A Multi-View Rumor Detection Framework Using Dynamic Propagation Structure, Interaction Network, and Content

Social networks (SN) have been one of the most important media for information diffusion in recent years. However, sometimes SN are used to spread rumors, which results in many social issues. Many researches have been done to detect rumors automatically. Previous works mostly exploit a single modality, especially the textual content, thus ignoring other modality such as the propagation structure and the interaction network of the rumor. However, the interaction network of users and tweets, and the propagation structure of a tweet, can provide important information to be used in rumor detection. In this paper, we propose a multi-view rumor detection framework (MV-RD) which captures multiple views of a tweet simultaneously including propagation structure, interaction network, and content. Previous works that considered propagation structure, mostly have used the final propagation structure at the end of information diffusion. Few related researchers have considered the dynamic evolution of propagation structures. In this paper, using partitioning of propagation structure over time, we have designed a propagation structure model that learns the evolution of the propagation structure of rumors over time. Besides, we take advantage of features of the rumor interaction network (modeling first-level interactions of tweets) for detecting rumors using the interaction network model. Also, a content model is learned to detect rumors using the tweet contents. Finally, these three models are fused into a unified framework. The results show the effectiveness of using multiple views in the rumor detection task. The proposed framework can detect rumors more effectively than other existing methods, even without using the tweet content. The proposed method achieved accuracies of 77.82%, 85.65%, and 88.26% by leveraging the propagation structure model alone, combining the propagation structure and interaction network models, and incorporating all three models, respectively. These results outperformed previous approaches and also demonstrated the method's capability to detect rumors earlier than existing methods.