Abstract: Large-scale sparse graph data appear in reality heavily,for example,collaborative graph,Laplacian matrix,and so on.Non-negative matrix factorization (NMF) has become a useful tool in data mining,information retrieval and signal processing.How to achieve the data clustering in large-scale data is an important issue.This paper used the two-stage method to realize data clustering.First of all,the Nystrm approximate sampling method is used.The initial profile of data is obtained from large data,and the similar matrix of user-user or movie-movie is obtained.The purpose of doing that is to reduce the original high dimensional space to a low dimensional subspace.Then convex non-negative matrix decomposition of low dimensional similarity matrix is used to get the center of the cluster and indicator.The center of the cluster represents the features of movies or users,and the indicator represents the weight of the features of mo-vies or users.The advantage of two-stage preference clustering method is that the approximation of initial data contour and convex non-negative matrix factorization have better robustness and anti-noise.On the other hand,the data of the subspace are derived from the real matrix,which makes the results of clustering preference have good interpretability.This paper utilized Nystrm method to solve the problem that large-scale data cannot be stored in the memory,saving memory and improving operation efficiency.Lastly,the test on movie data sets which containing 100000 ratings shows the effectiveness of the clustering algorithm.

Key words: Nystrm method,Convex non-negative matrix factorization,Preference clustering,Clustering center,Clustering indicator