A Comparative Study of Spatiotemporal Clustering and Classification Approaches for Security Incident Risk Assessment in UN Peacekeeping Operations

Abstract

This study presents a comparative analysis of spatiotemporal clustering and classification approaches applied to security incident risk assessment in United Nations peacekeeping operations. Drawing on 61,274 conflict event records from the Armed Conflict Location and Event Data Project (ACLED) spanning 2010 to 2024 across four active mission areas---MINUSCA, MONUSCO, UNMISS, and UNIFIL---and supplemented by geospatial features from PRIO-GRID, this research evaluates three spatial clustering methods (ST-DBSCAN, Kernel Density Estimation, and Getis-Ord Gi*) for hotspot identification and examines temporal periodicity through STL decomposition and autocorrelation analysis. The study compares Random Forest, XGBoost, and SVM classifiers under natural class imbalance using SMOTE, ADASYN, and SMOTE-Tomek resampling strategies. Results indicate that ST-DBSCAN demonstrates superior flexibility in delineating irregular hotspot boundaries, while XGBoost paired with SMOTE yields the highest Macro F1-Score (0.716) and Matthews Correlation Coefficient (0.683) among all configurations. These findings provide methodological references for data-driven risk assessment in peacekeeping contexts.