Answers are grounded in this paper and cite their sources.
Achieving sustainable urban mobility and climate resilience (SDG 11 and SDG 13) requires not only mitigating accident impacts but proactively preventing them, as traffic collisions induce significant secondary congestion, fuel waste, and emissions. However, existing data-driven risk assessment models predominantly rely on correlational learning, which limits their robustness under distribution shifts and their capacity to support actionable, policy-oriented interventions. To bridge this gap, this study proposes CausalBGT, a neural–symbolic framework that integrates causal inference, spatio-temporal graph learning, and large language models for proactive and interpretable urban traffic risk management. The framework uses quantitative parameters—traffic flow, speed, occupancy, historical accident frequency/severity, temperature, visibility, and spatio-temporal coordinates—as node and graph features for causal discovery and the causally guided spatio-temporal graph transformer, and qualitative parameters—weather, road-surface, and lighting categories plus unstructured incident reports, weather warnings, construction/signal-failure descriptions, and traffic-related social-media text—through a domain-adaptive BERT-based semantic encoder, whose outputs are aligned with sensor data via cross-modal attention and gated fusion to extract causal risk factors. The framework consists of three synergistic components: (1) a causally guided spatio-temporal graph transformer that embeds discovered causal priors into graph attention mechanisms to enhance generalization and structural robustness; (2) a domain-adaptive BERT-based semantic encoder that extracts causal risk factors from unstructured traffic text and aligns them with structured sensor data; and (3) a causally augmented large language model interpreter that translates predictive outcomes into structured, intervention-oriented recommendations grounded in causal pathways. Extensive experiments on multi-source datasets from the United Kingdom, New York City, and California demonstrate that CausalBGT achieves an AUC-ROC of 0.923, outperforming TrafficGPT by 2.2% and ST-Transformer by 3.8%, while reducing average treatment effect error by 55.6% and 68.4%, respectively. A real-world case study in Midtown Manhattan further validates a 28-min early warning lead time. By enabling proactive, evidence-based interventions, the framework provides a model-based projection for mitigating accident-induced congestion and emissions, thereby contributing to sustainable urban mobility and enhanced resilience of city transportation systems.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: