Robust Locally Linear Embedding and Application in High Dimensional Data

RLLE (robust locally linear embedding) is presented in this paper, which overcomes some deficiencies of LLE (locally linear embedding) such as sensitivity to noise and random selection of the neighborhoods. Some examples are given to compare RLLE with LLE. Experiments show that RLLE is insensitive to noise while discovering the intrinsic structure more clearly. Compared with other techniques of data manifolds, whose subjects are to remove noise, RLLE makes the most of the data local structure and unites reduction and noise removing. Besides, the neighborhood ball method, which RLLE uses to choose the neighborhood, can be transplanted into other nonlinear reductions.