Rumor Detection Based on Bi-directional Graph Attention Network

Information dissemination on social media is often accompanied by a large number of false or true rumors. The spread of these rumors on social network platforms does serious harm to social stability. Traditional rumor detection methods use text or image contents to identify rumors, but the performance is not satisfactory. Deep detection models combined with deep learning algorithms are developed recently. In this study, a new bi-directional graph attention network (Bi-GAT) model is proposed for rumor detection, where two directions of rumor information propagation are considered at the same time. Firstly, rumor information is transformed into a bi-directional graph structure form, simultaneously. Secondly, the feature of each node is extracted and input to graph attention network with the graph structure information, where the multi-head attention mechanism is adopted. Thirdly, after being processed via two graph attention layers, the two-directional data are concatenated and input to a fully connected layer. The results of rumor classification can be obtained according to the output of the fully connected layer. The simulation results on some open datasets demonstrate the validity of the proposed model.