Title: Conspiracy vs science: A large-scale analysis of online discussion cascades

Abstract: With the emergence and rapid proliferation of social media platforms and social networking sites, recent years have witnessed a surge of misinformation spreading in our daily life. Drawing on a large-scale dataset which covers more than 1.4M posts and 18M comments, we investigate the propagation of two distinct narratives--(i) conspiracy information, whose claims are generally unsubstantiated and thus referred as misinformation to some extent, and (ii) scientific information, whose origins are generally readily identifiable and verifiable--in an online social media platform. We find that conspiracy cascades tend to propagate in a multigenerational branching process while science cascades are more likely to grow in a breadth-first manner. Specifically, conspiracy information triggers larger cascades, involves more users and generations, persists longer, is more viral and bursty than science information. Content analysis reveals that conspiracy cascades contain more negative words and emotional words which convey anger, fear, disgust, surprise and trust. We also find that conspiracy cascades are more concerned with political and controversial topics. After applying machine learning models, we achieve an AUC score of nearly 90% in discriminating conspiracy from science narratives.
We find that conspiracy cascades are more likely to be controlled by a broader set of users than science cascades, imposing new challenges on the management of misinformation. Although political affinity is thought to affect the consumption of misinformation, there is very little evidence that political orientation of the information source plays a role during the propagation of conspiracy information. Our study provides complementing evidence to current misinformation research and has practical policy implications to stem the propagation and mitigate the influence of misinformation online.