While detection of rumours in online social networks has been intensively studied in the literature, mitigation of the spread of rumours has only recently gained attention and remains a challenging task. Some studies developed user opinion models to find top influential users as debunkers to spread the truth to counter rumour spread. Other studies designed an intervention framework to optimize the mitigation activities for given debunkers. The issue of optimizing the selection of debunkers in a dynamic environment where users’ beliefs and behaviour change remains under investigated. This paper addresses this issue by proposing a rumour mitigation approach based on the deep reinforcement learning framework. In particular, we model the changes in users’ beliefs with an epidemic model. We further employ deep reinforcement learning to train an agent to learn a multi-stage policy for selecting the optimal debunkers to inject truthful information under a budget constraint. Our model selects debunkers to inject truthful information at multiple stages with an overall objective to maximize the number of users who will believe in the true information (a.k.a number of recovered nodes), such that the spread of rumours is minimized. Our experiments on synthetic and real-world social networks show that our proposed method for rumour mitigation can effectively minimize the spread of rumours.

Understanding the spread of rumors on online social networks (OSNs) is crucial for designing strategies to detect and mitigate rumor propagation. Previous studies analysing rumor propagation have focused on summarising the static measurements of propagation-based information cascades. But static features are unable to capture the dynamic nature of information propagation across time. In this paper, we employed two generative models, Multivariate Hawkes process (MHP) and marked Hawkes process (marked HP) to model user influence and the dynamics of rumor propagation. Using the MHP model, we were able to derive a novel measurement of user influence in information propagation, namely, the influence rate. We then employed the marked HP model and considered various mark measurements including the proposed influence rate to provide new insights into differentiating between rumor and non-rumor propagation, and among different types of rumor propagation. Our analysis on Twitter rumor datasets clearly showed that users play different roles (i.e., possess different influence rates) across different categories of source tweets. Moreover, different categories of source tweets have different patterns of diffusion. In particular, rumor cascades typically attracted more influential users at the early stage of cascades, and they are more likely to generate more retweets than non-rumor cascades. Among different types of rumors, false rumors diffused faster than true rumors.