Deciphering terrorist attributions: a global terrorism data-driven comparison of machine learning models

Attributing terrorist incidents to responsible groups remains a critical challenge in counter-terrorism, complicated by the evolving complexity and decentralization of extremist operations. This study addresses the urgent need for automated, accurate, and scalable attribution mechanisms by proposing and rigorously evaluating two novel ensemble-based machine learning models: an adaptive weighted voting ensemble (WVE) and a hierarchical Stacking Classifier with optimized meta-learner selection. Leveraging structured incident data from the Global Terrorism Database, the research implements a comprehensive methodological framework encompassing domain-specific feature engineering, advanced class imbalance handling through SMOTE-Tomek, and systematic hyperparameter optimization. Robust nested cross-validation is employed throughout to ensure unbiased model selection and reliable performance estimation. The study conducts comparative analysis across ten established machine learning algorithms and the two proposed ensemble methods. Experimental results demonstrate that the Stacking Classifier achieves 73.70% accuracy (± 1.23) and an AUC-ROC of 0.892, while the WVE attains 73.38% accuracy (± 1.18) and an AUC-ROC of 0.889-representing a 15–20% improvement over conventional methods. Both models maintain robust precision-recall balance, with their superiority statistically validated (p < 0.01, Friedman-Nemenyi test) in multi-class attribution tasks; furthermore, integrated SHAP analysis (showing 94% correlation with traditional importance) enables transparent, instance-level interpretation essential for operational deployment in counter-terrorism contexts. The proposed framework advances the field by offering interpretability, operational scalability, and multi-dimensional attribution capabilities, equipping security agencies with effective tools for real-time threat identification and response.