Abstract: The basic features learned by nonnegative matrix factorization (NMF) are usually smooth because the objective function is non-convex for both nonnegative factors simultaneously. To enhance the sparseness of the basic features for NMF, feature re-factorization-based data sparse representation is proposed by looking into the essential similarity between original data and basic features. A set of rough features are obtained from parts of data by making use of several prior regularizations, which are some semantic face images. Applying NMF on these semantic faces, the latent features are further outstanding by holding salient features and eliminating overlapping between rough features. The results from synthetic swimmer dataset and face datasets show that the final latent features are sparser and more powerful for sparse representation than NMF.