摘要
对支持向量机(twin support vector machine,TWSVM)的优化思想源于基于广义特征值近似支持向量机(proxi mal SVMbased on generalized eigenvalues,GEPSVM),问题解归结为求解两个SVM型问题,因此,计算开销缩减到标准SVM的1/4.除了保留了GEPSVM优势外,在分类性能上TWSVM远优于GEPSVM,但仍需求解凸规划问题,并且,目前尚无有效的TWSVM的特征提取算法提出.首先,向TWSVM模型中引入正则项,提出了正则化TWSVM(RTWSVM).与TWSVM不同,RTWSVM保证了该问题为一个强凸规划问题.在此基础上,构造了TWSVM的特征提取算法(FRTWSVM).该分类器只需求解一个线性方程系统,无需任何凸规划软件包.在保证得到与TWSVM相当的分类性能以及较快的计算速度上,此方式还减少了输入空间的特征数.对于非线性问题,FRTWSVM可以减少核函数数目.
Twin support vector machine（TWSVM for short）,as a variant of GEPSVM,is based on the idea of generalized support vector machine（GEPSVM）,which determines two nonparallel planes by solving two related SVM-type problems,such that its computing cost in the training phase is 1 4 of standard SVM.In addition to keeping the superior characteristics of GEPSVM,classification performance of TWSVM significantly outperforms that of GEPSVM.However,the stand-alone method requires the solution of two smaller quadratic programming problems（QPPs）and there are few modifications of it that have been proposed to automatically select the input features.In this paper,through introducing a regularization term to the objective functions of TWSVM pair,the authors first propose a modification of TWSVM,called RTWSVM.Entirely different formulation from TWSVM,RTWSVM guarantees two QPPs are strong convex,implying that it can obtain the global but unique solution.Then,a feature selection method for RTWSVM is proposed,which is based on RTWSVM,and generates two planes directly from solving two sets of linear equations and requires no special optimization solvers.This method can obtain comparable classification performance to TWSVM,and have faster computational time,better suppression of input features and better ability to reduce the number of kernel functions.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2011年第6期1029-1037,共9页
Journal of Computer Research and Development
基金
国家自然科学基金项目(90820306)
江苏省自然科学基金项目(BK2009393)
关键词
TWSVM
GEPSVM
RTWSVM
无约束凸规划
特征镇压
twin support vector machine（TWSVM）
generalized support vector machine（GEPSVM）
regularized TWSVM（RTWSVM）
unconstrained convex programming
suppression of input features
