文献类型：期刊文献

英文题名：Selection of the suitable parameter value for ISOMAP

作者：Jing, Li[1]; Shao, Chao[1]

第一作者：景丽

通讯作者：Jing, L.

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

年份：2011

卷号：6

期号：6

起止页码：1034-1041

外文期刊名：Journal of Software

收录：EI(收录号：20113414255710);Scopus(收录号：2-s2.0-84860421123)

语种：英文

外文关键词：Computation theory - Data visualization - Dimensionality reduction - Graph algorithms - Graph theory - Visualization

摘要：As a promising dimensionality reduction and data visualization technique, ISOMAP is usually used for data preprocessing to avoid "the curse of dimensionality" and select more suitable algorithms or improve the performance of algorithms used in data mining process according to No Free Lunch (NFL) Theorem. ISOMAP has only one parameter, i.e. the neighborhood size, upon which the success of ISOMAP depends greatly. However, it's an open problem how to select a suitable neighborhood size efficiently. Based on the unique feature of shortcut edges, introduced into the neighborhood graph by using the unsuitable neighborhood size, this paper presents an efficient method to select a suitable neighborhood size according to the decrement of the sum of all the shortest path distances. In contrast with the straightforward method with residual variance, our method only requires running the former part of ISOMAP (shortest path computation) incrementally, which makes it less time-consuming, while yielding the same results. Finally, the feasibility and robustness of this method can be verified by experimental results well. ? 2011 ACADEMY PUBLISHER.