Deep reinforcement learning based ensemble model for rumor tracking

Abstract Fully automated rumor defeating is meaningful for reducing hazards of misinformation in social networks. As one of the automated approaches, content-based rumor defeating is a pipeline that could be divided into four sequential sub-tasks: detection, tracking, sentence classification, and veracity. Specifically, rumor tracking gathers relevant posts and filters unrelated posts for a potential rumor news, which is significant for rumor defeating and has not been studied extensively. However, the existing proposals only consider rumor tracking as an auxiliary task in multi-task learning without special optimization, therefore restraining the accuracy of tracking performance. To this end, we propose a deep reinforcement learning based ensemble model for rumor tracking (RL-ERT), which aggregates multiple components by a weight-tuning policy network, and utilizes specific social features to improve the performance. Finally, we conduct experiments on public datasets and the experimental results show the superiority of RL-ERT on efficiency and effectiveness.