In plain words
Facebook and Twitter are studied endlessly, but WhatsApp — where much of the world actually organises — is a research blind spot: end-to-end encryption means there are almost no open datasets of group chats, even as groups have been used to spread riots and radicalization.
This chapter proposes a framework that combines NLP, semi-supervised machine learning, pattern mining and topic detection to estimate the probability that a WhatsApp group poses a civil-unrest threat. It is deliberately a framework: designed to be field-deployable, with prediction quality that scales as more real data becomes available to feed it.
How it works
- Acquire what's possibleWorking within encryption constraints, available group text is collected and preprocessed.
- Detect the topicsTopic detection surfaces what a group persistently talks about.
- Learn with few labelsSemi-supervised learning stretches scarce labelled examples across the unlabelled majority.
- Score the groupPattern mining and classification combine into a threat probability for the group as a whole.
Why it matters
Radicalization migrated to encrypted platforms precisely because nobody could watch. A principled framework for threat scoring under those constraints — published in 2019, before this became mainstream concern — is why this remains my most-cited work.