Research On The Identification Method Of Opinion Leaders In WeChat Groups

As an important vehicle for people's daily pastime,information distribution,dissemination and execution,the stable and complex information flow network formed within We Chat groups contains symbols of the identity and status of group members,the infiltration of social capital and the control of the group's discourse.Opinion leaders in groups are important for the formation of relationships between group members and the publication and dissemination of information.In this thesis,by understanding the association between opinion leaders and the strength of relationships theory in social network theory,analysis the importance of relationship strength for We Chat groups and the identification of opinion leaders in We Chat groups,and investigate the characteristics of We Chat groups,arguing that We Chat group information flow networks are weak relationships networks based on strong relationships.Based on existing methods of social network centrality analysis,we propose an algorithm for identifying opinion leaders in We Chat groups and optimize the performance of the algorithm in terms of applicability and convergence.The research in this thesis is divided into the following two parts:1.The RPR algorithm for identifying opinion leaders of We Chat groups is proposed.The identification of existing We Chat group opinion leaders is mostly based on chat contents,but because of the alternation and drift between topics in We Chat group chats,it does not describe the interaction behavior among members well,thus leading to the inability to accurately establish interaction relationships among group members.To address this problem,this thesis proposes an RPR algorithm based on association strength,which avoids this situation while identifying the opinion leaders in the group.First,the We Chat group chat data is grouped in days,and each group represents a topic or unique attribute feature,then each group member will have its own unique attribute vector.The homogeneity between members is calculated using Jaccard similarity as the strength of the strong relationship between group members.Secondly,we use the strong relationship between group members to build a We Chat group information flow network,use the shortest path-based social network centrality analysis method as the measure of weak relationship between group members,and use the Page Rank algorithm as the basis and combine the strong and weak relationship to build an RPR algorithm for identifying We Chat group opinion leaders.Finally,the centrality analysis of nodes in social networks is carried out separately using the existing social network centrality analysis method and the RPR algorithm,and the obtained experimental results are compared with the SIR propagation model for group member influence,which strongly proves the effectiveness of the RPR algorithm.2.Algorithm performance improvement.The improvement mainly focuses on two aspects:applicability and convergence.First,in networks with degree power-law distribution or high modularity,the Page Rank algorithm will distribute most of the weights to a few nodes,leading to an increase in the localization of the Page Rank vector,making it difficult to distinguish between the remaining nodes.In order to solve this problem,this thesis proposes a HPR algorithm for weakening the phenomenon of network localization and gives the expression of transition probability matrix of this algorithm in directed and undirected networks,which well weakens the degree of localization of the original algorithm network.Secondly,the power iteration method is a common iteration method to solve the Page Rank problem,which is computationally cheap and easy to implement,but when the damping factor?of the Page Rank algorithm tends to 1,the second eigenvalue of its transition probability matrix are closer to the primary eigenvalue,and its convergence speed becomes slower and its performance is poor.For this reason,in this thesis,we propose a power iteration method based on the normalization of Lnorm and theoretically prove the convergence of this normalization method on the power iteration method.Finally,we experimentally demonstrate the effectiveness of the HPR algorithm based on Lnorm normalization,which both reduce the localization of the feature vector of the original algorithm to a certain extent and largely improves the convergence speed of the algorithm without changing the original centrality ranking of the nodes in the network.