GRU with Level-Aware Attention for Rumor Early Detection in Social Networks

Social networks have developed rapidly in recent years, which gives people the opportunity to access more information. Nevertheless, social networks are illegally used to spread rumors. To clean up the network environment, rumor early detection is urgent. But the detection accuracy of existing rumor early detection methods is still not high enough in the early stage of news spread. They neglected to focus on more critical tweets by measuring the contribution of each retweet node to news events based on the propagation structure feature. In this paper, a novel method based on Gate Recurrent Unit with Level-Aware Attention Mechanism is proposed to improve the accuracy of rumor early detection. In this method, text features and user features are extracted from tweets about a given news event to generate a unified node representation of each tweet. Meanwhile, the process of news propagation is simulated to encode the node’s forwarding level according to time span and forwarding hops between source tweet and retweet. In order to pay different attention to tweets according to the forwarding level of nodes, a new method based on attention mechanism is proposed to update node representation. Finally, the news event feature is learned from related tweet nodes representation in time sequence via a GRU-based classifier to predict the label (rumor or non-rumor). Extensive experimental results on two real-world public datasets show that the performance of the proposed model is higher than the baseline model in the early stage of rumors spread in the social network.