The Target User Oriented Collaboration Strategy Mining In Social Networks

With the advantages of real-time interaction,along with openness and convenience of speech,social network media have gradually become the main platform for the public to obtain news information,discuss current affairs and organize collective activities.For specific events,groups of users,driven by politics or interests,try to manipulate the direction of public opinion in a cooperative way.Therefore,how to understand the organized manipulation behavior of user groups in a specific political context and how to explore the cooperation strategy behind it pose a new challenge to content governance in cyberspace security.Existing industry research on group collaboration strategies mainly focuses on the detection of collusive group’s political roles,which uses the individual level of tweet semantics.However,there are limitations in the analysis of collaboration strategies only from the perspective of collusive groups.It is necessary to supplement the analysis of collaboration strategies in the view of target users and explore the correlation between the target users’ political inclinations and collaboration strategies,so as to verify the effectiveness of collusive group collaboration strategies.To sum up,this thesis studies the following three aspects of the problem of collaboration strategy mining.(1)We propose a multiple collaboration strategies and topic augmentation model for political role detection,which integrates various collaboration strategies and global common topics.The algorithm adopts a heterogeneous graph attention network for multi-level collaborative user information fusion,namely co-hashtag strategy,co-mention strategy and mention strategy,and increases the dimension of topic information through biased random walks between group-topic bipartite graphs.Finally,it can effectively model the cooperative relationship between collusive users,and realize the mining and fusion of the implicit cooperative behavior characteristics of collusive group in political role detection.The superiority of the algorithm is demonstrated by experiments on the Russian Trolls dataset and comparison with existing research methods.(2)This research proposes a target users’ political inclination detection algorithm based on local influence subgraphs.The algorithm considers the target user’s local influence subgraph and designs a variety of subgraph extraction methods.On this basis,the Graph Isomorphism Network is used to encode the network structure features and user semantic features of the local influence subgraph,and performs graph-level classification to verify the effectiveness of the collective group collaboration strategy.In addition,an ensemble tree model XGBoost is used to analyze the importance of statistical properties in the highly effective subgraphs.The rationality of the algorithm is proved by analyzing the structure strategy of the complex network level of collusive users from the perspective of the influence of target users.(3)Aiming at the portability and repeatability of collusive group cooperation strategy research in system application,this study integrates the MCSTA algorithm and LIS algorithm according to the characteristics of the actual system to design a target user-oriented group collaboration strategy mining system.The main functions include the detection of collusive users’ political roles and the analysis of complex network characteristics of local influence subgraphs of target users’ political inclinations.Finally,by testing the prototype system,the obtained results show that it is easy to understand the cooperation strategy of the group,which proves the usability of the system.