AI-BASED CRIME HOTSPOT PREDICTION SYSTEM USING MACHINE LEARNING

Abstract

Rising urban crime and the limitations of manual patrol scheduling have created a pressing need for data-driven policing tools. This paper presents an AI-based crime hotspot prediction system that applies supervised machine learning to historical crime records in order to forecast the localities most likely to witness criminal activity. Three ensemble learning algorithms — Random Forest, Gradient Boosting, and XGBoost — were trained on a structured crime dataset comprising incident type, location coordinates, time of occurrence, and sector-level metadata. The processed features were used to build a prediction engine that outputs a risk score for a queried location, which is then visualised through an interactive command dashboard, a hotspot tracking module, and a Google Maps-based geo-risk layer. Experimental evaluation shows that XGBoost achieved the highest overall accuracy of 73.1% along with a recall of 75.0%, narrowly outperforming Gradient Boosting and Random Forest, both of which stayed close behind on accuracy but showed competitive ROC-AUC scores of 76.4% and 74.9% respectively. The system demonstrates that even a modestly sized dataset, when combined with careful feature engineering and ensemble modelling, can support meaningful proactive policing decisions. The paper also discusses the system architecture, workflow, mathematical foundations, and the dashboard implementation, and closes with a discussion of limitations and directions for future enhancement using deep learning and real-time sensor integration.