect of distribution of classes to inter-class separability and hierarchical clustering. The efficiency of improved methods are proved by results of experiment.