改进的一对一支持向量机多分类算法  

Improved multi-classification algorithm of one-against-one SVM

作　　者：单玉刚[1,2,3,4] 王宏[1,2] 董爽[5] 

机构地区：[1]中国科学院沈阳自动化研究所 [2]沈阳中科博微自动化有限公司 [3]空军93303部队 [4]中国科学院研究生院 [5]北京联合大学管理学院

出　　处：《计算机工程与设计》2012年第5期1837-1841,共5页Computer Engineering and Design

基　　金：国家863高技术研究发展计划基金项目(2007AA041407)

摘　　要：支持向量机的一对一多分类算法具有良好的性能,但该算法在分类时存在不可分区域,影响了该方法的应用。因此,提出一种一对一与基于紧密度判决相结合的多分类方法,使用一对一算法分类,采用基于紧密度决策解决不可分区,依据样本到类中心之间的距离和基于kNN(k nearest neighbor)的样本分布情况结合的方式构建判别函数来确定类别归属。使用UCI(university of California Irvine)数据集做测试,测试结果表明,该算法能有效地解决不可分区域问题,而且表现出比其它算法更好的性能。Multi-class classification algorithm of one-against-one SVM show good performance,but the algorithm exists an unclassifiable region,which affects the application effect of the algorithm.Hence,a multi-classification algorithm of integration of one-against-one and affinity decision is presented.Firstly,the one-against-one multi-class classification algorithm is used to classify samples,and then the affinity decision is used to solve samples in the unclassifiable region and to determine categories of samples,which using the approach of distance between the sample and centers of classes and sample distribution based on kNN（k nearest neighbor） to create decision function.By adopting UCI data sets for testing,the results show that the algorithm can solve unclassifiable region issues,and show better performance than other algorithms.

关 键 词：K近邻 一对一支持向量机 多分类 不可分区 紧密度 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]