DADiffNet: Delay-aware diffusion networks with adaptive subgraphs for large scale traffic forecasting.

Rapid urban expansion and growing traffic demand make accurate traffic forecasting increasingly important. However, methods that rely solely on absolute flow values often fail to explicitly capture upstream-triggered perturbations and their delayed diffusion downstream. To address this issue, we propose the Delay-Aware Diffusion Network (DADiffNet), which reformulates spatiotemporal coupling from two complementary perspectives: traffic-flow increments and adaptive subgraph structures. DADiffNet models high-frequency perturbations using temporal differences, enabling explicit learning of trigger-response relations and inter-node propagation delays. Meanwhile, an adaptive local subgraph sampling strategy encodes large-scale topology efficiently with near-linear computational complexity. Experiments on eight real-world datasets show that DADiffNet consistently outperforms fifteen strong baselines, achieving average improvements of 8.04%, 7.65%, and 8.40% in MAE, RMSE, and MAPE, respectively. In addition, DADiffNet reduces memory consumption, offering an improved trade-off among accuracy, efficiency, and interpretability for large-scale traffic forecasting.