一种改进的渐进直推式支持向量机分类学习算法  

An Improved Learning Algorithm with Progressive Transductive Support Vector Machine

作　　者：廖东平[1] 魏玺章[1] 黎湘[1] 庄钊文[1] 

机构地区：[1]国防科技大学电子科学与工程学院 ATR 国家重点实验室,长沙410073

出　　处：《信号处理》2008年第2期213-218,共6页Journal of Signal Processing

基　　金：国防预研基金资助课题(41303040203)

摘　　要：基于支持向量机的直推式学习是统计学习理论中一个较新的研究领域。较之传统的归纳式学习方法而言,直推式学习往往更具有普遍性和实际意义。针对渐进直推式支持向量机学习算法存在的缺陷,提出了一种改进算法。该算法利用区域标注法取代前者的成对标注法,在继承了其渐进赋值和动态调整的规则的同时,提高了算法的速度;根据每个无标签样本的标注可信度自适应地对其赋予不同的影响因子,从而控制训练误差的传递和积累,提高了算法的性能。雷达实测数据实验结果表明该算法是有效的。Transductive inference based on support vector machine is a relatively new research region in statistical learning theory. Compared with traditional inductive learning, transductive inference is often more practical and can give results with better performance. In allusion to the shortcoming of progressive transductive support vector machine learning algorithm, an improved algorithm is presented in this paper. Using region labeling rule to substitute the former＇ s twin label rule, this algorithm inherits the gradually evaluate rule and dynamic adjust rule, and raises the speed at the same time. Every unlabeled example is adaptively given an effect factor according to it＇ s label reliability in this algorithm. That can control the transfer and accumulation of train error and improve the performance of this algorithm. Experiments with radar raw data show the validity of this algorithm.

关 键 词：统计学习理论(SLT) 直推式支持向量机(TSVM) 直推式学习 区域标注法 标注可信度 

分 类 号：TN915[电子电信—通信与信息系统] TP18[电子电信—信息与通信工程]