TheLLEandalinearmapping - 模式识别国家重点实验室.pdfVIP

Pattern Recognition 39 (2006) 1799–1804 /locate/patcog Rapid and brief communication The LLE and a linear mapping ∗ F.C. Wu , Z.Y. Hu National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, P.O. Box 2728, Beijing 100080, People’s Republic of China Received 23 September 2005; accepted 30 March 2006 Abstract The locally linear embedding (LLE) is considered an effective algorithm for dimensionality reduction. In this short note, some of its key properties are studied. In particular, we show that: (1) there always exists a linear mapping from the high-dimensional space to the low-dimensional space such that all the constraint conditions in the LLE can be satisfied. The implication of the existence of such a linear mapping is that the LLE cannot guarantee a one-to-one mapping from the high-dimensional space to the low-dimensional space for a given data set; (2) if the LLE is required to globally preserve distance, it must be a PCA mapping; (3) for a given high-dimensional data set, there always exists a local distance-preserving LLE. The above results can bring some new insights into a better understanding of the LLE. 2006 Pattern Recognition Society. Published by Elsevier Ltd. All rights reserved. Keywords: Locally linear embedding (LLE); Linear mapping; Principal component analysis (PCA) 1. Introduction