Time varying Granger causality guided delay aligned propagation and spatiotemporal adaptive embedding for traffic speed prediction

Abstract

Traffic speed prediction is vital for intelligent transportation systems. To achieve accurate forecasting, spatio-temporal graph neural networks have been widely adopted; however, despite recent advances, many existing methods still aggregate spatial features synchronously and rely on static or weakly adaptive dependency structures. This mismatch limits long-horizon accuracy because spatial influence in traffic networks is inherently directional, delay-sensitive, and time-varying. To address these limitations, we propose a unified forecasting framework: the Time-Varying Granger Causality-Enhanced Spatio-Temporal Adaptive Embedding Transformer (TVGC-STAEformer). Our framework utilizes a strictly causal Recursive Least Squares (RLS) procedure to estimate dynamic predictive influence and optimal propagation lags. These estimates guide a Delay-Aligned Dynamic Causal Message Passing (DADCMP) module to aggregate historical features at physically appropriate temporal offsets. Additionally, a Spatio-Temporal Adaptive Embedding (STAE) module explicitly encodes node-specific and time-varying heterogeneity. By precomputing these causal tensors, the model efficiently captures realistic directional dependencies directly from evolving traffic sequences without requiring strong interventional causation. Experiments on Daegu and METR-LA, supported by five-seed and congestion-transition analyses on Daegu, demonstrate the effectiveness and stability of TVGC-STAEformer, while a supplementary PeMSD7-M evaluation probes whether the observed long-horizon tendency is retained on a larger detector network. Specifically, compared to the strongest baseline (DCRNN) at the 60-min prediction horizon, our model reduces the MAE by 1.93% and 1.56% on the Daegu and METR-LA datasets, respectively. These results suggest that explicitly combining delay alignment, time-varying predictive influence, and node-time heterogeneity is a useful direction for traffic speed forecasting. These findings can support traffic agencies in selecting forecasting tools for proactive signal control, route guidance, and congestion management, particularly when reliable long-horizon speed estimates are required.