Considering that global dimensionality reduction methods lack useful discriminant information, and local dimensionality reduction methods have defects in measuring neighborhood relationships, a novel dimensionality reduction method based on margin, named Margin Discriminant Projection (MDP), was proposed. Depending on the neighbor structure of mean vector of classes, the boundary vector of the class edge was defined by the heterogeneous neighbor relation of the center mean of the class. On this basis, the between-class scatter matrix was redefined, and the within-class scatter matrix was constructed by the global method. The class margin criterion was established based on discriminant analysis, and discriminant information of samples in projection space was enhanceed by maximizing class margin. The expression recognition on JAFFE and Extended Cohn-Kanade data sets presented the comparison of MDP with PCA (Principal Component Analysis), MMC (Maximum Margin Criterion) and MFA (Marginal Fisher Analysis), and the experiment results show that the proposed method can extract more distinguishable low-dimensional features with relatively higher efficiency, and MDP has better classification accuracy than the other methods.