A Two-stage Prediction Method of News Popularity only using Content Features

News popularity prediction benefits election, investment decisions, news recommendation, etc. Existing prediction methods using news dynamic trends or propagation characteristics need an observation period after news publication, thus failing to meet the requirements of timely prediction. Aiming at predicting news popularity before its publication, we use features extracted only from news content. Considering differences existing between news columns (financial, sports, entertainment news, etc we propose a two-stage method to predict news popularity. Firstly, we select global features related to column information from all news content and classify the news to different columns. Secondly, we select local features related to news popularity separately from each column of news content, reconstruct news VSM with the selected features, and predict the popularity of each column of news. We collect S687 pieces of news including four columns over 6 months from Sohu News. The average popularity prediction F-value of the four columns is 91.4%. Compared with traditional method which predicts news popularity without considering column differences, our method improves by 2.2%. Furthermore, we find that popularity of social and financial news is more predictable than sports and entertainment news.