Time-Aware Crime Prediction: A Multi-Output AI Framework for Operational Risk Prioritization
IEEE Access, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3701892
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Anahtar Kelimeler: Crime prediction, ensemble learning, harm-based analysis, machine learning, risk ranking, spatio-temporal modeling, stacking, top-k evaluation
- Gazi Üniversitesi Adresli: Evet
Özet
This study proposes an artificial intelligence (AI)-supported operational crime prediction framework that integrates temporal, spatial, and harm-aware risk modeling for short-term operational prioritization. Unlike conventional binary crime prediction approaches, the proposed multi-output framework simultaneously estimates crime occurrence probability, expected crime volume, crime harm, and crime category. Experimental results demonstrate that the joint use of temporal and spatial features substantially improves predictive performance, increasing Precision-Recall Area Under the Curve (PR-AUC) from 0.1849 to 0.2346 despite severe class imbalance (positive-class prevalence = 0.0715), while Receiver Operating Characteristic Area Under the Curve (ROC-AUC) improved from 0.7236 to 0.7496. The stacking approach produced more balanced and generalizable performance compared to individual models. The framework also exhibited strong probabilistic calibration, achieving an Expected Calibration Error (ECE) of 0.0012 and a Brier score of approximately 0.06. The multi-output modules further achieved RMSE≈0.445 for expected crime count estimation and RMSE≈4.02 for expected harm prediction. Under realistic operational evaluation settings, the model achieved precision@500 of 0.7460, precision@5000 of 0.5418, and lift@5000 of 7.57, while maintaining stable performance under strict out-of-fold and rolling as-of backtesting conditions (PR-AUC≈0.260 and PR-AUC≈0.251, respectively). The findings indicate that the proposed framework provides both improved predictive capability and practical utility for operational risk prioritization and decision-support processes in short-term crime forecasting.