Structure-Guided Graph Attention Learning for Traffic Conflict Prediction under Heterogeneous Expressway Geometries

Abstract

Accurate traffic conflict prediction on urban expressways remains challenging because similar observed traffic states may carry different risk implications across merging, weaving, and diverging segments. Existing data-driven models often learn shared correlations or impose fixed structures under geometric heterogeneity. To address this issue, this study develops Causal-GAT, a structure-guided graph attention framework for heterogeneous traffic conflict prediction. Within the framework, SAM-inferred structures are incorporated as segment-specific structural priors to organize observed kinematic relationships and guide graph attention learning. Structural Agnostic Modeling (SAM) is first applied to observed traffic variables to infer segment-specific proxy structural topologies. These priors are then injected into both feature reconstruction and graph attention so that message passing is guided by segment-specific structure in addition to data similarity. Experiments on the MAGIC UAV trajectory dataset from Shanghai’s Inner Ring Elevated Road cover three representative segment types within one expressway corridor and a four-level traffic conflict classification task defined by Time Spent in Conflict (TSC). Across these segments, Causal-GAT achieves the best Macro F1 among the baselines, with the largest relative improvement appearing in the off-ramp segment, where geometric and traffic conditions are more stochastic. The inferred proxy structural topologies and attribution results are also broadly consistent with traffic-flow differences among merging, weaving, and diverging environments, while indicating that proxy structural importance and model-based predictive contribution are related but not identical. Overall, the findings suggest that structure-guided attention helps mitigate spurious associations in heterogeneous geometric conditions and provide a more interpretable basis for segment-sensitive traffic safety analysis.