Efficient Clustering Based On Tensor Nonnegative Decomposition

CPSS contains abundant heterogeneous networks.Clustering analysis of heterogeneous networks is helpful to find the shallow data features in the network,and the unified representation of heterogeneous networks by tensor is helpful to integrate multisource heterogeneous data.But the current clustering analysis of heterogeneous networks faces the following problems: the existing algorithms can not cluster different types of objects,and can not make full use of the sparse characteristics of heterogeneous networks.To effectively utilize the rich semantic information in heterogeneous networks and cluster various types of objects,this thesis first introduces the tensor representation of heterogeneous networks,which can represent different heterogeneous networks as tensors.In order to cope with the large-scale network representation and reduce storage,it also introduces the sparse storage of tensors.Then,from the perspective of optimization,the clustering model is established,and the equivalent relationship between the constrained tensor nonnegative decomposition and the clustering model is established.Based on different updating strategies,two clustering algorithms based on tensor nonnegative decomposition are proposed.Both algorithms can cluster all kinds of objects.The performance of the two algorithms is compared on simulated data and real data Experiments show that the clustering algorithm based on LS distance has better clustering performance than that based on KL distance.To reduce iterations caused by random initialization of two algorithms,a similarity measure method based on tensor is proposed by combining with the theory of Metapath,and then an initialization algorithm based on user guidance is given.Experiments show that the initialization algorithm can provide efficient initial solution,and reduce the number of iterations.According to the sparse characteristic of clustering task,the idea of transition matrix is extended to the clustering model based on LS distance,and a clustering algorithm based on sparse control is proposed.Experiments on real data sets show that the algorithm can effectively control the sparse degree of factor matrix through parameter .