Research On Social Influence Measure Based On Constrained Tensor Factorization

With the rapid development of Web 2.0 technology and the prevalence of online social networks(such as Twitter,Facebook,Twitter,etc.),there are a plenty of user generated content existing in social media,including:retweet,comments,etc.It is a superexcellent opportunity to analyze and measure the influence of users for us.Social influence occurs when one's emotions,opinions,or behaviors are affected by others.As a kind of important factors influencing the network structure and information dissemination,it has attracted the attention of many researchers.By studying the existing works,we find that the user social influence is determined by a variety of factors,but the measurement methods based on graph model or matrix have certain limitations in the fusion of many factors.Leveraging the advantages of constrained tensor factorization in fusing multiple factors,we propose a novel social influence measurement model based on constrained tensor factorization.The details of the research are proposed as followed:First,existing social influence analysis methods are limited at fusing multiple factors at the same time.Taking the advantages of tensor in expressing multidimensional data,this paper proposes a kind of social influence measurement model based on multiple factors constraint,which uses tensor to storage interactive relationships among users and the relationships contain opinion information.User popularity and topic relevance are in the form of constraint term to control tensor factorization process.And then we figure out the tensor factorization by alternative least square.The calculation of user influence and the influence polarity distribution is by means of tensor completion.The experimental results show that,compared with OOLAM,TwitterRank etc,the model improves 9%in average p@k at least.Second,The previous model we proposed is failed to take into account the user topic similarity.We propose another social influence measurement model based on Laplacian constraint.The user topic similarity is fused into tensor factorization by Laplacian Matrix to affect the tensor factorization process.And we give the convergence proof of the constrained tensor factorization method.The experimental results show that,compared with the previous method we proposed,this model increases by an average of 1%in average p@k.Third,in practice,with the increasing of data scale,the Laplacian constrained tensor factorization method has to estimate massive parameters,which slows down the tensor factorization process and the demand of computer's memory will increase sharply.This paper based on MapReduce programming framework parallelizes Laplacian constrained tensor factorization method to alleviate the issues.The experimental results show that the tensor factorization algorithm based on MapReduce has good scalability and can make use of multiple nodes to satisfy the memory requirements.