A classification method to detect if a tweet will be popular in a very early stage

Abstract

Timely prediction of the popular tweet is of great value in monitoring public opinion, marketing, emergency detection, personalized recommendations and other areas. This paper makes two improvements of predicting popularity of a tweet in Microblog. On one hand, we proposes some dynamic features, such as the retweet depth, retweet width, the total fans of the retweeters, to improve the prediction accuracy. On the other hand, we sharply shorten the time to detect whether a tweet will be popular by putting forward a method called LR-DT which combining the linear regression and the decision tree. We firstly use the linear regression method to predict the dynamic features' amount an hour later after the tweet transferred and then combine them with some static features into the decision tree classifier to detect the popularity of tweets. Our experiments are based on the real data set from Sina Weibo, which is the most popular micro blog service platform in China . The results show that we proposed method effectively identify the popularity of a tweet in less than 5 minutes while with little lose of accuracy.