Hierarchical Classification-Regression (HiClassR) to Improve Incident Clearance Time Prediction

Incident clearance time prediction is a key task in Advanced Transportation Management and Advanced Traveler Information Systems, as it informs mitigation and response strategies adopted by transportation agencies. Incident clearance time prediction is not a trivial task given that the duration of an incident is influenced by various factors, some of which are difficult to measure or not measured at all. Such constraints in the data limit the accuracy of the models developed for predicting clearance time. Over the past decades, researchers have employed various statistical methods and machine learning approaches to overcome these limitations and improve the incident clearance time prediction model's accuracy. While these efforts have been successful in improving the overall accuracy of models, there are very limited studies that have focused on assessing and improving the generalization ability of incident clearance time prediction models.

This paper proposes the hierarchical classification-regression (HiClassR) framework to improve the predictive performance of common machine-learning algorithms in predicting the incident clearance time. The HiClassR is a bi-level framework in which the incidents are first assigned to an incident class; then, a fine-grained class-specific model provides an estimate of the incident clearance time. While the HiClassR method is expected to improve error measures such as RMSE over the entire dataset, it is also expected to improve the generalization ability of the model across various incident scenarios.

In a case study, the HiClassR method was applied to six conventional predictive methods to develop incident clearance time prediction models for St. Louis, MO, USA. The case study results indicate that the HiClassR model consistently outperforms the conventional base models and effectively generalizes the overall prediction power of the model to various incident categories of different conventional predictive methods. On average, the preferred HiClassR model reduced the root mean square error (RMSE) from 22.7 min to 17.6 min for the random forest model, from 22.5 min to 19 min for Bayesian Regularization for Feed-Forward Neural Networks, from 23.7 min to 16.9 min for Support Vector Machines, from 25.2 min to 21.1 min for K-Nearest Neighbors, from 24.5 min to 19.7 min for XGBoost, and from 23.9 min to 18.3 min for Neural Networks.