On predicting event propagation on Weibo

The enormous amount of tweets generated every day contain remarkably theoretical and practical value in numerous researches and applications. Although the prediction of information propagation has been extensively studied, most previous work mainly focuses on whether the event will break out and when it will burst out. It is still a huge challenge to predict the final size due to the great uncertainty. This study aims at finding out the best combination of the method and features to predict the final size of the certain event. In the best parameter settings, we compare the method of Gradient Boosting Decision Tree (GBDT) with other methods both in same and different strategies, and find that the GBDT performs much better. In the process of feature extracting and computing, we find that the amount of the early uids, the amount of fans, and the field of the topic matter in the early stage of the prediction while other features begin to show differences over time. That is to say, we achieve the more accurate prediction with fewer features in the early stage of the event which make sense.