How to Get Endorsements? Predicting Facebook Likes Using Post Content and User Engagement

摘要

We view the prediction of Facebook likes as a content suggestion problem and show that likes can be much better predicted considered post content or user engagement. Experiments presented are based on a dataset of over 4 million likes collected from over seventy thousands of users in fan pages. The proposed model adopts the similarity metric to appraise how a user may like a document given his or her liked documents, as well as the Restricted Boltzmann Machine (RBM) to estimate whether a user would like a document given the like records of all users. The model achieves a precision of 5-10% and a recall of 2-55%. The commonly used label propagation model is implemented and tested as a baseline. Different models and settings are compared and results show superiority of the proposed model.