A relevance index-based method for improved detection of malicious users in social networks

The phenomenon of “trolling” in social networks is becoming a very serious threat to the online presence of people and companies, since it may affect ordinary people, public profiles of brands, as well as popular characters. In this paper, we present a novel method to preprocess the temporal data describing the activity of possible troll profiles on Twitter, with the aim of improving automatic troll detection. The method is based on the zI, a Relevance Index metric usually employed in the identification of relevant variable subsets in complex systems. In this case, the zI is used to group different user profiles, detecting different behavioral patterns for standard users and trolls. The comparison of the results, obtained on data preprocessed using this novel method and on the original dataset, demonstrates that the technique generally improves the classification performance of troll detection, virtually independently of the classifier that is used.