Conflict Analytics: a Predictive Model to Forecast Violent Conflicts in Ethiopia for Improved Early Warning Systems
No Thumbnail Available
Date
2023-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
In the context of an increasing number of violent political conflicts across nations and societies, understanding and predicting violent conflicts that lead to significant economic, social and humanitarian consequences is both an academic interest and a moral obligation. Research on conflict prediction is critical to offer input for policymakers to see potential conflicts and devise strategies for conflict early warning. However, the application of data analytics tools to analyze the dynamics of violent conflicts in Ethiopia has hardly existed. With the motivation to fill this knowledge gap, this research aims to develop a model using supervised machine learning algorithm that can best forecast violent political conflicts in Ethiopia in terms of the dominant conflict types/categories and regions at risk of conflict incidents. Methodologically, this research has employed an experimental research design and adopted the CRISP- ML framework. Predictive analytics tools, as well as three algorithms (random forest, gradient boosting and Gaussian naive Bayes), are used. Open-source software called Jupyter Notebook is used for analysis. The research combined past and recent conflict incidents data with political, economic, social and environmental data from 2007 to 2022. After collecting data on both independent and dependent variables from open databases of research institutes and international organizations, models were built, compared and assessed using a new dataset of variable values projected for the next five years (2023- 2028). The finding shows that the Gradient Boosting machine learning algorithm has a better performance than Random Forest and Gaussian naïve Bayes in predicting the location and types/categories of conflict classes individually. In predicting types of violent conflicts, while Gradient Boosting has a testing accuracy of 57%, both the Random Forest and Gaussian naïve Bayes has 55%. Yet, in terms of predicting the location of conflict incidents, all have 48% testing accuracy. However, when both type/category and location of conflict incidents were considered together the predictive performance of the selected algorithms declined to 30% testing accuracy. In conclusion, the performance of conflict-predicting models is determined by whether the classes (type and locations) are merged and predicted concurrently or separately and independently. The best performing algorithm(57%), such as the Gradient Boosting machine learning algorithm, predicts that Ethiopia will continue to be at risk of violent conflicts for the coming five years, 2023- 2028. The location and type of violent political conflicts shift yearly. With its original findings, this research can make valuable contributions to policy-making and academic fields, such as Information Science, Conflict Studies and others.
Description
Keywords
Predicative Analytics, Machine Learning Algorithms, Conflict Prediction, Conflict Early Warning, Ethiopia