计算机工程与应用 ›› 2010, Vol. 46 ›› Issue (36): 29-33.DOI: 10.3778/j.issn.1002-8331.2010.36.008

• 研究、探讨 • 上一篇    下一篇

一种无约束凸规划多平面修正TWSVM

徐金宝1,业巧林2,业 宁3,吴美红1   

  1. 1.南京工程学院 计算机工程学院,南京 211167
    2.南京理工大学 计算机学院,南京 210094
    3.南京林业大学 信息技术学院,南京 210037
  • 收稿日期:2010-05-24 修回日期:2010-07-12 出版日期:2010-12-21 发布日期:2010-12-21
  • 通讯作者: 徐金宝

Unconstrained convex programming multi-surface modified twin support vector machine

XU Jin-bao1,YE Qiao-lin2,YE Ning3,WU Mei-hong1   

  1. 1.School of Computer Engineering,Nanjing Institute of Technology,Nanjing 211167,China
    2.School of Computer Science and Technology,Nanjing University of Science and Technology,Nanjing 210094,China
    3.School of Information Technology,Nanjing Forestry University,Nanjing 210037,China
  • Received:2010-05-24 Revised:2010-07-12 Online:2010-12-21 Published:2010-12-21
  • Contact: XU Jin-bao

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(Proximal SVM based on Generalized Eigenvalues,GEPSVM)。该算法将传统SVM问题分解为两个凸规划问题,使得训练速度缩减到原来的1/4。对TWSVM做了修正,基于新的优化准则设计了一种特殊TWSVM(GTWSVM),在此基础上,提出了快速GTWSVM(FGTWSVM),其将GTWSVM转换为无约束凸规划问题求解。该算法在保证得到与TWSVM相当的分类性能以及较快的计算速度的同时,还减少了输入空间的特征数以及内存占用。对于非线性问题,FGTWSVM可以减少核函数数目。

关键词: 对支持向量机(TWSVM), 近似支持向量机(GEPSVM), 多类问题, 无约束凸规划, 特征数, 核函数数目

Abstract: The optimization thought of TWSVM(Twin Support Vector Machine) originates from GEPSVM(Proximal SVM based on Generalized Eigenvalues).This algorithm divides traditional SVM problem into two convex programming ones,reduces training speed to the original 1/4.TWSVM is modified and a special TWSVM(GTWSVM) is designed based on a new optimization criterion.On this basis,fast GTWSVM(FGTWSVM) is proposed,which transforms GTWSVM to an unconstrained convex programming problem.This algorithm guarantees the same performance of TWSVM,faster computing speed,reduction of eigen-values number in input space and memory occupation.As far as non-linear problem is concerned,FGTWSVM requires fewer kernel function numbers.

Key words: Twin Support Vector Machine(TWSVM), Proximal SVM based on Generalized Eigenvalues(GEPSVM), multi-class problems, unconstrained convex plan, eigen-value number, kernel function number

中图分类号: