Feature Mining And Popularity Prediction For Information Diffusion In Online Social Networks

With the rapid development of Internet technology and the rapid growth of the scale of mobile users,the role of users in information dissemination has gradually changed from passive receivers to active producers and disseminators,bringing new challenges to the network public opinion governance in the Internet era.As an important part of social network analysis,information dissemination prediction is of great significance and application value.However,the current researches on information dissemination feature mining is limited to a single public opinion event,and lacks quantitative analysis of information dissemination environment and dissemination continuity;researches on popularity prediction often has poor prediction accuracy,and there are few studies on the prediction of information reinforcement effect.Focusing on the information dissemination process in large-scale online social networks,this thesis conducts compressive researches on the dissemination popularity prediction from the perspectives of information dissemination scale,information coverage and reinforcement effects.The main contents include:(1)A systematic and comprehensive method for mining spatiotemporal evolution features of information dissemination is proposed.To address the shortcomings of the previous information dissemination feature construction methods in potential user scale mining,a systematic and comprehensive information dissemination evolution feature mining method is constructed based on an analysis on a large number of real microblog dissemination link data.This method takes the information dissemination spatiotemporal evolution features into account,and provides a comprehensive measurement of information attributes and dissemination process.Multi-dimensional features are extracted from the perspectives of information diffusion space structure,information diffusion time evolution,and information text attributes,including the network structure features based on the following network,retweet network,and boundary user network,the time series features based on information dissemination links,and the topic features extracted from informative texts.It shows that the potential user scale is mined by the structured forwarding coefficient indicator,and the continuity of information dissemination is effectively quantified by the indicators of information activity and forwarding activation characteristic.(2)A feature enhancement-based information dissemination scale prediction algorithm is constructed.Transforming the information dissemination scale prediction problem on the microblog platform into a machine learning problem,an accurate information dissemination scale prediction model is constructed,the feature importance and error factors in the prediction problem are also analyzed.First,two prediction tasks are set up,i.e.to predict the information dissemination scale based on the forwarding process within hours after the message(PS1H)and the previous forwarding situation(PSn R).Next,a feature enhancement-based information dissemination scale prediction algorithm(FEPS-M)is proposed to predict the spread range,forwarding breadth and forwarding depth of microblogs,by taking the spatiotemporal evolution feature set of information dissemination as the input.Finally,experiments are designed to explore feature importance and the factors that influence the predictability of the scale of information dissemination.Results shows that for the PS1 H tasks,the FEPS-M algorithm can achieve accurate dissemination scale prediction,and the mean square error is only 92.09,the goodness of fit was 0.95.For the PSn R tasks,a better classification of the final information dissemination scale can be achieved by using only the first ten forwarding data,which results in a high AUC of 0.815.(3)An information coverage and reinforcement effect prediction algorithm based on feature fusion is proposed.First,considering the importance of information coverage and redundancy effect in describing the influence of information dissemination and strengthening the behavior of audience users,the coverage scale and redundancy scale in the process of information dissemination are defined from the perspective of deeply excavating the effect of information dissemination.Next,the correlation features with information coverage and redundancy phenomena are extracted based on the empirical analysis of popular information dissemination links.Finally,an information coverage and reinforcement effect prediction algorithm based on feature fusion(FFPR-M)is proposed,which yields a goodness of fit of 0.831 and 0.812 for predicting the coverage scale and redundancy scale,respectively.This study provides a new perspective for information diffusion measurement methods and prediction problems,which helps deeply explore the factors that affect information diffusion.This study is of theoretical significance for grasping the laws of information dissemination,predicting the scope of information dissemination and reinforcement effect,and is of important application value for network marketing and public opinion analysis on online social platforms.