MMSPP: Multimodal Social Media Popularity Prediction

摘要

People post and share news on social media in daily lives. Popular posts get attention and further influence the society. It’s helpful to know in advance if certain posts will become popular. In order to achieve this, several challenges need to be resolved: (i) more posts become multi-modal – smartphones make it easy to capture and share images and videos with texts in posts, and (ii) the popularity of posts depends not only on the quality of the content, but also on the popularity of the post author themselves. To this end, we propose MMSPP, a multimodal social media popularity prediction system incorporating image, text and author social features of a post to predict its retweet counts and favorite counts in Twitter as the measurements for popularity. Specifically, two novel multimodal fusion methods are proposed. We evaluate MMSPP with the two fusion methods in a Tweeter dataset with 3448 image-text pairs of posts. MMSPP is able to beat three baselines by at least 12.50% and 7.69% MSE for retweet counts and favorite counts respectively.