LLM-enhanced causal graph learning for real-time crash risk prediction

Area of Science:

• Road safety
• Artificial Intelligence
• Causal Inference

Background:

• Accurate traffic crash risk prediction is vital for road safety.
• Existing methods often lack a balance between predictive performance and causal interpretability.
• There is a need for advanced frameworks to enhance collision risk prediction.

Purpose of the Study:

• To propose a closed-loop framework integrating causal discovery, semantic enhancement, and spatio-temporal prediction for traffic risk assessment.
• To improve the balance between predictive accuracy and causal interpretability in collision risk models.
• To enable real-time risk prediction for enhanced traffic management.

Main Methods:

• Utilized transfer entropy for causal discovery from dangerous driving scenarios to build causal graphs.
• Employed a GPT-2 language model fine-tuned with Low-Rank Adaptation (LoRA) for semantic graph enhancement.
• Integrated Graph Attention Networks (GAT) and Long Short-Term Memory (LSTM) networks for spatio-temporal prediction.

Main Results:

• Achieved high performance metrics: 0.956 accuracy, 0.872 F1-score, and 0.985 AUC on the highD dataset.
• Demonstrated significant outperformance compared to traditional baseline methods.
• Ablation studies confirmed the crucial role of GPT-2-based causal enhancement with LoRA.

Conclusions:

• The proposed framework offers a powerful solution for precise and interpretable collision risk prediction.
• The integration of causal discovery and advanced language models significantly boosts predictive accuracy.
• The framework shows strong potential for real-time traffic management systems and improving road safety.