Discriminant Graph Based Linear Embedding

Bo Li, Jin Liu, Wenyong Dong, Wensheng Zhang

Abstract: LLE is a nonlinear dimensionality reduction method, which has been successfully applied to data visualization. Based on the assumption of local linearity, LLE can compute the weights between the KNN nodes using the local least reconstruction errors, which increase the computational cost. In this paper, a method titled Discriminant Graph Based Linear Embedding (DGBLE) is proposed to set the weights between the nodes in the KNN graph directly to reduce the computational expense. Moreover, label information can also be taken into account to improve the discriminant power of the original LLE. Experiments on some benchmark data show that the proposed method is feasible and effective.