一种基于收缩超平面的支持向量分类算法  

A Novel Support Vector Classification Algorithm Based on Shrunk Hyperplane

作　　者：凌萍[1] 荣祥胜[2] 高大金[2] 

机构地区：[1]江苏师范大学计算科学科学与技术学院 [2]空军勤务学院训练部

出　　处：《小型微型计算机系统》2014年第12期2717-2726,共10页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61105129 61100167)资助;江苏省自然科学基金项目(BK20131130)资助

摘　　要：支持向量机(Support Vector Machine,SVM)的在二分问题上表现优异,而在多分问题上易受到基本分类器性能不稳定、集成方式依赖于具体问题等多种因素的影响,因此表现一般.鉴于此,提出一种基于收缩超平面的支持向量分类算法(A Novel Support Vector classification Algorithm based on Shrunk Hyperplane,SVASH).SVASH摈弃了二分思想,通过为各类分别构造穿过其密集分布区的超平面(命名为收缩的超平面),获知各类的鉴别性信息,并根据数据与收缩超平面的投影距离确定其类别.文中提出并证明了收缩超平面的几何性质,以此说明算法的有效性.文中设计了快速训练算法,以提高算法效率.实验表明,SVASH在多分问题上显示出优于同类算法的性能;在二分问题上也有接近最优性能的表现.In spite of excellent behaviors in binary-classification problems, in multi classification, Support Vector Machine （SVM） is caught by the unsteady performance, heavy dependence on concrete problems, etc. To address these problems, a novel Support Vector classification Algorithm based on Shrunk Hyperplane （ SVASH ） is proposed in this paper. In each class SVASH constructs the individ- ual hyperplane around which class-members are clustered （ named as the shrunk hyperplane）, and labels the query according to the projection distance between the query and the shrunk hyperplanes. SVASH completely discards the binary-classification idea but holds the all-addressed spirit. The geometric properties of shrunk hyperplanes are proposed and proofed to guarantees the validation of SVASH. SVASH is equipped with a fast training method to bring high efficiency. Empirical evidence on benchmark and real datasets indicates in multi-classification SVASH exhibits better performance than the peers, and in binary-classification it is of the competitive performance with the optimal results.

关 键 词：支持向量机 收缩的超平面 切割超平面 几何性质 快速训练算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]