CausalST: directed delay-aware graph neural networks for traffic flow prediction

Abstract

Much of the recent work on traffic forecasting relies on two implicit simplifications: representing the road network as an undirected graph and adopting temporally symmetric encoders within the historical observation window. These choices may obscure asymmetric predictive relations among sensors and the chronological ordering of intermediate temporal representations. CausalST addresses these limitations jointly. For spatial modeling, separate source and destination embeddings are assigned to each sensor to construct a learned directed adjacency under road-topology constraints. On the temporal side, dilated causal convolutions serve as a temporal inductive bias, ensuring that the representation at each historical position depends only on current and earlier observations. Delay-aware graph propagation aggregates current and lagged neighbor states over predefined discrete lags. A parallel spatial attention branch captures nonlocal dependencies beyond the physical topology, while a gated fusion unit adaptively integrates the temporal, graph-propagation, and spatial-attention representations. Experiments on PeMS03, PeMS04, and PeMS08 show that CausalST achieves the lowest MAPE among all compared methods while remaining competitive in MAE and RMSE. Ablation results reveal non-additive interactions between directed weighting and delayed aggregation, with the full configuration attaining the lowest MAE on both evaluated datasets.