局部敏感判别直推学习机  

Locality sensitive discriminant transductive learning

作　　者：潘俊[1,2] 孔繁胜[1] 王瑞琴[2] 

机构地区：[1]浙江大学计算机科学与技术学院,浙江杭州310027 [2]温州大学物理与电子信息工程学院,浙江温州325035

出　　处：《浙江大学学报（工学版）》2012年第6期987-994,共8页Journal of Zhejiang University：Engineering Science

基　　金：浙江省自然科学基金资助项目(Q12F020047)

摘　　要：为了充分挖掘样本内在的几何结构和蕴含的判别信息来指导样本数据分类,提出一种局部敏感的判别直推学习方法.该方法将局部敏感辨析(LSDA)的基本原理引入到直推学习中,在直推学习的正则化框架中同时引入有助于分类的样本局部结构信息和判别信息,在判别信息指导下构建了类内图和类间图来刻画类内紧性和类间散性,从而在每个局部邻域中进一步最大化类间样本的间隔.同时,用数学的形式给出了目标函数的解析表达,在几个典型数据集上的实验结果表明,相较传统的基于图的半监督学习算法,该方法能取得更高的分类效果.A novel approach called locality sensitive discriminant transductive learning（LSD-TL） was developed in order to utilize the underlying geometric structure and the discriminant information to full extent in the process of data classification.Inspired by the basic theories of the locality sensitive discriminant analysis（LSDA）,LSD-TL directly incorporates the discriminative information as well as the local geometry of the data space into the regularization framework of transductive learning.LSD-TL constructs the within-class graph and the between-class graph under the guidance of the discriminative information,so that the margins between the data of different classes in each local manifold are maximized.In addition,the analytical solution for the problem was obtained.Experimental results on real world datasets demonstrated that compared with several traditional graph-based semi-supervised algorithms,the new approach improved the classification accuracy.

关 键 词：局部敏感辨析 直推学习 图方法 正则化 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]