A Matrix Factorization Method With Graph Regular Term For Clustering In Heterogeneous Information Networks

In the context of big data,information technology is developing at a rapid pace,and there is an increasing amount of data available for research.Many complex systems in the real world can be abstracted into the form of networks by these rich data,and since these ubiquitous networks are closely related to our lives,the research work on complex networks is very urgent.Mining useful and stable communities in the network has important research value for mining,recommending and predicting network information.At present,most community discovery methods are carried out on homogeneous networks,but whether they are objects in the real network or the relationship between them can be diverse,and the heterogeneous network can obviously perform the real world more accurately.Abstraction,but how to effectively mine the community structure in heterogeneous networks also poses new challenges for community discovery.Through research,the heterogeneity in heterogeneous networks leads to the following problems in the current community discovery methods:(1)the interaction noise in heterogeneous networks is many,resulting in reduced performance of the algorithm;(2)between different types of nodes in the same dimension The interaction noise is too large,resulting in the inability to distinguish the actual community.Aiming at the above problems,this paper takes a large number of existing star networks in heterogeneous networks as the research object,and proposes a community discovery algorithm based on non-negative matrix decomposition of graph regular terms.The key technical work of this paper is as follows:First,combined with the traditional non-negative matrix factorization algorithm,a joint optimization framework that can fuse heterogeneous information of each sub-network is designed.The framework learns the consensus matrix reflecting the common structure of different sub-networks,and uses each row vector in the consensus matrix as the central type node's membership distribution in each community,and continuously optimizes the coefficient matrix by using iterative matrix.And consensus matrix.Here.in order to achieve the purpose of differently treating different sub-networks,the model adopts an automatic weight learning strategy that assigns different weights to different importance sub-networks,balancing the impact of different sub-networks.Secondly,based on the comprehensive comparison and analysis of the current main heterogeneous network community discovery algorithm,the internal connection relationship between the central type subspace and the attribute type subspace is introduced as a constraint item into the regularization joint optimization framework through the regularization of the graph,combined with the multiple sub-s The manifold constraint of space effectively utilizes the topology of each sub-network,and adds a priori information to the graph's regular item weight matrix,which makes the framework shift to semi-supervised type,and the prior information is utilized.Thirdly,based on the actual data set.the joint optimization framework and community discovery algorithm proposed in this paper are experimentally tested and analyzed.The framework construction,algorithm implementation and result visualization are completed by Matlab software platform.Through experiments,compared with the previous algorithm,the superiority of the proposed algorithm performance is also verified.