In recent years several methods for the analysis of nonlinear data sets have
been proposed, including Locally Linear Embedding (LLE) [1]. LLE has already
been successfully applied to many problems, including face recognition [2], pre-
diction of membrane protein types [3] and the analysis of micro array data [4].
The algorithm assumes linearity in the local area centered on each data point.
Each area is mathematically characterized by a set of coecients (weights) which
correlate the particular data point with its n nearest neighbors. The aggrega-
tion of all areas can be intuitively thought as an assemblage of linear patches
which approximates the nonlinear data structure. The high-dimensional data is
then projected into a lower-dimensional space while preserving the coecients
between neighboring data points.