Investigation of Freeway Incident Duration Using Classification and Regression Trees Based on Multisource Data.

Area of Science:

• Traffic Engineering
• Data Mining
• Machine Learning

Background:

• Incidents on freeways cause significant harm and delays.
• Targeted contingency measures are effective in mitigating incident impact.
• Predicting incident duration is crucial for effective emergency response.

Purpose of the Study:

• To predict and quantify incident duration using the Classification and Regression Trees (CART) data mining technique.
• To compare the predictive accuracy and interpretability of CART with other machine learning models.
• To develop data-driven rules for effective incident assessment and emergency planning.

Main Methods:

• Utilized multisensor data from the Hangzhou freeway (2019-2021).
• Constructed an eight-level, 14-leaf node regression tree using CART.
• Extracted 14 rules from the CART model to inform contingency measures.
• Compared CART with XGBoost, Random Forest (RF), and Accelerated Failure Time (AFT) models.

Main Results:

• The CART model demonstrated superior prediction accuracy compared to XGBoost, RF, and AFT.
• The CART method provides strong interpretability, capturing interactions among up to seven variables.
• The derived rules facilitated accurate incident assessment and effective emergency plan implementation.

Conclusions:

• CART is a highly effective and interpretable method for predicting incident duration on freeways.
• The study's findings support the engineering application of machine learning in traffic management.
• Data-driven insights from CART improve incident response and reduce overall harm.