Achieving Participatory Smart Cities by Making Social Networks Safer

Abstract

Cases of organised disinformation campaigns on Twitter, including those reported by the social network itself in its Transparency centre, continue unabated. The negative consequences of these attacks in processes of great importance to societies, such as electoral processes or vaccination campaigns, have sparked research into detecting this type of malicious user. State-of-the-art models for bot detection use numerous information collected from profiles, tweets, or network archi-tecture to obtain competitive outcomes. On the other hand, these models allow for post-hoc detection of such users because they rely on fixed training datasets to classify users based on their previous activities. In contrast, we propose a proactive technique that uses user records to predict dangerous attacks before they occur as a measure to make social networks safer, fairer and less biased. For this purpose, our method uses a model that predicts malicious assaults by projecting users' embedding trajectories before completing their actions. We employed a Dynamic Directed Multigraph representation of temporal inter-actions between people in the Twittersphere for the experiments. By comparing them in the same data, our model outperforms state-of-the-art methods by 40.66% in F-score detecting malicious users preemptively. In addition, we propose a model selection study that evaluates the usefulness of several system components.