Towards an extensible, heterogeneous graph-based modeling of conflict

Abstract

Forecasting armed conflict remains challenging, and many approaches emphasize spatio-temporal patterns while under-representing human actors and the relations through which violence propagates. We reinterpret conflict event datasets as heterogeneous, time-indexed graphs. These graphs jointly encode actors, civilians, and countries alongside their mutual interactions. Using the UCDP Georeferenced Event Dataset, we show that centrality- and embedding-based analyses of the resulting graphs' topology capture meaningful patterns, including (i) actor groupings by aggression neighborhood, and (ii) a separation between countries that subsequently experience violence and those that do not. We conclude by outlining extensions to our representation incorporating additional data sources and findings from process-oriented research, and by discussing its future use as a foundation for graph-based predictive models, with the potential to improve the performance of conflict forecasts.