Research On Feature Modeling For Community Detection In Dynamic Complex Networks

Data modeling is an important research problem in the field of machine learning and artificial intelligence.Clustering analysis can mine the inherent patterns and laws of data effectively.As a new research hotspot,complex network(graph data)plays an important role in understanding social complex system,which can be effectively used to recommend system,terrorist rganization mining and other issues.In recent years,the dynamic evolution of network structure has brought new challenges to community detection modeling.How to use efficient learning model to mine the internal laws of dynamic complex network and realize community detection and evolution analysis has become the focus of dynamic complex network research.The project plans to study the feature modeling and analysis for community detection in dynamic complex network from the perspectives of node importance,community structure,feature representation,high-order similarity,etc.,and to realize the joint modeling of node feature,network feature and community feature in dynamic network by using the unified non-negative matrix factorization architecture.The specific work includes:In view of the challenge of the interaction between the importance of nodes and community structure,a joint model of the change of node importance and community structure is proposed.Specifically,based on the dynamic non-negative matrix factorization model,an evolution matrix is introduced to model the evolution process of node importance.At the same time,another evolution matrix is introduced to model the evolution process of community structure,and a dynamic network multi-structural feature joint model is proposed.Furthermore,based on the gradient descent method,an optimization method of the model is proposed.Experimental results show that the model has good performance in time-varying community detection and key node identification tasks.In view of the dynamic variety of local characteristics of nodes,a unified framework of learning regular community structure modeling using network representation is proposed.Specifically,based on the factorization of symmetric non-negative matrix to integrate the current topological information of the network,and based on the evolution clustering to introduce the evolution matrix to integrate the historical community structure information of the network.At the same time,the network representation learning method is used to describe the dynamic local features of nodes,and the dynamic network feature model is integrated into a regular form.Then a unified framework of community structure modeling based on network representation learning is formed,and its corresponding optimization algorithm is given.The experimental results show that the framework is robust and effective.Aiming at the sparsity of network and the effectiveness of high-order representation,a community evolution model is proposed,which integrates the high-order features of dynamic network and the regular first-order similarity.Specifically,in view of the sparsity of real networks,the high-order representation strategy of dynamic networks is introduced,taking into account the manifold structure of network data,combining with the advantages of multi-structural feature modeling of dynamic networks,the first-order similarity regularity is introduced to propose a community evolution model integrating the high-order features of dynamic networks.Experimental results show that the model has good performance in time-varying community detection and key node recognition.Aiming at the important problems in dynamic complex network mining,this paper proposes a set of theoretical methods for community detection,network evolution and node importance analysis,which improves the application scenarios and theoretical methods of dynamic complex network.