Predicting News Popularity by Mining Online Discussions

Georgios Rizos, Symeon Papadopoulos, Yiannis Kompatsiaris

Abstract: The paper presents a framework for the prediction of several news story popularity indicators, such as comment count, number of users, vote score and a measure of controversiality. The framework employs a feature engineering approach, focusing on features from two sources of social interactions inherent in online discussions: the comment tree and the user graph. We show that the proposed graph-based features capture the complexities of both these social interaction graphs and lead to improvements on the prediction of all popularity indicators in three online news post datasets and to significant improvement on the task of identifying controversial stories. Specifically, we noted a 5% relative improvement in mean square error for controversiality prediction on a news-focused Reddit dataset compared to a method employing only rudimentary comment tree features that were used by past studies.