Community Evolution Prediction Based On Deep Learning Model

Complex networks exist widely in people's production and life.An important research direction in this field is the division and evolution analysis of network community structure,which is helpful to recognize the structure and function of complex network,and has important significance in application fields such as advertising,information dissemination management,personalized recommendation,etc.But the accuracy of community discovery is not high because of the problems of high matrix dimension and lack of local information when using classical clustering algorithm to community discovery.In the current community evolution prediction research,there are two main problems.The first is that the prediction feature set constructed by the model usually only contains the features reflecting the static attributes of the community,and did not consider the dynamic attributes in the process of community evolution,which makes the timing information not fully used,and the prediction results were not accurate enough.The second is that the model described the community evolution by extracting the evolution chain,which can not directly show the evolution of community division and integration.In view of the above problems,this thesis proposes a community discovery method based on deep learning model.The community evolution predicted method improves the accuracy of community discovery and evolution prediction by using deep learning model in two aspects: community discovery and community evolution prediction.(1)In order to reflect the local information of each node in the network,the adjacent matrix of the network was transform into the similarity matrix by using hop connection.A deep sparse self-encoder network model with attention mechanism was established to extract low dimensional features of similarity matrix in this thesis.It reduced the complexity of clustering tasks,and improved the expression ability of the total community structure of network topology.Finally,the K-means clustering algorithm was used to divide the community,which improved the accuracy of community discovery.Through the experimental verification,the results showed that the community discovery method proposed in this paper has better community division ability and can find good community structure.(2)Based on long-term memory network,a community evolution predicted method was proposed.Firstly,the multi-element community features were extracted from the core node attributes,community structure,time sequence and behavior,and the evolution feature set was constructed.Secondly,this thesis detected the evolution events in different time windows and constructed the evolution tree,which can show the evolution events in the whole life cycle of each community intuitively.Thirdly,for the features extracted from the model,long-term and short-term memory networks were used for training to obtain the long-term dependence in the process of community evolution.The experimental results show that the proposed method has high accuracy.There are 20 pictures,15 tables and 82 references in this thesis.