As a manifold reduced dimensionality technique, neighborhood preserving discriminant embedding (NPDE) and its variant version have been proposed recently. But NPDE and its variant version have the so-called small-sample-size (SSS) problem. In this paper, the NPDE method is taken as the representative and an exponential neighborhood preserving discriminant embedding (ENPDE) is proposed to address the SSS problem. The main idea of ENPDE is that the matrix exponential is introduced to NPDE. ENPDE has two superiorities. First, ENPDE avoids the SSS problem. Second, ENPDE has a diffusion effect on the distance between samples belonging to different classes in the neighborhood, and then the discrimination property is emphasized. The experiments are conducted on three face databases: Yale, CMUPIE and AR. The proposed ENPDE method is compared with the global method, including PCA, LDA, EDA, and the unsupervised and supervised neighborhood preserving embedding methods, including NPE, ENPE, NPDE, and the two-dimension NPDE methods, including 2DDNPE, B2DNPDE. The experiment results show that the performance of ENPDE are better than those of the above methods.