For the problem of the sparse coding methods not making full use of the label information in data representation, an algorithm, the supervised learning sparse coding, is proposed which can be applied to data representation. Firstly, the proposed algorithm can build the graph via the label information. Thus it directly extracts the discriminate information of the data and then tries to learn the basis which can best fit the discriminate vector. Therefore, it can find a basis set embedding the discriminant information of the samples which are individually for sparse representation. The experiments on the COIL20 and PIE image data sets demonstrate that the proposed algorithm can provide a better representation and classification than the traditional unsupervised matrix factorization algorithms.