Urban theft prediction via LLM-empowered spatiotemporal transformer

Abstract With the deepening of urbanization, the spatiotemporal heterogeneity of theft crimes in New York City has become prominent, creating a demand for more accurate prediction. Existing models face limitations in capturing nonlinear correlations, integrating multi-source data, and generalizing to dynamic scenarios. This study proposes an LLM-enhanced Spatiotemporal Transformer (LLM-STT) model, which integrates multi-source spatiotemporal features (including taxi passenger flow proxy) and Gemma3-12B embeddings, with a lightweight fine-tuning scheme for Gemma3-1B. Its main explorations include LLM-based semantic encoding, quantifying feature coupling, and balancing performance and deployment feasibility. Experiments on hourly neighborhood-scale theft prediction in New York City show the model achieves an AUC of 0.91 and an F1 score of 0.83, demonstrating competitive performance against baselines. LLM embeddings and dynamic population features contribute positively, and the lightweight fine-tuned model outperforms the random baseline. These findings offer preliminary support for targeted crime prevention in similar urban contexts, with broader generalization requiring further validation.