集合分类中的鉴别式局部信息距离保持映射

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。

Discriminant local information distance preserving projection for set classification

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).