On Local Kernel Polarization

摘要：

The problem of evaluating the quality of a kernelfunction for a classification task is considered.Drawn from physics,kernel polarization wasintroduced as an effective measure for selectingkernel parameters,which was previously donemostly by exhaustive search.However,it only takesbetween-class separability into account but neglectsthe preservation of within-class local structure.The ‘globalityof the kernel polarization may leave lessdegree of freedom for increasing separability.In thispaper,we propose a new quality measure calledlocal kernel polarization,which is a localizedvariant of kernel polarization.Local kernelpolarization can preserve the local structure of thedata of the same class so the data can be embeddedmore appropriately.This quality measure isdemonstrated with some UCI machine learningbenchmark examples.

作者: Tinghua Wang Shengfeng Tian Houkuan Huang

作者单位: School of Computer and Information Technology Beijing Jiaotong University,Beijing 100044,PR China

会议类型: 国际会议

会议名称: 2008 3rd International Conference on Intelligent System and Knowledge Engineering(第三届智能系统与知识工程国际会议)(ISKE 2008)

会议地点: 厦门

会议语种:英文

页码: 1304-1309

在线出版日期: 2008-11-17（万方平台首次上网日期，不代表论文的发表时间）