Key points are not available for this paper at this time.
Probabilistic traffic flow forecasting is fundamentally concerned with characterizing the intrinsic uncertainty of future traffic states, thus delivering robust and reliable decision-support information for urban traffic management. However, conventional spatiotemporal models typically aggregate features from a rapidly increasing number of neighboring nodes without distinction. This indiscriminate aggregation mixes true propagation patterns with irrelevant background noise, leading to the curse of dimensionality and substantially weakening critical traffic signals. To address these limitations through a principled framework, this work introduces a causal-guided decoupled learning framework designed to selectively identify truly influential intersections without expanding the feature space. The framework adopts a decoupled learning architecture to estimate the parameters of a Gaussian Mixture Model, enabling accurate modeling of the nonstandard traffic flow distributions at individual intersections. Underpinned by a rigorous network disentanglement strategy, the model explicitly separates local feature extraction from global structure discovery. Specifically, a specialized temporal encoder models the intricate interactions among nearby intersections, producing robust sequential representations. In parallel, to capture global and long-range dependencies, a dynamic causal graph is constructed via PTE and a temporal PageRank algorithm. As a result, the framework identifies both adjacent and multi-hop influential intersections and seamlessly fuses this global long-range influence with underlying local effects to generate the final probabilistic forecasts. Experiments on real-world traffic datasets demonstrate that the proposed model consistently outperforms state-of-the-art baselines, with ablation results suggesting that the causal disentanglement mechanism effectively reduces spurious spatial correlations.
Zhang et al. (Mon,) studied this question.