Traditional locality preserving projection (LPP) algorithm directly uses the spatial information of original data, which leads to inaccurately select neighborhood, and ignores data categories information of projection of LPP algorithm. To solve these problems, a face recognition method is proposed based on adaptive neighborhood LPP. In the feature extraction, the objective function is constructed based on the variable similarity, neighborhood information, and data categories information, so that the same class samples are close and different class samples are far away from each other in projected subspace. Adjacency matrix and projection matrix are adaptably optimized by minimizing objective function. Optimized projection matrix is used to reduce the dimension of high-dimensional face data, and low-dimensional data is used to classify and recognize face samples. The experimental results on Yale B, PIE, MSRA and CAS-PEAL databases validate the effectiveness of the proposed algorithm.