Trend Prediction On Social Network

The explosive growth of online social networks has been one of the most exciting events in recent years.It provides people convenient platforms to share information such as Twitter,Facebook and Weibo.Mining popular content from huge information has attracted researchers' attention.Predicting the popularity of content in social network is significant for online marketing,flow control or real-world outcome prediction.This thesis analyzes social relationship and social influence in social network,and propose a framework to predict popularity trend of contents and user reaction.The main contents are as following:1.In this thesis,we analyze dataset from typical social network platform(Weibo),and find out after content publish,users who repost the content(we call them reposters)contributes a lot to its popularity.And the higher status the reposter is,the more important he is.Through analyzing popularity of those data,we also find out that part of the content are potential,which means they are not very popular in the early stage but rise over others in the later stage.Then we propose an idea of users' importance to measure reposters' contribute in content popularity as the basis of prediction.2.This thesis proposes a framework to predict popularity trend of social network content based on reposters in repost series.First we divide the series into T slices according to time,then extract reposter's importance in each slice as T dimension features.We use multiple linear regression and regression tree to predict final popularity.Experiments on Weibo dataset prove that our method preforms better than state-of-art algorithms(increases 36.8% in MAE,and 2.9% in tau).And our method can achieve good results in predicting potential content in early stage.3.This work analyzes social influence in social network.We find out local network can present neighbor's influence to users and model users' repost action as a classification problem.This model uses global social influence and user importance to predict if user will repost a content or not.Experiments on Weibo dataset show that user importance plays a critical role in this problem.Our method performs better than method that only consider user relevance(increses 20.6% in precision).And we also discuss how neighbor's influence aggregate,then conclude that the longer time between neighbor node's repost time and predict time,the bigger influence it has on the target user node.