D6.1 Computational Behaviour of radicalisation

This deliverable summarizes research efforts of BGU concerned with computational models of deradicalisation. We focussed this analysis on integrating data from social media with a lexical analysis of the I-GAP spectrum, following our collaboration with WP3 and WP4. The focus of work in the design of novel computational tools for radicalisation detection from social media. Social media sites are increasingly being used by radical organizations as platforms to broadcast their ideology and recruit followers and finance. Prior work has established a ’radicalization pipeline’ in YouTube, driven in part by its recommendation algorithm, that potentially exposes users to increasingly radical content, in some cases leading to verbal and physical violence. Our goal was to provide a computational model for early detection of such individ- uals. We study two research questions. First, how does extremism portray in users’ ac- tivities on YouTube and how does this activity vary over time? Second, can we predict whether users are at risk of radicalization; that is, will users with a history of activity in communities with milder versions of radical ideologies, transition to participate in more extreme communities? We find that there exists a significant rise in extremism portrayed in users’ com- ments relating to key issues known to drive polarization and violence, and that users at risk of radicalization exhibit significantly different engagement behavior on the site. We compare the performance of different machine learning models for predicting risk of radicalization among individuals using features that are informed by users’ com- menting and engagement behavior. We show that combining both of the feature families leads to best performance, and that the learnt model is able to detect relevant users at risk in the upcoming 12 months with just a week’s worth of activity data. Thus we can potentially support those providing support for people at risk of radicalization in real time. We also provide a detailed description of the integration of the computational work with with the Drad I-GAP spectrum (Section 7)