Combating Misinformation on Social Media by Exploiting Post and User-level Information

Misinformation on social media has far-reaching negative impact on the public and society. Given the large number of real-time posts on social media, traditional manual-based methods of misinformation detection are not viable. Therefore, computational approaches (i.e., data-driven) have been proposed to combat online misinformation. Previous work on computational misinformation analysis has mainly focused on employing natural language processing (NLP) techniques to develop misinformation detection systems at the post level (e.g., using text and propagation network). However, it is also important to exploit information at the user level in social media, as users play a significant role (e.g., post, diffuse, refute, etc.) in spreading misinformation. The main aim of this thesis is to: (i) develop novel methods for analysing the behaviour of users who are likely to share or refute misinformation in social media; and (ii) predict and characterise unreliable stories with high popularity in social media. To this end, we first highlight the limitations in the evaluation protocol in popular rumour detection benchmarks on the post level and propose to evaluate such systems using chronological splits (i.e., considering temporal concept drift). On the user level, we introduce two novel tasks on (i) early detecting Twitter users that are likely to share misinformation before they actually do it; and (ii) identifying and characterising active citizens who refute misinformation in social media. Finally, we develop a new dataset to enable the study on predicting the future popularity (e.g. number of likes, replies, retweets) of false rumour on Weibo.