Abstract

Conventional urban safety mechanisms like CCTV and manual alerts are fundamentally reactive. To enable proactive hazard mitigation, this research proposes a predictive spatio-temporal risk analytics framework paired with an Explainable AI (XAI) engine and an interactive decision-support dashboard. The system establishes a dynamic Compound Risk Scoring Index (RSI) combining temporal cycles, spatial isolation, illuminance grades, crowd density, and historical incident rates. An XGBoost classifier is trained against classical baselines (Logistic Regression, Decision Tree, and Random Forest) to categorize urban zones into granular risk tiers, while Tree-SHAP provides real-time local feature attributions to resolve the black-box dilemma. The architecture is deployed via an asynchronous REST API backend and Folium-based geospatial heatmaps. Evaluated on a procedural synthetic benchmark to demonstrate end-to-end feasibility, the proposed XGBoost model achieved 93.80% accuracy on a stratified split and 94.60% on a chronological split, with a Macro F1-Score of 0.9379-0.9456 and sub-millisecond inference latency, establishing a scalable, transparent foundation for proactive urban safety management.

Authors

Vinit Kumar Shukla, Priyanka Kumari, Sofi Mittra, B.M. Reddama, S.T. Sadiya
Nagarjuna College of Engineering and Technology, India

Keywords

Predictive Risk Analytics, Explainable Artificial Intelligence (XAI), Tree-SHAP, Gradient Boosting, Spatio-Temporal Modeling, Smart Cities