The Locality Preserving Projection (LPP) algorithms have been extensively applied for feature extraction of high dimensional face images, but selecting the neighborhood size and defining the affinity weight have a significant impact on the efficiency of LPP algorithms. In this paper, a new sparse manifold learning method was proposed, called Sparse Similarity Preserving (SSP), for dimensionality reduction of face images. It adaptively selected the similarity relation in the global structure of the data and constructed non-negative sparse graph using the sparse coefficients which reserved the global sparsity and non-linear manifold structure of face images, effectively extracting the low dimensional discriminant features. Experiments on two popular face databases (Extended Yale B, and CMU PIE), whose recognition rate reached 87.35% and 90.09%, demonstrated the effectiveness of the presented SSP algorithm.