We propose a novel manifold learning approach, called Neighborhood Discriminant Projection (NDP), for robust face recognition. The purpose of NDP is to preserve the within-class neighboring geometry of the image space, while keeping away the projected vectors of the samples of different classes. For representing the intrinsic within-class neighboring geometry and the similarity of the samples of different classes, the within-class affinity weight and the between-class affinity weight are used to model the withinclass submanifold and the between-class submanifold of the samples, respectively. Several experiments on face recognition are conducted to demonstrate the effectiveness and robustness of our proposed method.