集合分类中的鉴别式局部信息距离保持映射  

Discriminant local information distance preserving projection for set classification

作　　者：何亮[1] 栗志意[1] 蔡猛[1] 刘加[1] 

机构地区：[1]清华大学电子工程系清华信息科学与技术国家实验室,北京100084

出　　处：《清华大学学报（自然科学版）》2011年第7期1010-1016,共7页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(90920302;61005019);国家"八六三"高技术项目(2008AA040201)

摘　　要：该文提出鉴别式局部信息距离保持映射,以解决一类集合分类问题。鉴别式局部信息距离保持映射假设集合所对应的概率密度分布位于统计流形上,选取Fisher信息距离作为概率密度分布间的距离,并将最小化同类点的信息距离、最大化异类近邻点的信息距离作为目标函数,利用特征值分解的方法,求解线性映射矩阵。基于美国国家标准技术署于2008年公布的说话人识别数据库的实验结果表明:鉴别式局部信息距离保持映射优于无用分量投影和鉴别式无用分量投影。A discriminant local information distance preserving projection（DLIDPP） was developed for the set classification problem.With the DLIDPP assumption,the hidden probability sample distributions lie on a statistical manifold.The Fisher information distance is used as the distance measurement on the statistical manifold.The goal of the DLIDPP is to minimize the information distance between the same class sets as well as to maximize the information distance between neighbouring sets of different classes.A linear mapping matrix is obtained by solving an eigenproblem.Tests on a speaker recognition evaluation data corpus released by the American National Institute of Standards and Technology in 2008 show that the DLIDPP has better recognition than nuissance attirbute projection（NAP） or discriminant nuisance attribute projection（DNAP）.

关 键 词：集合分类 流形学习 信息距离 局部保持映射 说话人识别 无用分量投影 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]