Towards Resilient Transportation: A Conditional Transformer for Accident-Informed Traffic Forecasting

TL;DR

ConFormer introduces a conditional Transformer for accident-informed traffic forecasting, integrating graph propagation with Guided Layer Normalization to dynamically adapt spatial-temporal dependencies under incident conditions. The approach is trained and evaluated on two large incident-rich datasets (Tokyo and California), outperforming STAEFormer and other baselines while maintaining scalability and efficiency. Key contributions include the Spatiotemporal Condition Propagation, GLN-guided conditioning, and extensive ablations showing the value of incident data and dynamic normalization. The results demonstrate the method’s practical potential for resilient urban traffic management, including large-scale deployment scenarios.

Abstract

Traffic prediction remains a key challenge in spatio-temporal data mining, despite progress in deep learning. Accurate forecasting is hindered by the complex influence of external factors such as traffic accidents and regulations, often overlooked by existing models due to limited data integration. To address these limitations, we present two enriched traffic datasets from Tokyo and California, incorporating traffic accident and regulation data. Leveraging these datasets, we propose ConFormer (Conditional Transformer), a novel framework that integrates graph propagation with guided normalization layer. This design dynamically adjusts spatial and temporal node relationships based on historical patterns, enhancing predictive accuracy. Our model surpasses the state-of-the-art STAEFormer in both predictive performance and efficiency, achieving lower computational costs and reduced parameter demands. Extensive evaluations demonstrate that ConFormer consistently outperforms mainstream spatio-temporal baselines across multiple metrics, underscoring its potential to advance traffic prediction research.

Figures (7)

• Figure 1: Illustration of our motivation: enhancing traffic forecasting towards resilient transportation.
• Figure 2: ConFormer enhances the spatiotemporal transformer with Graph Propagation and Guided LayerNorm (GLN). The graph propagation, combined with an MLP, generates the mean ($\beta$), variance ($\gamma$), and amplitude factor ($\alpha$) to guide layer normalization and residual connections, regulating the latent space distribution at each node.
• Figure 3: Comparison of model efficiency and effectiveness across spatiotemporal models on the Bay Area dataset.
• Figure 4: Case studies of traffic prediction in a subarea of Tokyo. Left: Analysis of an accident scenario, where ConFormer effectively captures the evolution of node distributions in latent space before and after the accident. Right: Analysis of a traffic regulation scenario, where graph propagation demonstrates learned differential attention scores in response to the implemented traffic regulations.
• Figure 5: Distribution of predictions and ground-truths under normal and accident scenarios on Tokyo dataset.