文献类型：期刊文献

英文题名：Selection of the neighborhood size for manifold learning based on Bayesian information criterion

作者：Shao, Chao[1]; Wan, Chunhong[1]

第一作者：邵超

通讯作者：Shao, C.|[10484123a513c99]邵超;

机构：[1] School of Computer and Information Engineering, Henan University of Economics and Law, Zhengzhou 450002, China

第一机构：河南财经政法大学计算机与信息工程学院

通讯机构：[1]School of Computer and Information Engineering, Henan University of Economics and Law, Zhengzhou 450002, China|[1048412]河南财经政法大学计算机与信息工程学院;[10484]河南财经政法大学;

年份：2012

卷号：8

期号：7

起止页码：3043-3050

外文期刊名：Journal of Computational Information Systems

收录：EI(收录号：20122215069884);Scopus(收录号：2-s2.0-84861418898)

语种：英文

外文关键词：Learning algorithms - Principal component analysis

摘要：To select a suitable neighborhood size for manifold learning algorithms efficiently, this paper presented a new method based on Bayesian Information Criterion (BIC). Due to the locally Euclidean property of the manifold, the Principal Component Analysis (PCA) reconstruction errors of all the neighborhoods in the neighborhood graph remain small and fall into one cluster when the neighborhood size is suitable; however, once the neighborhood size becomes unsuitable, the PCA reconstruction errors of the neighborhoods with shortcut edges become rather large, which makes all the PCA reconstruction errors fall into two distinct clusters, which can be detected by BIC. This method only requires running weighted PCA and computing BIC incrementally, which makes it more efficient than those methods based on residual variance. The effectivity of this method can be verified by experimental results well. ? 2012 Binary Information Press.