Summary: In order to make use of the local structure information and the label information of limited labeled data, a constrained nonnegative matrix factorization based on local learning (CNMFLL) is proposed for data representation. To take consideration of the local structure information in the data, a predictor is constructed by the neighborhood of each point and its label information is estimated. In addition, the label information of the labeled data is as hard constraints so that the samples sharing the same label in high dimensional spaces have the same coordinate in new representation spaces. Therefore, this algorithm not only makes use of the geometry structure information and discriminate structure information, but also considers the label information of labeled data. Thus, CNMFLL has more discriminate power than traditional NMF. The experimental results on 20 Newsgroups text databases and ORL face databases show that the proposed algorithm can improve the accuracy and normalize mutual information.