As an effective data representation method, non-negative matrix factorization has been widely used in pattern recognition and machine learning. To obtain a compact and effective data representation in data dimension reduction, unsupervised non-negative matrix factorization usually needs to discover the latent geometry structure information of the data. A similarity graph constructed by reasonably modeling similarity relationships between data samples is typically used to capture spatial distribution structure information for data samples. Subspace learning can effectively explore the subspace structure information inside the data, and the obtained self-expressive coefficient matrix can be used to construct a similarity graph. In this paper, a nonnegative subspace clustering algorithm is proposed to explore the subspace structure information of data which is used to guide the non-negative matrix factorization, so as to obtain the effective non-negative low-dimensional representation of the original data. At the same time, an effective iterative strategy is developed to solve the problem of non-negative subspace clustering. The results of clustering experiments on two image datasets demonstrate that utilizing the subspace structure information of data can effectively improve the performance of non-negative matrix factorization.