Data Science and Genocide Prevention

Data science has promise to be a useful tool in the prevention of genocide and related atrocity crimes. The generally agreed upon definition from the United Nations’ 1948 Genocide Convention has significant weight behind it, but attempts to operationalize it has proven to have mixed success. In this paper, we introduce our program toward genocide prevention which, in part, requires quantitative approaches in addition to more traditional methods in genocide studies. We apply data-scientific approaches to operationalize this definition of genocide, in particular, to predict the onset of genocide, and identify preventive policy options. Toward forecasting, we reaffirm some of the risk metrics and methodologies of the Early Warning Project, attaining accuracy scores in retrospective analysis of the forecasting problem, applying a bootstrapping methodology to achieve an aggregate median test-set accuracy of 80% across bootstraps of (F1 score interquartile range: [0.750, 0.875]) with a random forest classifier. Toward operationalizing a genocide definition and quantifying similarity between atrocities, we code a subset of 251 incidents in the Early Warning Project on three guidelines and analyze quantitative predictability of such codings: the labeling task in the test set had median accuracies of 84.9% for Coercive Control, 85.4% for identification of target groups, and 87.4% for Existential Threat (F1 scores: 0.69, 0.68, and 0.51). These reflect varying degrees of significance in operationalizability, particularly when compared against a random label permutation test. Beyond quantitative indicators, we discuss our work analyzing the Shoah Archive, a multimedia corpora of genocide events, as an a way to check genocide definitions externally. Ultimately, the goal of our program is to provide a data storytelling toolkit that communities can use to advocate for policy changes that can prevent genocide and related episodes of identity-based violence. We contextualize our preliminary results and outline directions for the future of the project.