Research Of User Influence In Social Network Based On Bayesian Statistical Inference

In recent years,social network presents its explosive development,and it has deeply influenced our real society. There are millions of users in social network,along with massive data.A research on quantification of user influence in social networks will help us explore the laws of human behavior,and the results can be applied to a lot of areas,such as business promotion,the supervision of public opinion,personalized search,link prediction and so on.The PageRank algorithm can be applied to measure the overall influence of users in directed social networks.But to only take the topology structure into account and assume that the partial influence between users is equal make this algorithm not so accurate.By abstracting the behavior of retweet into multinomial-distribution,the partial influence can be measured as retweet probability,the unknown parameters. Due to the shortcomings of classical statistical inference,we use Bayesian statistical inference to solve this problem.Firstly,we define the concept of Similarity by using the topology structure of social network,and give the calculation formula.Similarity of users is regarded as prior information.Combined with the sample information of retweets,we obtain the calculation formula of partial influence by using Dirichlet distribution which is a conjugate distribution.Finally,we improve the PageRank algorithm by using the value of partial influence which is calculated from Bayesian statistical inference above as assignment weight of influence,and propose an algorithm called BPRank to measure the overall influence of users.Personalized PageRank algorithm can be used to measure the positive influence which is a kind of relative influence of a specified user.We thus define a concept of Inverse Influence,which is another kind of relative influence.According to the difference between the two kinds of relative influences,we give the basic definition of inverse influence based on local random walk,and then propose an algorithm called BWPRank to calculate it.It consists of three steps.Firstly,to generate a tree data structure represents path.Secondly,to find the appropriate paths we need.And lastly calculate the transition probability of each path by using the Bayesian value of partial influence above,then thesummation is the value of Inverse influence.BWPRank algorithm can be applied to link prediction.At the last of this paper,by using the real data of Sina weibo,we design a series of comparative experiments for the above two algorithms respectively.The result shows,the quantification of partial influence between users is thorough and reasonable,and both of BPRank algorithm and BWPRank algorithm are valid and accurate.