中文摘要：

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.

英文摘要：

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.