计算机工程与应用 ›› 2008, Vol. 44 ›› Issue (36): 24-28.DOI: 10.3778/j.issn.1002-8331.2008.36.006

• 博士论坛 • 上一篇    下一篇

新的稀疏支持向量回归机算法及实验研究

陈晓峰1,王士同1,曹苏群1,2,马培勇3   

  1. 1.江南大学 信息学院,江苏 无锡 214122
    2.淮阴工学院 机械系,江苏 淮安 223001
    3.中国科学技术大学 工程科学学院,合肥 230026
  • 收稿日期:2008-07-07 修回日期:2008-10-10 出版日期:2008-12-21 发布日期:2008-12-21
  • 通讯作者: 陈晓峰

Novel sparse Support Vector Regression and its experimental study

CHEN Xiao-feng1,WANG Shi-tong1,CAO Su-qun1,2,MA Pei-yong3   

  1. 1.School of Information,Jiangnan University,Wuxi,Jiangsu 214122,China
    2.Department of Mechanical Engineering,Huaiyin Institute of Technology,Huaian,Jiangsu 223001,China
    3.School of Engineering Science,University of Science and Technology of China,Hefei 230026,China
  • Received:2008-07-07 Revised:2008-10-10 Online:2008-12-21 Published:2008-12-21
  • Contact: CHEN Xiao-feng

摘要: 支持向量回归机是一种解决回归问题的重要方法,其预测速度与支持向量的稀疏性成正比。为了改进支持向量回归机的稀疏性,提出了一种直接稀疏支持向量回归算法DSKR(Direct Sparse Kernel Support Vector Regression),用于构造稀疏性支持向量回归机。DSKR算法对ε-SVR(ε-Support Vector Regression)增加一个非凸约束,通过迭代优化的方式,得到稀疏性好的支持向量回归机。在人工数据集和真实世界数据集上研究DSKR算法的性能,实验结果表明,DSKR算法可以通过调控支持向量的数目,提高支持向量回归机的稀疏性,且具有较好的鲁棒性。

关键词: 支持向量回归机, 核方法, 稀疏核学习

Abstract: Support Vector Regression is an important kind of method for regression problems.The predicting speed of Support Vector Regression is proportional to its sparseness.In order to increase sparseness of support vector regression,in this paper,a sparse support vector regression method named DSKR(Direct Sparse Kernel Support Vector Regression) is proposed to construct sparse Support Vector Regression.DSKR adds a non-convex constraint to ε-SVR (ε-Support Vector Regression),and then obtains Support Vector Regression with better sparseness using iterative optimization.Experimental comparisons are made with several Support Vector Regression methods on both synthetic data sets and real-world data sets,the comparisons show that the proposed DSKR gives promising results.It can improve sparseness and adjust number of support vectors with perfect robust performance.

Key words: Support Vector Regression(SVR), kernel method, sparse kernel learning