Influence Diffusion Inference In Social Networks

Nowadays,with the development of computer science and the popularity of mobile devices,online social networks have replaced the traditional media as the main platform for information dissemination and diffusion.As the explosive growth of users in online social networks platforms,information can be spreaded to all over the world at an unprecedented speed.Each node in the online social network can be the source node,and starts an influence propagation.Therefore,the dissemination of information in social networks often brings tremendous impact on people's daily life,social economy and national security.These factors make the influence diffusion research in the social network get a lot of attention both in academia and in the industry.Focusing on understanding the phenomenon of influence propagation in social networks,this thesis aims to study the key issues of influence propagation inference.We discuss the modeling of influence propagation mechanism,the location of propagation source nodes,the discovery of key nodes in diffusion process and the inference of influence propagation in social networks one by one.The method proposed in this thesis is verified through experiments.Concretely speaking,this article mainly conducts the research from the following aspects.(1)Modeling the influence propagation process in social networks.In this thesis,a cascade model with influence decay is proposed to describe the influence propagation in social networks by considering the randomness,cumulative effect and attenuation of influence over time.Firstly,based on the influence of nodes and neighbors,the mathematical definition of node activation probability is defined.Secondly,a diffusion model that reflects the main characteristics of influence spreading is proposed.The model is an extension of the classical independent cascade model and the linear threshold model.(2)The problem of locating the source node is studied.Based on the Bayes theorem,this thesis proposes a Bayes framework for solving posterior probability of propagation source direction from observation nodes.Based on the idea of random walk theory,a method of backtracking propagation source node based on posterior probability is proposed.This research is based on the idea of random walk backtracking,which is free from the complete network structure and node information,and makes the backtracking process accord with the randomness of influence propagation.The framework based on Bayes theory improves the accuracy of source node localization.(3)We studied the problem of finding key nodes for influence diffusion.This thesis discusses two different source node discovery methods based on the maximization of impact propagation and propagation information.Firstly,we prove the Submodularity of the propagation maximization objective function based on the cascaded model with influence decay and give a greedy algorithm to locate the key nodes.Secondly,we propose a key node estimation method based on propagation information,and prove the unbiasedness of the estimation.Based on the propagation maximization method,the key node set which contributes greatly to the propagation of influence can be obtained by considering the interaction among nodes.However,the method based on information dissemination does not depend on the propagation model and network structure,and solves the more universal propagation key nodes in the network.(4)We have studied the method of influence propagation inference in social networks.Based on the results of source node solution and key node discovery,this thesis improves the cascade model with influence decay,and gives a complete method to infer the influence propagation according to the source node and the improved propagation model.Based on the observation results,we continually correct the inference result.This research solves three key problems of propagation inference,which are propagation mechanism modeling,propagation source nodes location and propagation key nodes discovery.Experiments on a real influence propagation data on Twitter show that the proposed method effectively solves the key problems in influence propagation inference,and the inference result are very close to the real propagation result.