基于可靠性测度的多类支持向量机

多类分类问题是我们经常遇到的问题,常用的方法是将多类问题转化为若干个二类问题,然后利用二类支持向量机(support vector machine,SVM)进行分类,如一对余SVM,一对一SVM,决策树SVM等.在这些方法中,大都没有考虑所生成的多个分类器之间的可靠性和重要性问题.为了改进这一点,本文以一对余SVM为例,提出了两种基于可靠性测度的多类分类算法,算法的思想可用于一对一SVM,决策树SVM等其他多种分类器中.为了检验所提算法的有效性,本文进行了比较试验,实验结果表明所提算法不仅提高了分类准确度,而且具有更为广泛的推广能力.

Multi-class SVMs Based On Reliablity Measures

Multi-class classification problems are often encountered in the practical application. A common method that we deal with a multi-class classification problem is to translate it into some binary classification problems and then uses binary support vector machine （SVM）, such as one-against-all SVM, one-against-one SVM, decision tree SVM and so on. However, the reliability and significance a- mong multiple classifiers produced by using the kind of methods are not considered. In order to improve this point, this paper takes one-against-all SVM for example and proposes two multi-class algorithms based on reliability measures. The idea in this paper can be used for OAO-SVM, decision tree SVM and so on. We perform the comparing experiments to test the effectiveness of our algorithms. Experimental results show that our algorithms not only can improve the classification accuracy but also have a goodr generalization ability.