Explainable Temporal Heterogeneous Graph Learning for Crash Risk Regime Discovery: A Data-Driven Engineering Framework

Understanding how road crashes group into recurring patterns is a prerequisite for targeted safety management, yet most crash-analysis methods treat each crash as an independent record, ignore the relational structure linking crashes through shared roadway, environmental, vehicle and temporal contexts, and depend on predefined labels such as injury severity. This study proposes a Temporal Attention Heterogeneous Graph Transformer (TA-HGT) framework for discovering latent crash risk regimes without predefined categories. A total of 13,729 crash records from 2013 to 2024, described by 57 attributes, were converted into a heterogeneous crash graph with 13,795 nodes (13,729 crash nodes and 66 contextual entity nodes) and 219,664 edges across 16 relation types. TA-HGT combines relation-aware projection, heterogeneous attention aggregation and a feature-wise temporal attention gate to produce 32-dimensional crash embeddings. Clustering identified four regimes containing 60.54%, 19.78%, 10.31% and 9.37% of crashes. TA-HGT achieved a silhouette score of 0.523, a Davies-Bouldin index of 0.548 and a Calinski-Harabasz index of 28,208, versus 0.345, 1.294 and 3,042 for a static heterogeneous graph transformer (HGT); a dynamic HGT that adds time information without temporal attention performed no better than the static model. A random forest recovered the regimes from the embeddings with 89.7% accuracy on a stratified hold-out set (macro F1-score = 0.84). Two regimes were temporally stable and two were volatile, moving in opposite directions with pronounced shifts around 2020. SHapley Additive exPlanations (SHAP) showed that a few latent dimensions dominate and that regimes rely on different dimensions. Adaptive temporal weighting, rather than the mere presence of time information, appears to drive well-organized crash representations. Limitations concerning external validation, semantic interpretation of latent dimensions and the choice of regime number are discussed.