Predicting High-Risk Conflicts on Highways Using Hybrid Deep Learning Resampling and Explainable Artificial Intelligence

Abstract Accurately predicting high-risk conflicts on highways is essential yet challenging due to complex traffic maneuvering under heterogeneous conditions. This challenge is further complicated in real-world data by the prevalence of non-high-risk conflicts relative to high-risk conflicts, resulting in an extremely imbalanced dataset. To address this challenge, the present study developed three machine-learning algorithms with six resampling techniques, including a novel deep learning framework that integrates conditional tabular generative adversarial network (CTGAN)–based synthetic data generation with strategic undersampling. High-risk conflicts were identified using the surrogate safety measure modified time to collision ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="upper M upper T upper T upper C less than 1 normal s" display="inline" overflow="scroll"> <mml:mi>MTTC</mml:mi> <mml:mspace width="-0.2em"/> <mml:mo>&lt;</mml:mo> <mml:mspace width="-0.2em"/> <mml:mn>1</mml:mn> <mml:mtext> </mml:mtext> <mml:mi mathvariant="normal">s</mml:mi> </mml:math> ). Spatiotemporal traffic variables (i.e., traffic density, traffic flow, standard deviation in speed and acceleration, lane change frequency, etc.) were extracted for the 5-s window prior to each conflict within a two-dimensional influence zone using high-resolution trajectory data. Results show that the random forest model outperformed others when trained on CTGAN-synthesized data combined with random undersampling (RU). This hybrid approach improved sensitivity by 30% and increased the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="upper G" display="inline" overflow="scroll"> <mml:mi>G</mml:mi> </mml:math> -mean by 11% compared to the original imbalanced dataset, validating the proposed data-balancing approach. Sensitivity analysis indicated that CTGAN-RU performs optimally when the RU ratio is 2:1 (non-high-risk: high-risk) and the CTGAN-RU ratio becomes 1:1. Further, Shapley additive explanations identified traffic density and speed variability as the two most influential variables increasing the likelihood of high-risk conflicts, while higher flow reduced risk. By addressing both data imbalance and critical preconflict dynamics, the proposed framework significantly enhances high-risk conflict prediction and supports more effective driver alerts in heterogeneous traffic conditions.