Graph Attention Networks for Modeling Multi-Agent Decision Dynamics at Unsignalized Intersections

Abstract

Gap acceptance at unsignalized intersections is a safety-critical decision in which a driver on a stop-controlled approach must judge whether the available time headway in a conflicting traffic stream is sufficient to cross safely. Conventional models treat this as a single-vehicle problem, estimating a critical gap threshold from ego-vehicle kinematics alone---an approach that ignores the multi-agent reality of intersection negotiation. This paper reframes gap acceptance as a multi-agent interaction problem and applies a Spatiotemporal Graph Attention Network (ST-GAT) to model it. Using the INTERACTION Dataset v1.2 across four unsignalized intersection scenarios, we extract 24,627 labeled gap acceptance events and construct per-event scene graphs where nodes represent agents encoded with ego-relative spatiotemporal features and edges connect agents within a 10-meter spatial threshold. ST-GAT achieves AUC values between 0.793 and 0.911 across scenarios, outperforming GCN in all four and remaining competitive with logistic regression. Critically, attention weight analysis reveals three behavioral patterns consistent across scenarios: proximity drives attention allocation, non-motorized road users are disproportionately attended during gap rejections relative to acceptances, and accepted gaps are associated with faster-moving attended agents. These findings generalize across two intersection geometries and three class balance conditions, demonstrating that graph attention mechanisms can surface interpretable behavioral structure in human gap acceptance decisions that single-agent and non-attentive graph models cannot.