Research And Analysis Of Multi-class Support Vector Machine

In the field of machines learning, Support Vector Machines (SVM) has always been a most important technology. It has been very successful in the past ten years and has made a lot of good applications in many domains including text classification, face recognition etc. SVM's biggest discrimination from other machines learning methods is that it is based on the statistical learning theory, such as structure risk minimization. The fast development of SVM depends on the combination of kernel function and SVM. SVM is originally designed for the two-class problems. But in the real life, multi-class problem is prevailing. Therefore, how to solve multi-class problem with two-class SVM is one of the most important research works which make SVM be used effectively.This dissertation researches and analyzes the theory and performance of existing multi-class SVMs and describes the classification principle of one-class-SVM. After that, two multi-SVM improvement methods are proposed. One of them combines one-class SVM with two-class-SVM. It is a method which trains dataset with one-class-SVM and revise the classification with two-class-SVM. The other method minimizes the region in which data can not be classified correctly in one-versus-one SVM. It is a method which trains the original dataset with one-versus-one method and trains the dataset which can not be classified correctly by one-versus-one method with one-versus-rest method again to reduce the region in which data can not be classified.After the proposal, we have done the following work:(1)Firstly, we implement the two new methods on the libsvm platform.(2)Secondly, we choose training and predicting dataset and experiment with these new methods. Compare the experiment results with the original method.(3)Thirdly, we get the conclusion that these two new methods are effective through analyzing.