Modeling Information Diffusion In Social Networks

Online social networks have changed ways of communications between people and propagation of information among people.Understanding information diffusion in social networks has great scientific and marketing values.However,the fundamental mechanism behind information diffusion is complex;users' roles and behavior patterns in diffusion process are hard to be quantized;the interplay between individual behaviors at micro level and diffusion phenomena at macro level remains unclear.To address these problems,this thesis utilizes technologies of both data mining and machine learning,along with cross-disciplinary knowledge,to construct behavioral analysis and design diffusion models from different angles.At the micro level,to study individual user's retweet behavior,this thesis proposes social role-aware information diffusion model(RAIN),which integrates social role min-ing and diffusion process modeling into the same framework.Comparing with traditional models,RAIN reduces computational complexity by assuming that users with the same social roles have similar diffusion parameters.Meanwhile,RAIN handles cold-start is-sue by leveraging structural information of each user.Moreover,this thesis studies how opinion leaders,who take central positions in a given network,and structural hole span-ners,who bridge otherwise disconnected communities,influence information diffusion differently.This study is further extended into an emotion diffusion scenario.At the macro level,to study how diffusion scales evolve over time,this thesis proposes a dynamic model,which describes the dynamic process of both user status and information popularity simultaneously.The proposed model balances the conflict between model complexity and granularity,and is able to derive a power-law distributed diffusion scale by assuming more popular posts will receive more retweets.Finally,this thesis proposes a uniform framework to explain the interplay between individual behaviors and heavy-tailed phenomena.In theoretical,with different param-eters,the proposed framework is capable to obtain diffusion scales following different heavy-tailed distributions,such as power-law distribution and log normal distribution.