基于集成图的保局投影算法

Graphs ensemble based locality preserving projection

为了克服保局投影方法(locality preserving projection,LPP)对噪音敏感,有效性依赖于近邻图构造等缺点,提出一种基于集成图的保局投影方法(graphs ensemble based LPP,GELPP)。该方法先根据鲁棒统计原理定义出对噪声鲁棒的样本间相似性度量,再以该度量为基础构造多个近似的最大生成树;然后利用集成学习泛化能力强的优点来组合多个树为一个集成图;最后通过替换LPP的近邻图和相似性度量来进行保局投影。在高维人脸图像上的降维实验结果表明,该方法对噪声鲁棒,以及在集成图上降维的有效性。

Locality preserving projection（LPP） is sensitive to noise and its effectiveness relies much on the construction of neighborhood graph.To tackle these problems,a graphs ensemble based locality preserving projection（GELPP） method is proposed.GELPP first derives robust similarity between samples via robust statistic.Secondly,it constructs several approximate maximum spanning trees（MST） via this similarity,then combines these trees into an ensemble graph by the virtues of strong generalization ability of ensemble learning.Finally,it replaces similarity metric and neighborhood graph of LPP with robust similarity and this ensemble graph respectively.Experiments on high dimensional face databases prove not only GELPP’s robustness to noise and outliers,but also its effectiveness in dimensionality reduction on graphs ensemble.