The supervised local linear embedding algorithm (SLLE) maps the high dimensional data in the low dimensional feature space through the label information of the data points. In the process of homogenizing the high dimensional data distribution and minimizing the reconstruction cost and for the situation that the sparse discrete data ignored in-class deviations from the population distribution may be incorrectly projected in other hyperplanes during the linear reconstruction, the Kmeans ++ algorithm is introduced to adjust the distance between the samples, and the selection of the optimal neighbor points making the data more efficiently reflect the actual distribution in the high-dimensional space, so that the reduced dimension of the data has better separability. Through the simulation of ORL and Yale data set, the proposed method has stronger generalization ability and higher recognition rate.