Crash severity analysis: A data-enhanced double layer stacking model using semantic understanding.

Area of Science:

• Traffic Safety Engineering
• Data Science
• Machine Learning

Background:

• Traffic crash severity analysis is vital for prevention and resource allocation.
• Existing models often underutilize semantic information in crash data.
• Small sample sizes and data imbalance hinder prediction accuracy.

Purpose of the Study:

• To propose a semantic understanding-based, data-enhanced model (EnLKtreeGBDT) for improved crash severity prediction.
• To leverage inherent semantic information for deeper understanding of crash factors.
• To address data imbalance and reduce reliance on large datasets.

Main Methods:

• Developed a semantic enhancement module for multi-dimensional feature extraction from crash data.
• Implemented a data enhancement module using denoising and migration techniques to combat data imbalance.
• Constructed a two-layer stacking model combining linear and nonlinear classifiers for complex relationship learning.

Main Results:

• The EnLKtreeGBDT model demonstrated superior prediction precision compared to state-of-the-art methods on UK road safety datasets.
• Ablation studies confirmed the critical contribution of both semantic and data enhancement modules.
• The model effectively predicts crash severity on complex urban roads.

Conclusions:

• The proposed EnLKtreeGBDT model significantly enhances traffic crash severity prediction accuracy.
• Semantic understanding and data enhancement are crucial for robust crash prediction models.
• This approach offers a promising solution for improving road safety and resource allocation.