无偏置ν-SVM分类优化问题研究  

Study on ν-svm for Classification Optimization Problem without Bias

作　　者：丁晓剑[1] 赵银亮[1] 

机构地区：[1]西安交通大学电子与信息工程学院,西安710049

出　　处：《电子与信息学报》2011年第8期1998-2002,共5页Journal of Electronics & Information Technology

基　　金：国家863计划项目(2008AA01Z136)资助课题

摘　　要：在高维空间中,分类超平面倾向于通过原点,即不需要偏置(b)。为了研究在ν-SVM分类问题中是否需要b,该文提出了无(b)的ν-SVM的对偶优化问题并给出了其优化问题求解方法。该方法通过有效集策略将对偶优化问题转化为等式约束子优化问题,然后通过拉格朗日乘子法将子优化问题转化为线程方程组来求解。实验表明偏置(b)的存在会降低ν-SVM的泛化性能,ν-SVM只能得到无(b)ν-SVM的次优解。In the high-dimensional space,the classification hyperplane tends to pass through the origin and bias（b） is not need.To study whether ν-svm for classification needs（b）,dual optimization formulation of without（b） is proposed and the corresponding method of solving the optimization formulation is presented.The dual optimization formulation is transformed into equality constraint sub-optimization formulation by the active set strategy in this method,then the sub-optimization formulation is transformed into the linear equation by lagrange multiplier method.The experimental results show that the existence of（b） would reduce the generalization ability of ν-svm and ν-svm can only obtain the sub-optimal solution of ν-svm without b.

关 键 词：V-支持向量机 偏置 泛化性能 有效集 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]