Urban theft prediction via LLM-empowered spatiotemporal transformer

Area of Science:

• Urban Criminology
• Artificial Intelligence
• Geospatial Analysis

Background:

• Urbanization increases spatiotemporal crime heterogeneity, necessitating advanced prediction models.
• Existing models struggle with complex data integration and dynamic urban environments.
• Accurate crime prediction is crucial for effective urban planning and public safety.

Purpose of the Study:

• To develop an LLM-enhanced Spatiotemporal Transformer (LLM-STT) model for accurate, neighborhood-scale theft prediction in New York City.
• To integrate multi-source spatiotemporal data, including taxi passenger flow and LLM embeddings (Gemma3-12B), into a unified prediction framework.
• To evaluate the model's performance, focusing on semantic encoding, feature coupling, and deployment feasibility.

Main Methods:

• Proposed an LLM-STT model incorporating Gemma3-12B embeddings and a lightweight fine-tuning approach for Gemma3-1B.
• Integrated multi-source spatiotemporal features, including dynamic population proxies.
• Quantified feature coupling and assessed the balance between model performance and deployment feasibility.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.91 and an F1 score of 0.83 for hourly theft prediction.
• Demonstrated competitive performance against baseline models, highlighting the contribution of LLM embeddings and dynamic population features.
• The lightweight fine-tuned model significantly outperformed the random baseline.

Conclusions:

• The LLM-STT model shows significant promise for enhancing the accuracy of urban crime prediction, particularly for theft.
• LLM embeddings and dynamic population data are valuable additions for spatiotemporal crime analysis.
• Findings support the development of targeted crime prevention strategies in urban settings, though further validation is needed for broader applicability.