Research Of Social Network Circle Identification

With the continuous development of technology and network,real-life interactions are gradually abstracted to social networks for research.Social circles are the basic structural features on social networks.Detecting the community structure of social networks plays an important role in uncovering complex social network features,and also provides guidance for understanding real-life social behavior such as precise marketing,personalized recommendation and knowledge dissemination of social circles.Community detection is one of the most important means to identify social network circles.Realworld social networks usually have multi-layer properties and dynamic properties,and can be portrayed through multi-layer networks and dynamic networks.Therefore,in order to study and analyze the circle division of social networks,this thesis will focus on the community detection of multi-layer networks and dynamic networks:(1)Concerning the issue that the complex inter-layer relationships of multilayer networks make network consensus communities difficult to obtain,this thesis proposes a multilayer network community detection algorithm based on prior information and network reduction,using network prior information based on non-negative matrix method and node similarity to reduce the influence of network redundancy and noisy information in the clustering process on consensus communities,which improves the accuracy of multi-layer network community detection.Firstly,the algorithm uses a non-negative matrix decomposition method to generate a consensus prior information layer of multilayer network,and uses the network reduction strategy based on nodedegree to compress the network layers to reduce the time complexity.After that,the prior information of network is combined with the decomposition-based multi-objective evolutionary algorithm,and new chromosomes are generated to guide the chromosome population evolution by the uniform crossover operator as well as the neighbor-based mutation operator.Moreover,for the mis-divided nodes in the reduced network,the network repairing strategy based on the similarity prior information is proposed to correct the community detection after each iteration,making full use of the information of network node features to reduce the interference of network noise and improve the robustness of multilayer network community detection.Comparative experiments on real and artificial networks demonstrate the accuracy and robustness of the algorithm.(2)Concerning the issue that the complex temporal characteristics of dynamic networks make the global information of network difficult to characterize,this thesis proposes a dynamic network community detection algorithm based on long-short term memory networks and contrast learning,which combines the mutual information maximization strategy and the network smoothing strategy to cyclically update the low-dimensional representation of nodes,and reduce the impact of information differences of neighboring snapshot networks on the global representation ability of the networks,and achieves high-quality detection of dynamic network community structures under unsupervised learning.Firstly,the algorithm uses a feature aggregation strategy based on network nodes relevancy to calculate the network feature matrix,and based on this,a mutual information maximization strategy is used to inscribe the global low-dimensional representation matrix of the nodes on a single snapshot network,and the cross-entropy loss is used to maximize the mutual information between the local and global feature of dynamic network.Then,in order to reduce the computational overhead in the optimization process,the weight parameters of the graph convolutional neural network in each snapshot are updated using long-short term memory networks.Finally,this thesis designs a network smoothing strategy based on the contrast learning to minimize the feature difference between neighboring dynamic network nodes.The superiority of the proposed algorithm is verified by experiments on real and artificial networks.