A Research Of Network Structure-Based Community Detection And Evolutionary Characteristic Analysis

With rapid development of economy and society,interactions between people have become increasingly closer.Such close interaction gives rise to complex systems.As an abstraction of real world,complex networks represent individuals as nodes and their relationships as edges.By studying the structural characteristics of these nodes and edges,various inherent properties of the system can be revealed.Therefore,studying the structural characteristics and evolution laws of complex networks is of great significance for comprehending abstract concepts in the real world.As one of the most fundamental tasks in complex network research,community detection aims at uncovering closely connected groups in the network.Such a task helps understand node interaction modes,which ultimately reveal structural properties and functional features of the network.However,the exponential growth of real networks scale poses some new challenges.On one hand,nodes in complex networks often belong to multiple communities simultaneously,making it challenging to detect overlapping community structures in overlapping networks.On the other,the research challenge of dynamically evolving network has gained great attention.Therefore,this thesis delves into overlapping community detection,dynamic community detection,and analysis of dynamic network evolution characteristics.The main work of this thesis can be concluded as follows:(1)For overlapping community detection: the overlapping community detection algorithm based on label propagation only uses the topology to detect communities in the network.It does not need a priori information about the community,nor does it need a predefined fitness function.Therefore,label propagation algorithm has been an active research direction in the past decade.However,the overlapping community detection algorithm based on label propagation does not distinguish nodes,which results in unreliable nodes interfering with the message transmission process in the network,and randomly selecting alternative labels makes the results of each propagation different.In response to these two issues,this thesis first proposes a node influence calculation method in the initialization stage,which comprehensively considers local and location information in the network to calculate the influence of nodes.In the initialization stage,only nodes with high influence are given labels,and the labels of nodes are updated according to the influence sequence,in order to reduce the impact of unreliable nodes on the label propagation process.Subsequently,in the label propagation stage,the influence of the label is measured by its own importance and the importance of its location.During the label propagation process,the label with the highest influence is selected,replacing the random selection process of the label.The experimental results on 44 LFR networks and 5 real networks show that the proposed method effectively improves the modularity and mutual information of the detected overlapping communities.(2)For dynamic community detection: Graph neural networks have strong representation capabilities for low dimensional features of nonlinear data,making them widely used in graph data mining tasks.In the existing research of community detection based on graph neural networks,most of the work obtains the community structure by processing the low order information of the network(such as adjacency matrix).However,the information outside the network topology structure is of great help to the identification of community structure,and only relying on the low order structure information can not obtain high-quality community division.In response to this issue,this thesis proposes an initial partition of the network,which uses the modularity after the initial partition,first-order similarity between nodes,and second-order similarity to strengthen the connections between node pairs,in order to represent the structural information of the network.Subsequently,a three-layer autoencoder is used,and the network structure information matrix obtained in the previous step is used as the input for the stacked autoencoder.At the same time,the temporal smoothing term is used to model the time dependence of the dynamic network,and the embedding vector of the network is obtained by minimizing the structural information matrix and the reconstruction error between adjacent.Finally,clustering algorithms are used to obtain the community structure of the network for the low dimensional vectors.The experimental results on 4SYN networks and 4 real networks validate the effectiveness of the proposed algorithm in this thesis.(3)Aiming at the problem of evolution analysis in dynamic network,this thesis proposes an analysis method based on the evolution characteristics of mobile base station network,and analyzes the evolution characteristics of the network from the "individual perspective" and "collective perspective".First,the mobile base station data in Shulan City is mapped to a dynamic network structure.From the "individual perspective",this thesis proposes a series of "sequence descriptors" to analyze the static and dynamic characteristics of nodes and edges in the mobile base station network.From the "collective perspective",this thesis uses the group evolution discovery method to explore the community evolution events of the mobile base station network and analyze the characteristics of the group evolution in the network.Finally,by comparing and analyzing the evolution characteristics of the mobile base station network in combination with major social events,this thesis discusses the changes in the Shulan citizen mode during the epidemic.