A holistic system for troll detection on Twitter

摘要

Various techniques based on artificial intelligence have been proposed for the automatic detection of online antisocial behaviors, both in existing systems and in the scientific literature. In this article, we describe TrollPacifier, a holistic system for troll detection, which analyses many different features of trolls and legitimate users on the popular Twitter platform. In this system, the most known and promising approaches and research lines are applied, along with original new ideas, in a form that fits such a large public platform. In particular, we have identified six groups of features, based respectively on the analysis of writing style, sentiment, behaviors, social interactions, linked media, and publication time. As its main scientific contributions, this work provides: (i) an up-to-date analysis of the state of the art for the problem of troll detection; (ii) the systematic collection and grouping of features, on Twitter; (iii) the description of a working holistic system for troll detection, with a very high accuracy (95.5%); and (iv) a comparison among the different features, with a machine learning approach. Our results demonstrate that automatic classification can be useful in the whole process of identification and management of online anti-social behaviors. However, a multi-faceted approach is required, in order to obtain an adequate accuracy.