Research On Popularity Prediction Of Online Social Media

Popularity prediction of online social media is aimed at predicting online items' popularity in the future based on the interactive data between users and items. Research on populairty prediction of online social media has great meaning in theory and practice.From the point of theoretical view, popularity prediction can help to understand what factors may result in items to be popular, how the dynamic evolution and microscopic behavior influence items' popularity. From the point of practical view, popularity prediction of online items can help online users filter useless information more efficiently,help online service providers organize and manage their platforms better, and provide advertisers more appropriate strategies. Traditional popularity prediction methods always predict popularity of items at a macro view, but these methods can not predict “potential items” which are not popular ones at current but will be in the near future, and also have little contribute to interpret why items become popular. Therefore, we analyze art of the state of popularity prediction, propose some approches to solve the above problems by introducing recommendation algorithms and social relationships. The main works are as follow:1. Based on the basic workflow of popularity prediction, we deeply analyze the commonly evaluation index; summarize commonly popularity prediction methods, and indicate the application scenarios and workflow of different menthods. Lastly, popularity prediciton methods are studied and their advantages are given.2. This thesis introduces recommendation algorithms to popularity prediction, and proposes a framework of recommendation-based popularity prediction method. This method firstly explores users' future interests on items by recommendations algorithms,then accumulates differnet user's interests as the predicted popularity. All predictors are verified on three dataset(Movielens, Netflix and Digg), it's found that our method has better performance on digging “potential items”. Moreover, analysis on the dataset of Digg shows that the prediction performance especially for “potential items” can be further enhanced if the user social network is known.3. This thesis proposes a social-based popularity prediction method. This method makes use of the early social relationships data, constructs social network, and extracts features from the network to improve the accuracy of popularity prediciton. Moreover,we define and quantify spread leader for the first time, extract features based on the defined spread leader, then apply different machine learning models to predict. Our method is verified on Digg, it's found that the performance of prediction can be imporved by introducing network features and spread leader features. What's more, we find that votes between friends go against that items become popular, and there is a positive correlation between our spread leaders' selections and the popularity of items. The above results throw light on why items to be popular from the aspect that how user's choice influences item's popularity.