An integrated data- and theory-driven crash severity model.

Area of Science:

• Transportation Science
• Traffic Safety Engineering
• Data Science

Background:

• Traditional crash severity models face a trade-off between predictive accuracy (data-driven) and interpretability (theory-driven).
• Existing methods like one-hot encoding in econometric models fail to capture semantic relationships between crash variables.
• Machine learning models offer high predictability but often lack transparency in explaining crash severity factors.

Purpose of the Study:

• To propose an integrated model, TVR-EF, that combines the strengths of data-driven and theory-driven approaches for crash severity modeling.
• To enhance the interpretability of crash severity factors by leveraging learned embeddings.
• To improve the flexibility and reduce prior knowledge dependency in crash severity outcome analysis.

Main Methods:

• Developed the Embedded Fusion model based on Text Vector Representations (TVR-EF).
• The data-driven component utilizes learned embedding weight matrices for enhanced interpretability.
• The theory-driven component implements a multinomial logit model as a 2D-Convolutional Neural Network (2D-CNN).

Main Results:

• The TVR-EF model demonstrated superior predictive performance compared to traditional econometric and machine learning models.
• The integrated approach significantly improved the interpretability of the relationships between crash characteristics and severity.
• Benchmarking was conducted using a crash dataset from Guangdong Province, China.

Conclusions:

• The TVR-EF model effectively bridges the gap between predictability and interpretability in crash severity modeling.
• This integrated approach offers a more comprehensive understanding of traffic crash dynamics.
• The findings suggest a promising direction for developing advanced traffic safety analysis tools.