Rumor Detection in Social Network via Influence Based on Bi-directional Graph Convolutional Network

Nowadays, social media has become a convenient and prevalent platform for users to communicate with others and share their opinions publicly. In the meantime, due to the rapid growth of social media, the circulation of untrue and irresponsible statements is also boosted, making it harder to detect rumors in the massive amount of social data. Existing deep learning-based approaches detect rumors by modeling the way they spread or their semantic features. However, most of them ignore the different levels of influence when various users participate in the spread of rumors. Hence, we define the influence power of users, which is related to the popularity of their posts, as influence factors, and users with higher influence factors are more likely to determine the direction of public opinion, which can also make rumors spread more quickly and widely. In this paper, we propose a novel graph model named Influence-based Bi-Directional Graph Convolutional Network (IBi-GCN) to capture the influence of users and the way a rumor spreads. First, our model uses an information entropy-based approach to calculate the local and global influence of users, respectively, and obtain the overall influence factors of users in the form of a weighted sum. Second, we combine the overall influence factor with the two main features of rumor propagation and diffusion. Finally, we use a bi-directional graph convolutional neural network to learn a high-level representation for rumor detection.