Modeling And Measuring Social Influence

Social influence occurs when one's opinions or behaviors are affected by others.It forms a prevalent,complex and subtle force that governs the dynamics of social networks.With the rapid proliferation of online social networks such as Twitter,Facebook,Yelp,Amazon and so on,modeling the influence diffusion mechanism and quantitatively mea-suring social influence between people becomes more and more critical for applications such as friend recommendation,expert finding,behavior prediction,and can also bene-fit the development of virtual marketing and supervision by public opinions.Recently,social influence has attracted tremendous attention from different communities and has been extensively studied before.However,the underlying mechanism is still unclear and a thorough investigation is thus needed.Studying the structural characteristics of social influence can help understanding the mechanism of influence diffusion more deeply.When there are multiple pairwise influence,existing research usually simply combines them together,which ignores the structural characteristics of the network formed by those who exert pairwise influence.To study the structural characteristics,we first quantify the metrics of structural diversity of multiple pairwise influence,and furthermore,propose a formulation of structural influence;second,we formalize conformity influence in terms of a utility function based on the conformity theory,and further incorporate the utility function into an application-oriented probabilistic model,to describe the correlation between conformity and structure characteristics.The dynamics of a social network increases the difficulty of modeling and measuring social influence.While existing research usually considers a static social network and ignores the effect from the network dynamics.In this paper,given a dynamic social network,we propose a time-delayed influence model to describe the diffusion process between the formation of relationships in a network.The model also incorporates the triad structures where the formation of neighboring relationships influence each other.Based on the model,we propose an application of friend recommendation maximization,which not only considers one-step acceptance rate,but also takes care of the effect of acceptance diffusion beyond one-step recommendation.When social network becomes extremely large,the efficiency of the existing mea-suring methods is limited.In this paper,we first propose a sampling algorithm to quickly measure pairwise influence between people.A theoretical proof about the lower bound of the sampling size is provided.An extensive empirical study on a large network of 1 billion edges shows that the proposed algorithm can be 300x faster than the state-of-the-art meth-ods;second,we propose three sampling algorithms to quickly discover frequent influence structures by scanning users' activity logs.A theoretical guarantee about the unbiased estimation is provided.An experiment on a microblogging network of 2 million nodes,0.3 billion edges and 20 million retweet activities shows that the proposed algorithms can achieve a 10× speedup compared to the exact method,with an average error rate of only 1.0%.