An explainable AI framework for spatial hotspot detection and temporal forecasting of urban crime

Abstract Urban crime analysis is most useful when it answers two practical questions: where crime is likely to concentrate, and how crime volume may change over time. This paper presents a unified and reproducible framework for studying both problems across three major U.S. cities: Chicago, Los Angeles, and New York City. For hotspot detection, eight machine learning classifiers and a stacked ensemble are tested on approximately 1 km 2 spatial grids under a strict chronological split. XGBoost gives the best ROC-AUC in all three cities. SHAP (SHapley Additive exPlanations) analysis is applied to the best-performing XGBoost classifier in each city to identify and visualise the most influential predictors of hotspot probability. For daily crime forecasting, four deep learning models - Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Transformer are compared with an AutoRegressive Integrated Moving Average (ARIMA) baseline. The deep learning models consistently outperform ARIMA on daily forecasting, with LSTM performing best in Chicago and GRU performing best in Los Angeles and New York City. The study also includes category-level forecasting for the five most common crime types in each city and an ablation study of hybrid deep learning architectures.