Dissertation Defense - Department of Geography and Geoinformation Science

PhD Candidate: Pamela Kanu

PhD of Science, Earth Systems and Geoinformation Sciences

Department of Geography and Geoinformation Science 

Title: Predicting rebel group conflict By using a Bayesian Neural Network and a Bayesian Hierarchical Model

Dissertation Chair:  Dr. Liping Di (GMU, GGS, CSISS)

Committee Members:

Dr. Maction Komwa (GMU, GGS)

Dr. Taylor Anderson (GMU, GGS)

Dr. Seiyon Lee (GMU, Department of Statistics)

Abstract: 

Mali has been ravaged by armed conflicts since the early 21st century, leading to the displacement of thousands of civilians and the proliferation of armed conflict across the country. The conflict is characterized by a complex web of state and non-state actors, each with their own agenda, leading to a complicated security situation and humanitarian disaster. In order to support the peacekeeping, humanitarian relief, and normal civilian socioeconomic activities in Mali, there is an urgent need to develop a risk map that depicts the likelihood of encountering rebel groups at various geographic locations in the country.

This dissertation research proposes to create such a map with the spatial resolution at the observation level by using innovative Bayesian statistical and machine learning models. The research activities include 1) understanding the proliferation of armed conflict across Mali and its evolution, the underlying factors that drive it, the actors involved, and their intent; 2) developing a new index, Rebel Group Risk Index (RGRI) that captures a comprehensive measure of the risk of encountering rebel groups in a given geographic location based on the understanding; 3) creating the RGRI maps with two Bayesian approaches, a Bayesian hierarchical model and a Bayesian Neural Network model, based on past incidents of rebel activity and conflict and the related occurrence of fatalities. Negative binomial regression with random effects is used to account for the spatial and temporal dependencies in the data and Integrated Nested Laplace Approximation-Stochastic Partial Differential Equation and stochastic variational techniques are used to estimate the posterior distributions; and 4) validating the models, evaluating the accuracy of RGRI maps, and determining the performance of the individual models through model diagnostic runs on historical data.

The validated models and the RGRI map created with the models for Mali can have implications to serve as a useful tool for policymakers and stakeholders in conflict prevention to guide resource allocation towards areas with the highest risk of violence.