Hot Topics Diffusion Analysis And Prediction In Social Networks

Online social networks are now recognized as an important platform for the diffusion of information.In recent years,network events have emerged in endlessly in our country,and network public opinions have been blowout.Under this background,analyzing and predicting the diffusion of hot topics are not only necessary to grasp the popularity evolution trend of hot topics from a macro perspective,but also need to grasp the changes and source of public opinions in the diffusion.Studying the related technologies of social network hot topics diffusion analysis and prediction will greatly help us to understand social behavior and social phenomenon.Besides,it is of great significance to national security and social stability.In this dissertation,we focuse on the three core issues,including hot topics detection and tracking,popularity of hot topics and source locating of hot topics.We carry out our work on hot topics dynamic detection and tracking on large-scale short text stream,popularity evolution analysis and prediction of hot topics based on deep learning,diffusion analysis and source locating of hot topics based on diffusion relationship and social network hot topics diffusion analysis and prediction platform,taking microblog as an example.The main work and contributions are listed as follows.(1)Traditional methods of topic detection and tracking do not work well in large-scale short text stream.To solve this problem,we propose a method of hot topics dynamic detection and tracking on large-scale short text stream.Through the empirical analysis of hot topic datasets,we explore the characteristics and rules of hot topics.Our proposed method implements spatiotemporal anomalies detection and related hot topics dynamic detection and tracking,solving the sparse of short text and improving the processing efficiency of large-scale short text.(2)Popularity evolution analysis and prediction mainly focus on predicting the popularity at a given time or final popularity.They only analyze how different factors affect the final popularity,without considering that different factors play different roles in different stages of popularity evolution.To understand popularity evolution pattern under information diffusion,we propose a three-dimensional feature(average,trend,cycle)model to fit and predict popularity,and a popularity prediction model based on deep learning.First of all,the evolution law of hot topics' popularity is analyzed,and a three-dimensional feature model is defined:average,trend and cycle.Based on the three-dimensional feature model,a time series model is constructed to fit the evolution of hot topic popularity.The short-term popularity of hot topics is predicted.Compared with SpikeM and SH models,our model proposed in this paper has higher fitting degree and accuracy.Then,from the three key stages of popularity evolution:burst,peak and fale,a popularity prediction model based on deep neural network is proposed.The different roles of each factor in the process of popularity evolution are analyzed in detail,and the active period of popularity is predicted.Compared with SpikeM and SVR models,our method proposed in this paper has better effect and performance in terms of timeliness and accuracy of prediction.(3)The existing source-locating methods have some shortcomings,such as poor adaptability,unreasonable assumptions and poor operability,under these condition including complex structure of social networks,evolution of topics at any time,incomplete observation and others.To solve these problems,we propose a topic diffusion analysis and source locating method based on diffusion relationship.First of all,diffusion relationship is analyzed qualitatively and quantitatively.Secondly,diffusion network of hot topics are constructed based on forwarding cascade,to find key users and information sources in single source diffusion network.Thirdly,the method identifies the possible source and the suspected superior node of each source by judging the diffusion relationship whether the diffusion is certain or uncertain.Then.the correlation strength between each source and the suspected superior node is described.Finally,each source is judged whether it is topic source by comparing the correlation strength with the threshold value,locating the real topic source accurately.(4)A platform for hot topics diffusion analysis and prediction is designed and implemented.Based on the existing microblogging data crawling system,we implement some methods proposed in this thesis for hot topics dynamic detection and tracking,hot topics popularity analysis and prediction,trend analysis and topic source locating.