Interpretable multi-variable transformer network for regional-level short-term bicycle crash risk prediction.

Area of Science:

• Urban planning and traffic safety
• Artificial intelligence and machine learning
• Transportation engineering

Background:

• Short-term bicycle crash prediction is vital for urban safety interventions.
• Existing methods struggle with complex data, class imbalance, and interpretability.
• The role of bicycle infrastructure in crash risk is underexplored.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate short-term bicycle crash prediction.
• To address challenges in spatiotemporal dependency modeling and class imbalance.
• To identify critical bicycle infrastructure features influencing regional crash risk.

Main Methods:

• Proposed an Interpretable Multi-variable Transformer Network (IMTN) using Transformer encoder blocks.
• Implemented a novel approach for class imbalance by predicting risk per region.
• Utilized an improved Layer-wise Relevance Propagation (LRP) for model interpretability.
• Incorporated diverse data sources including crash records, traffic, weather, and 48 infrastructure features.

Main Results:

• IMTN outperformed baseline models, reducing false positive rate (FPR) by up to 9.08% and improving AUC by 3.49%.
• The model achieved peak performance at a 1-hour temporal resolution, suggesting effectiveness in high-resolution settings.
• Interpretability analysis highlighted segregated cycle lanes, Sheffield stands, and colored path markings as significant infrastructure variables.

Conclusions:

• IMTN offers an effective and interpretable solution for short-term bicycle crash prediction.
• The proposed class imbalance handling method enhances performance at fine temporal resolutions.
• Identified infrastructure features provide actionable insights for improving urban cycling safety.