Improved Early Violence Detection through Dynamic Graph Classification

Early detection of intracountry outbreaks of violence or intercountry conflicts plays animportant role in geopolitics. Generally, it is based on the use of handcrafted countryfeatures (e.g., demographics, economics) that may serve as proxy to the event ofinterest. While such features matter, it is often their evolution over time both at thecountry level and among countries. As such, we propose to use a graph-basedapproach that allows us to capture static, structural and sequential measures. Usingwell-known data sets, we show how this approach outperforms both baseline andstate-of-the-art approaches.