Abstract: In order to capture the local geometric structure of multi-faceted data and improve the clustering performance, a nonlinear matrix factorization data clustering method based on manifold learning is proposed. A Pnearest neighbor graph was constructed for each relationship to capture two different types of closely related objects, so as to accurately learn the internal relations and multiple manifolds generated by the internal relations of data. And we stably kept the learned manifold when mapping to a new low dimensional data space with nonlinear matrix factorization. The clustering results of multiple data sets show that the method can fully mine the partial representation of various related types, and has certain advantages in accuracy and efficiency.