Social Media Rumour Detection Through Graph Attention Networks

Rumours are unverified statements or news that spread quickly across the Internet. The global ubiquity of social media platforms provides the perfect conditions for the spread of rumours. Such rumours can have global consequences. Tools for detection of rumours are therefore needed. Diverse methods have been applied to discover rumours through approaches based on text mining, propagation patterns and user networks and their interactions. Such approaches treat user interactions in discussions equally. In this paper, we propose a model to extract information from user interactions based on Graph Attention Networks. In the propagation graph, the nodes represent the user text content and the edges represent the reply interactions. The attention mechanism is implemented to determine the edge weights between node pairs. We conduct experiments using Twitter15, Twitter16, and PHEME datasets and achieve state of the art results.