CBI-GNN: Copula–Birnbaum Importance Gated Spatiotemporal Neural Network for Crime Hotspot Forecasting

This study introduces CBI-GNN, a Copula–Birnbaum Importance Gated Spatiotemporal Neural Network for crime hotspot forecasting. The model integrates heterogeneous city data—crime, point of interests, public housing, and weather—to predict weekly robbery risk in New York City. A CBI Gate captures nonlinear inter-feature dependence via a Gaussian copula and applies reliability-based FiLM scaling to dynamically weight features. Spatial and temporal dependencies are captured by GraphSAGE and GRU, respectively. Trained with a dual loss (cross-entropy and MSE), the proposed model achieves PR-AUC = 0.522, ROC-AUC = 0.928, MAE = 0.146, RMSE = 0.358, demonstrating that copula-informed reliability gating enhances interpretability and stabilizes spatiotemporal crime prediction.