Mining Semantic Variation in Time Series for Rumor Detection Via Recurrent Neural Networks

Social media has promoted the dissemination of information, which becomes an ideal platform for rumor diffusing. Due to debunking rumors automatically is crucial to minimize the adverse effects of rumors, numerous hand-crafted features based approaches have been proposed. However, as the information spreading in social media is usually large scale, feature engineering is time-consuming and labor-intensive. Besides, social media is dynamic and complicated which makes features difficult to cover the potential features in new states. In this paper, we propose a novel method which learns the variation of semantics over the entire life-cycle of events for rumor detecting. In this model, a bidirectional recurrent neural network (RNN) model with attention mechanism is utilized to learn high-level semantic representations automatically. And another RNN model captures the variation of semantics. Experimental results on datasets from two real-world microblog platforms show that (1) our model outperforms other state-of-the-art rumor detection models; (2) by using hierarchical attention mechanism, our method can select informative posts and distinctive words as features; (3) our model can detect rumors more accurately at an early stage.