A Generative Model of Grid-cell Level Violence

• In Event: Forecasting Conflict

Abstract

There has been an increasing focus on conflict predicting and forecasting across international relations. However, researchers have been forced to choose between either simple, low performing conceptually-justified models or flexible, high performing black box algorithms. This leaves us with a scientific gap in conflict prediction. If we are relying more and more on flexible models, that act as black boxes, how are we to understand and draw insights from these workflows? To understand the signals that are being learned, then we need to build models that encode these complex signals. Thus, for scientific and policy reasons, it is important to build generative models of violence at useful temporal and spatial scales, alongside more flexible black box models. these generative models are by definition explainable because they were a) built by humans, and so each component has a qualitative justification, and b) they can be used to simulate new realizations and scenarios from first principles. We set out a new generative model of violence at the grid-cell level. We can then illustrate how racing a theoretically driven Bayesian model against a flexible black-box classification algorithm can lead to new insights and improvements in our understanding of conflict that would not be possible if black-box strategy or the theoretically-informed model was utilized in isolation. Our results highlight how advances in machine learning can accelerate social science research on conflict prediction.

Authors

• Michael P. Colaresi, University of Pittsburgh

• Kevin T Greene, University of Pittsburgh

• Leonardo Fernandes, University of Pittsburgh