全局保持的流形学习算法对比研究  

Contrasting research of global preserving manifold learning algorithms

作　　者：曾宪华[1] 罗四维[2] 

机构地区：[1]重庆邮电大学计算机科学与技术研究所,重庆400065 [2]北京交通大学计算机与信息技术学院,北京100044

出　　处：《计算机工程与应用》2010年第15期1-6,89,共7页Computer Engineering and Applications

基　　金：国家自然科学基金No.60773016;重庆邮电大学博士启动基金项目(No.A2009-24)~~

摘　　要：全局保持的流形学习算法主要是基于保持高维观测空间和内在低维流形的全局几何特性。详细比较了全局保持的典型流形学习算法的特点及其相互之间的联系,标明了它们的优点与缺陷。实验说明这些方法发现的内在维数和内在低维流形的差异。最后提出了一些新的流形学习研究方向。Global preserving manifold learning algorithms are mainly based on preserving global geometric properties between high-dimensional observed space and intrinsic low-dimensional manifold.This paper compares in detail the characters and interre-lations of several classical manifold learning algorithms based on preserving global properties.Some advantages and shortcomings of these algorithms are shown.Experimental results demonstrate the differences of these algorithms about intrinsic dimension and intrinsic low-dimensional manifold.Finally,several new research directions of manifold learning are proposed.

关 键 词：全局保持 谱方法 核主成分分析 等度规映射 最大方差展开 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]