Crime Pattern Detection and Hotspot Prediction Using Data MiningTechniques for Forensic Applications

Abstract

Rising urbanization and increasingly sophisticated criminal activity present mounting challenges for law enforcement agencies worldwide. Traditional crime analysis approaches, reliant on manual examination of historical records, are ill-equipped to handle the volume, velocity, and variety of modern crime data. This paper proposes and evaluates a comprehensive data mining framework—integrating K-Means and DBSCAN clustering with Decision Tree, Random Forest, and Naive Bayes classification—to detect spatial, temporal, and behavioral crime patterns from historical open-source datasets (Chicago, San Francisco, Boston). The framework generates predictive hotspot maps and investigative intelligence that support proactive policing and forensic decision-making. Experimental results on the Chicago Crime Dataset (2001–2023, n = 7,941,282 records) demonstrate that the Random Forest classifier achieves 91.4% accuracy, 89.7% precision, 88.3% recall, and an F1-score of 0.890, outperforming the Kernel Density Estimation (KDE) baseline by 14.6 percentage points on the Prediction Accuracy Index (PAI). DBSCAN clustering reveals 47 statistically significant crime hotspots, while temporal analysis identifies strong weekly and seasonal periodicity. Beyond operational policing, the framework provides forensic analysts with case-linkage intelligence, evidence-prioritization support, and modus-operandi profiling. The study rigorously addresses ethical considerations—data anonymization, bias mitigation, and transparent reporting—ensuring outputs serve as investigative aids rather than definitive legal conclusions. These findings advance computational criminology and contribute a deployable, explainable forensic intelligence tool for evidence-based law enforcement.