A novel framework for crash frequency prediction: Geographic support vector regression based on agent-based activity...

Area of Science:

• Spatial analysis
• Traffic safety
• Machine learning applications in transportation

Background:

• Traffic crash data exhibits spatial dependence and heterogeneity, impacting predictive accuracy.
• Understanding the influence of diverse factors on crash incidence is crucial for road safety improvement.

Purpose of the Study:

• Introduce the Geographical Support Vector Regression (GSVR) framework to analyze spatial variations in traffic crashes.
• Evaluate the impact of traffic, infrastructure, socio-demographic, travel demand, and land use on total and fatal-or-serious injury (FSI) crashes.
• Enhance machine learning models for traffic safety analysis with improved feature selection and spatial impact incorporation.

Main Methods:

• Utilized the Melbourne Activity-Based Model (MABM) dataset, analyzing 50 indicators for peak hour traffic and commuting modes.
• Developed a GSVR framework incorporating generated distance matrices to assess spatial effects.
• Applied a feature selection technique to enhance machine learning model capabilities.

Main Results:

• Active transportation (walking, cycling) is linked to higher crash risks, indicating vulnerability.
• Car commuting shows a lower impact on crashes compared to active transport, suggesting an imbalance in road safety.
• Public transport is generally safer, but risks exist near stations; tram stops affect total crashes, while intersections impact FSI crashes.
• Land use mix plays contrasting roles in FSI versus total crashes.

Conclusions:

• The GSVR framework offers a novel method for integrating spatial impacts into machine learning for traffic safety.
• Findings highlight disparities in road safety, emphasizing the need for improved infrastructure and policies for vulnerable road users.
• The study provides a dynamic approach to distance matrix generation for tailored spatial analysis in crash prediction.