Predicting armed conflicts : a machine learning approach

Details

Supervisors

Rosoux, Valérie ; Jungers, Raphaël M.

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

Political conflicts may not be the first cause of death worldwide, they constitute nevertheless a major source of destruction and prevent many countries from developing themselves and providing stability, health, and security to their citizens. Assessing the risk of conflicts in every country with evidence-based predictions is now a duty for the governments and international organisations. In this thesis, we address the problem of conflict prediction through a machine learning approach. We aim at predicting whether there will be a conflict which corresponds to a quarrel with at least 25 casualties within the next four years, either at national or subnational scale. The prediction is based on social, demographic, economic, geographic and political national indicators measured on a yearly basis. First, we determine which variables are significant using analysis of variance. Then, using mutual information, we measure the strength of the relationship between pairs of input variables and between input variables and the outcome. Using mutual information in the context of conflict prediction is a contribution of this work. After this, we use logistic regression to classify the instances in two categories : future peace or future conflict. Then, with a random forest model, we obtain better performance than previously reported for the dataset that we use. In the last section, we address the problem of the interpretability of the results using SHAP values, a method able to generate an explanation model such that it relates the predicted probability of conflict and the input variables by a linear function. We show that the lack of democracy is not a determining variable for our algorithms whereas it is frequently used to justify international interventions by foreign countries. In addition, our observations suggest that the best predictors for future onset are the current intensity of conflict, the amount of structural constraints and the government's effectiveness when the fight is for national power. When the conflict is at subnational level, the best predictors are the current intensity of conflict, the population size and the level of repression.