A Subpattern-based Weighted Neighborhood Maximum Margin Criterion (SP-WNMMC) algorithm is proposed for face recognition. In order to enhance the robustness to facial pose, expression and illumination variations, SP-WNMMC method firstly operates on sub-patterns partitioned from an original whole face image. The contribution of each sub-pattern can be adaptively computed through the class information of neighborhood. Secondly, WNMMC is adopted in each sub-pattern to extract local features. WNMMC can preserve the local geometric structure of database. The objective function of WNMMC leads to the enhancement of classification capacity by using the linear reconstruction coefficients. Thirdly, for a new face image to be recognized, all the likelihoods in all the subpatterns are fused together for the final recognition result. Experiments show that our method can effectively extract the local feature while preserving the non-linear structures in sub-pattern sets. It can consistently outperform other recognition methods based on Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA) and WNMMC.