Unsupervised Post-Time Fake Social Message Detection with Recommendation-aware Representation Learning

This paper deals with a more realistic scenario of fake message detection on social media, i.e., unsupervised post-time detection. Given a source message, our goal is to determine whether it is fake without using labeled data and without requiring user interacted with the given message. We present a novel learning framework, Recommendation-aware Message Representation (RecMR), to achieve the goal. The key idea is to learn user preferences and have them encoded into the representation of the source message through jointly training the tasks of user recommendation and binary detection. Experiments conducted on two real Twitter datasets exhibit the promising performance of RecMR, and show the effectiveness of recommended users in unsupervised detection.