Modeling And A/B Testing Based On Social Network Data

In recent years,the rapid development of social networks has brought sweeping changes to human daily life,and the huge amount of information generated has become the focus of research in many disciplines.In the context of statistics,scholars mainly carry out research work based on the structural properties and modeling of social networks,among which,social network modeling is the verification and deepening of the understanding of its formation mechanism and evolution law based on the analysis of network structural properties.By simulating the interaction behaviors between people in real social networks,we construct evolutionary models with corresponding attributes and study the influence of specific social behaviors on the network structure.At present,the relevant research content mainly focuses on the structural characteristics of the network,information dissemination,and user behavior analysis.Based on this,this thesis,from the perspective of social network analysis,firstly,explores the behavioral changes generated by social influence on user interaction,and thus builds a relevant model to characterize this process;secondly,further considers the estimation effects of different parameter estimation methods on network models;finally,in the context of commercial applications,discusses how to differentially deliver game versions according to the different characteristics of network users,to achieve the purpose of accurate Finally,in the context of commercial application,it is discussed how to deliver game versions differently according to different characteristics of online users,to achieve the purpose of accurate marketing.The specific work carried out in this thesis is as follows:(1)To portray the interdependence of behaviors or views among individuals in the network structure,a network autocorrelation model based on Bayesian posterior inference is developed.Specifically,the potential locations of nodes are estimated using a potential space model to construct social influence to solve the problem that classical autocorrelation models cannot characterize the weights between two users.The proposed model is also applied to longitudinal data on adolescent friendship networks and smoking,and its effectiveness is explored in terms of precision and recall,and the results show that the model can better characterize the social influence generated by group interaction.(2)Variational EM algorithm inference under potential space-based models is investigated.Specifically,the classical MCMC algorithm often faces problems such as large computational costs and slow convergence speed in estimating model parameters under large-scale networks.To this end,this thesis discusses the theoretical framework of the variational EM algorithm and applies the Bayesian approach and the variational EM algorithm to the social network data of Instagram users to compare the estimation efficiency of the two,and the results show that the difference in estimation effectiveness between the two types of algorithms is small,but the latter can effectively save time cost and has a greater advantage under large-scale networks.(3)A/B tests on the selection problem of tweeting schemes in social network data are investigated.Specifically,two types of statistical tests are used to examine the A/B scheme selection problem: one is based on the Bayesian theory of posterior expectation loss to determine the advantages and disadvantages of different schemes;the other is to calculate the sequential probability ratio to select between schemes.The data of the level set in the mobile puzzle game "Cookie Cats" are also used to illustrate the applicability of the above two methods.The results show that the two methods are consistent,and the sequential A/B test has excellent results in small samples,saving time and cost,and is more suitable for companies with less information about users to make business decisions.In summary,this thesis explores the intrinsic value of network data from two aspects of network structure modeling and scheme pushing,respectively.In terms of technical route,it enriches and promotes the connotation of Bayesian posterior inference,variational EM algorithm,and A/B test of the network model to a certain extent;in terms of research significance,it draws on some new ideas of current Bayesian methods with certain application value,thus realizing the purpose of statistics for socio-economic services.