Existing traffic conflict models face challenges in handling minority class samples and capturing dynamic interactions in complex traffic scenarios. These limitations hinder model generalization and real-world applicability. This study employs an enhanced Two-dimensional Time-to-collision (2D-TTC) metric combined with vehicle interaction relationships to predict traffic conflicts of multiple patterns. To address imbalance in conflict and non-conflict events, both undersampling and oversampling techniques are employed, while a generative adversarial network with self-attention layers is leveraged to overcome the shortcomings of oversampling methods. Indeed, this approach proved highly effective, elevating the model's F1-score from 76.35 % with undersampling alone to 94.21 %. Additionally, several machine learning and deep learning models are compared, with the hypergraph attention network combined with Shapley additive explanations (S-HGAT) demonstrating the strongest learning capability. Furthermore, vehicle speed is identified as the most influential factor associated with traffic conflicts. A comprehensive re-evaluation of feature combinations reveals that the top six features-vehicle speed, the number of vehicles ahead, the standard deviation and the average of vehicle speeds within the traffic flow, distance with the road markings, and peak traffic hour indicators-result in the highest model F1-score of 98.41 % and accuracy of 97.66 %. Finally, the real-world implications of these findings are discussed.
Guan et al. (Mon,) studied this question.