一种快速的支持向量机多类分类算法  

Fast Multi-class Classification Algorithm of Support Vector Machines

作　　者：秦玉平[1] 罗倩[1] 王秀坤[2] 王春立[3] 

机构地区：[1]渤海大学信息科学与工程学院,锦州121000 [2]大连理工大学电子与信息工程学院,大连116024 [3]大连海事大学信息科学技术学院,大连116026

出　　处：《计算机科学》2010年第7期240-242,共3页Computer Science

基　　金：国家自然科学基金项目(60603023);国家基础研究重大项目(973)研究专项(2001CCA00700)资助

摘　　要：提出了一种快速的支持向量机多类分类算法。首先用每类训练样本的样本数作为权值构造最优二叉树,然后对每个非叶子结点训练两类分类器。分类时,从二叉树根结点开始逐层向下分类,直到某一叶子结点,该结点对应的类别即为待分类样本的类别。在Reuters 21578标准数据集上进行的分类实验表明,该算法具有较好的性能,在一定程度上克服了现有的支持向量机多类分类算法分类速度较慢的缺点,尤其在类别数较多、各类样本规模相同的情况下,采用该算法能够较大幅度地提高分类速度。A fast support vector machines multi-class classification was proposed. Firstly, the number of every class training samples is used as weight to construct Huffman binary tree, and then train sub-classifiers for every non leaf node in the binary tree. For the sample to be classified, the sub-classifiers that between the root and a certain left are used to classify, the left is the class of the sample. The classification experiments on the Reuters 21578 database are done with this algorithm. The experimental results show that it has better performance and partly overcomes the flaw of existing multi-class classification algorithm of support vector machines, which is slow in the process of classification. This algorithm can remarkably increase the speed of classification, especially in the case of more classes and the scale of every class is uniform.

关 键 词：支持向量机 多类分类 二叉树 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]