1374084 - High-Resolution Crash Risk Modeling at Intersections Enabled by Tree Ensembles with Contrasting Sampling

Crash prediction modeling has been generally conducted at aggregated spatiotemporal levels, largely due to the scarcity of crash data (i.e., crashes being rare events) and limitation of classical statistical methods in handling large datasets with high-dimensional features. This inevitably reduces the resolution and sensitivity of the models, leading to undesirable biases in risk inference. As a result, these classic statistical models have been commonly used for safety analysis at the planning or preliminary design levels and are not suitable for real-time location-specific crash prediction or risk inference. This paper addresses this modeling challenge by introducing a contrastive sampling technique, which enables the exploration of the existing statewide large and feature-rich datasets for high-resolution crash risk inference. Specifically, crash risk modeling is cast as a binary classification problem based upon the fine-grained spatiotemporal scale (spatial resolution: all intersections in a state; temporal resolution: 15-minute time intervals). As a case study, data from the state of Texas is utilized. The state-of-the-art tree ensemble methods are employed to deal with tabular data with high-dimensional features. Three popular ensemble models are developed and compared, including Random Forest, CatBoost, and Neural Oblivious Decision Ensembles (NODE). Random Forest is a representative parallel ensemble method while CatBoost is the state-of-the-art sequential ensemble method. In contrast, NODE is a novel deep learning architecture, which generalizes ensembles of oblivious decision trees and enables hierarchical representation learning with end-to-end gradient-based optimization. The performance evaluations on the test dataset indicate the CatBoost as the best-performed model for crash prediction, which achieved the accuracy of 96.23%, F1 score of 96.31% and Area Under the Curve (AUC) of 96.22%.