Hybrid Spatio-Temporal Crime Forecasting with Graph Neural Networks

Abstract

Traditional crime forecasting models often struggle with the complex, interconnected dynamics of urban crime, primarily because they analyze spatial and temporal factors in isolation. This paper introduces a novel hybrid deep learning framework designed to overcome these limitations by modeling crime as a dynamic spatio-temporal neural network. Our approach integrates a SpatioTemporal Graph Neural Network (ST-GNN) to capture complex inter-location dependencies and a robust Enhanced Long Short-Term Memory (eLSTM) to model intricate temporal patterns. This synergistic fusion enables a more holistic understanding of crime propagation and frequency. We evaluate this model on a five-year urban crime dataset, where it demonstrates superior performance over standalone forecasting methods, achieving a 10–15% reduction in Mean Absolute Error (MAE) and a hotspot precision of 0.88. The results of our ablation studies underscore the importance of combining these complementary learning paradigms. The proposed framework provides a highly accurate and scalable tool for proactive policing and resource allocation.