Predicting the Multi-Factor Trigger Probability of Major Urban Security Incidents Using the XGBoost Ensemble Algorithm

Abstract

Accurate prediction of major urban safety incidents is essential for proactive emergency management and intelligent infrastructure protection, particularly in modern smart cities supported by distributed sensing and communication networks. To address the limitations of conventional statistical and machine learning approaches in modeling nonlinear interactions among heterogeneous risk factors, this study proposes a multi-factor trigger probability prediction framework based on an improved Extreme Gradient Boosting (XGBoost) algorithm. A comprehensive “human– machine–environment–management” factor system is established to characterize complex coupling mechanisms underlying urban safety events, and adaptive sample weighting, feature interaction enhancement, and hierarchical regularization strategies are incorporated to improve prediction performance under imbalanced data conditions. Experimental evaluation using six years of real-world incident data demonstrates that the proposed model achieves an accuracy of 92.3%, a precision of 89.7%, a recall of 91.2%, an F1-score of 90.4%, and an AUC of 0.935, outperforming conventional logistic regression, support vector machine, random forest, and traditional gradient boosting models. Feature importance analysis further identifies extreme weather frequency, equipment aging rate, and personnel violation frequency as the dominant contributors to incident occurrence. Beyond urban emergency management, the proposed framework provides a reliable data-driven methodology for risk assessment and adaptive decision support in electromagnetic sensing networks, wireless monitoring systems, and intelligent communication infrastructures requiring real-time multi-source information fusion.