Considering the current trends in radical thinking and radicalism, it’s crucial to understand how radicalization works on online social networks.
A study on radicalization influence has been presented at the Web Science conference 2018 in Amstedam by Miriam Fernandez, Moizzah Asif, Harith Alani (Open University, KMI, UK), based on “roots of radicalization” model taken from social science models. The work revolved around three steps:
Three levels have been considered:
- Micro-influence: internal, personal influence, represented by posts on Twitter posted by the person.
- Meso-influence: small-group influence, represented by other people posts, shared.
- Macro-influence: global influence, represented by public web sites (this part was not covered because most of the site are actually blocked by law enforcement agencies)
Radicalization has been studied by analyzing radical speech style (n-gram model), and people have been compared to that through cosine similarity for the language.
Using a Kaggle dataset of 112 general and 112 pro-ISIS radical users, two analyses were run.
Can I understand if a user is actually radicalized?
The first analysis is about classification of radical and non-radical users, based on personal and social (meso level). Results show that the two aspects are quite correlated:
Users can thus be rather easily classified in the two categories, even with pretty simple classifiers. Naive-Bayes classification already reached 90% precision.
Can I predict if a user is going to be influenced and thus radicalized?
The second analysis looked into prediction of radicalization in time. To do that, the users’ timelines were split in two time-dependent sets, the first 80% of the post are used training and the newest 20% for testing. As a result, for radicalization prediction the precision is higher for the neutral user group, and recall is higher for the pro-ISIS radical group, but in any case was not so great. This is possibly due to the possible flexibility that should be considered in the time window for radicalization.