鲁棒的加权孪生支持向量机

来源期刊:中南大学学报(自然科学版)2015年第6期

论文作者:花小朋 丁世飞

文章页码:2074 - 2081

关键词:孪生支持向量机;局部几何结构;噪声点;鲁棒性;分类

Key words:twin support vector machine; local geometric structure; outliers; robust; classification

摘    要:基于局部信息的加权孪生支持向量机(WLTSVM)借用类内及类间近邻图分别表示类内样本的紧凑性和类间样本的分散性,克服孪生支持向量机(TWSVM)欠考虑训练样本间相似性的缺陷,并且在一定程度上降低二次规划求解的计算复杂度。然而,WLTSVM仍不能充分刻画类内样本潜在的局部几何结构,并且存在对噪声点敏感的风险。基于以上不足,提出一种鲁棒的加权孪生支持向量机(RWTSVM)。与WLTSVM相比,RWTSVM的优势在于:选用热核函数定义类内近邻图权值矩阵,可以更好地刻画类内样本潜在的局部几何结构及蕴含的鉴别信息;用类间近邻图选取边界点,同时结合类内近邻图使得超平面远离边界点中权重较大的样本,降低算法对噪声点敏感的风险。人造数据集和真实数据集上的测试结果验证算法RWTSVM的有效性。

Abstract: Weighted twin support vector machine with local information (WLTSVM), as a variant of twin support vector machine (TWSVM), uses within-class graph and between-class graph to characterize the intra-class compactness and the inter-class separability, respectively. This makes WLTSVM improve the generalization capability of TWSVM by mining as much underlying similarity information within samples as possible and reduces the time complexity of TWSVM by reducing the support vectors for each class. Despite these advantages, WLTSVM can not fully reflect the local geometry manifold within samples because of using the within-class graph whose weight matrix is defined simply. Moreover, WLTSVM ignores the possible outliers because it chooses boundary points in the contrary class to construct the constraints. Thus a novel method, robust weighted twin support vector machine (RWTSVM) was proposed, which can better characterize the underlying local geometric structure and the descriminant information by using a hot kernel function to define the weight matrix of within-class graph, and reduce the influence of outliers by considering within-class weight of the boundary points used to construct constraints in WLTSVM. The experimental results on the artificial and real datasets indicate the effectiveness of RWTSVM method.

有色金属在线官网  |   会议  |   在线投稿  |   购买纸书  |   科技图书馆

中南大学出版社 技术支持 版权声明   电话:0731-88830515 88830516   传真:0731-88710482   Email:administrator@cnnmol.com

互联网出版许可证:(署)网出证(京)字第342号   京ICP备17050991号-6      京公网安备11010802042557号