Dottorato in DATA SCIENCE

Titolo della tesi: Defining a Framework for Conflict Susceptibility Mapping

The primary aim of this research is to develop a comprehensive framework for Conflict Susceptibility Mapping using publicly available data and advanced machine learning techniques. The objective is to create a methodology for identifying areas susceptible to conflict based on different conditioning factors related to socio-economic, political, climate change, environmental, and various other indicators. This research used only publicly available historical conflict and conditioning factor datasets as sources. Central to this research is the methodological approach that includes data search, collection, preprocessing, and the application of machine learning models. This highly flexible methodology can accommodate various data types and formats from publicly available data sources. The use of public data also requires thorough data preprocessing practices to ensure data quality suitable for machine learning models. This multi-tiered approach includes data preprocessing, machine learning applications, and geospatial visualization. It provides a blueprint adaptable to different geographical areas and can be modified to various conditioning factors relayed to different conflict types and specific conflicts. Moreover, the methodology can also accommodate a variety of machine learning algorithms, depending on specific requirements or data formats. Due to Somalia's complexity and protracted conflict history, Somalia's public datasets were selected as a case study and application of this framework. An academic literature review was conducted on applying the conditioning factors and past machine learning projects for conflict and peace studies. Additionally, a methodology was developed for applying machine learning techniques to forecast conflict likelihood. It includes data preprocessing, model development, training and validation, and geospatial visualization. The data experiments and fine-tuning of the methodology were conducted using non-homogeneous publicly available data for Somalia. The effective data preprocessing and data standardization addressed the challenges associated with the non-homogeneity of the datasets. The framework was applied to Somlai public datasets and implemented using machine learning models, including the Random Forest Classifier (RFC), Support Vector Machine (SVM) Classifier, and Gradient Boosting Classifier (GBC). These models were instrumental in identifying complex conflict patterns leading to potential conflicts. The results of predictions of machine learning models were validated using metrics like Area Under the Curve (AUC) of Receiver Operating Characteristics (ROC) and Precision-Recall Curves (PRC) and provided positive results. The proposed framework's primary characteristics are flexibility and scalability, with the integration of diverse machine learning algorithms and publicly available datasets. Ultimately, the goal is to produce a prediction tool to assess the likelihood of conflicts based on a set of conditioning factors and display areas susceptible to conflicts on a map. This work is also intended to contribute to the broader field of conflict studies by providing machine learning tools and frameworks for conflict analysis and conflict mediation.