Research On Prediction Methods Of Content Popularity In Social Networks

Recent decades have witnessed the rapid development of online social networks.The rise of social network has changed the way people communicate and share information,and social network is gradually becoming the mainstream media in the era of information.Currently,a large amount of user-generated content(UGC)is available from online social network sites.Popularity prediction has been studied in diverse online contexts with demonstrable practical,sociological and technical benefit.Predicting the popularity of online content is an important issue for uncovering rules governing collective human behaviors in networks.Also,content popularity prediction finds application in an array of areas.However,it is not trivial to make predictions due to the myriad factors that influence a user's decision to reshare content.In view of this,we use the interdisciplinary ideas and methods to study the prediction of the final popularity,popularity dynamic,cascade increment size and popularity ranking of online content in social networks.Our work focus on exploring adaptive peeking window,and inherit the advantages from both the generative model and the feature-driven method to predict the eventual size of the information cascade,which leads to a good solution of the tradeoff between prediction accuracy and the interpretative result.Our work may help to understand the evolution process of content popularity in social networks,to understand the complex group behaviors on the network,and also provide some of exploratory theoretical results for the study of human behavior.This research was funded by the National Natural Science Foundation of China(No.61172072,No.61271308),the National Science Foundation for Young Scientists of China(No.61401015)and the Fundamental Research Funds for the Central Universities(Grant No.2017JBZ107).Our work makes the following contributions:1.We propose a generative approach using a Hawkes process,the purpose of which is to predict the eventual popularity of online content from the observed initial period of the information cascade.Our approach is distinguished from existing approaches via its ability to explore an adaptive peeking window and its added remarkable predictive power that was inherited from the feature-driven method.Further,we proposed a procedure for exploring an adaptive peeking window,which allows us to answer the question,"How can we obtain the most effective part of the history to make an accurate prediction?" And the added predictive layer bridges the gap between the generative method and the feature-driven method.Empirical studies on real world datasets demonstrate that the proposed method significantly outperformed the existing approaches.2.From the perspectives of "event" granularity and "time" granularity,we study the methods for predicting popularity dynamics.From the "event" granularity perspective,we develop a statistical model based on the theory of Hawkes processes.From the "time" granularity perspective,as single time series prediction models are usually targeted for homogeneous popularity patterns,majority of them are incapable of handling real-world popularity dynamics whose patterns may be heterogeneous.By considering multi-class regressions and historical prediction performance of each sub-model,we generate the combined weights of each sub-model for future time prediction,and establish a combined prediction framework that combines the predictive capabilities of multiple traditional time series models(ARIMA,ML,SVR).With the help of multi-class regression,the weights of predictors in this framework can be obtained in real-time based on their accuracy for current workload.3.This paper presents a novel method for predicting the increment size of the information cascade based on an end-to-end neural network.Learning the representation of cascade in an end-to-end manner circumvents the difficulties inherent to hand-crafted features design.An attention mechanism,consisting of the intra-attention and inter-gate module,is designed to obtain and fuse the temporal and structural information learned from the observed period of the cascade.The experiments are performed on two real-world scenarios,i.e.,predicting the size of retweet cascades on Twitter and predicting the citation of papers in AMiner.The efficiency of the proposed method is evaluated with respect to several state-of-the-art cascade prediction methods.4.This paper proposes a new approach to predict the popularity of content in the Chinese microblogging website Sina Weibo.There are four operations in Sina Weibo,including post,repost-only,repost-and-comment,and comlent-only.We model these operations as a bipartite graph,which takes the temporal factor into account by assigning edge weight as an exponential decay function.We then propose a regularization framework on this model to predict the original post's future popularity.Experimental results show that our method outperforms other methods in predicting the post5s future popularity,especially for short-term prediction.