Spatiotemporal subgraph learning for abnormal urban congestion diffusion prediction

Abnormal traffic congestion diffusion is assumed to be the main reason for large-scale congestion, which severely impairs urban traffic efficiency. However, predicting such events remains challenging due to complexity nature of urban traffic flow, where congested flows exhibit diffusion characteristics distinct from those of well-studied free flow. Existing studies focus on unified traffic flow dynamics for road congestion event prediction, little attention has been paid to the spatiotemporal subgraph diffusion pattern of congested flows. In this paper, we propose a traffic spatiotemporal subgraph neural network (TST-SNN) to extract multiscale congestion diffusion features for abnormal congestion diffusion prediction. At the node level, the congested flow diffusion features are modeled as position-aware spatiotemporal dependencies. The local congestion structure changes are captured as topology-aware spatiotemporal dependencies. At the subgraph level, the congestion subgraph evolution features are extracted as temporal dependencies. Finally, abnormal congestion subgraph evolution prediction is conducted using its evolution features. Experimental results on morning and evening peak in Beijing and Shanghai demonstrate that the proposed TST-SNN outperforms current state-of-the-art methods by 9.78% on average in terms of F1 score. Ablation test shows that the congestion propagation perception process can improve the performance of TST-SNN by 31.6% on average. Our findings highlight the importance of congestion subgraph identification for abnormal congestion diffusion prediction, which may help the intelligent management of traffic health.