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

Abstract: 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 interrelations 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.

Key words: global preserving, spectral method, kernel Principal Components Analysis, ISOmetric MAPping, Maximum Variance Unfolding