Spatio-temporal violent event prediction using Gaussian process regression

Large volume, data-driven violent conflict research is now possible using publicly available data sets. This work analyzes the predictive ability of data-derived Gaussian process models compared to a generalized linear model. Societal violence is a highly nonlinear process and the available data sets have high dimensionality that yield observation totals in the hundreds of thousands to millions. These challenges make machine learning modeling difficult without significant dimensionality reduction. We develop a computationally intensive Gaussian process modeling approach that exploits the size and complexity of the violent conflict dataset to identify appropriate basis vectors for the model. We develop our models using gridded monthly violent event counts for sub-Saharan Africa from 1980 to 2012. Our resulting Gaussian process models modestly improve the accuracy and predictive ability of existing generalized linear models. Despite this improvement, the accurate prediction of violence in sub-Saharan Africa at a relatively fine resolution spatial grid of 1 $$^\circ$$ ∘ latitude/longitude remains a challenging problem.