Landmark-Based Local Patches Alignment Embedding

Abstract

A novel embedding method, called landmark-based local patches alignment embedding (LLPA), is proposed. LLPA first searches a set of landmarks which preserve the global structure of data set well and constructs overlapping patches based on these landmarks. Then, global isometric mapping and multidimensional scale are applied respectively to derive the low-dimensional coordinates of the landmarks and local patches. Finally, we yield the resulting global coordinates by patches alignment technique combined with a set of landmarks in low-dimensional space as reference points.