Discovering state-conditioned multi-scale spatial dependence in urban road networks: a hop-aware graph attention approach

Multi-scale spatial dependence, in which local perturbations and regional trends jointly shape traffic states, is central to understanding urban road networks. Most graph neural networks compress neighbourhoods into a single adjacency scale, making dependence across scales and traffic states difficult to distinguish. We propose the Multi-scale Adaptive Weighting Spatio-Temporal Graph Attention Network (MSAW-STGAT). Within non-overlapping 1–4-hop neighbourhoods defined by shortest-path distance, intra-scale attention (α) characterises neighbour importance, while state-conditioned scale weights (β) characterise the relative roles of different scales. Together, they represent spatial dependence through neighbour differences and scale allocation. Experiments on public traffic datasets and independent data-side regression provide consistent evidence that the model identifies predictive dependence concentrating on directly adjacent nodes under congestion. Deletion tests among these nodes further support the ability of intra-scale attention to identify neighbours that matter more for prediction. Comparisons and ablations across datasets show RMSE comparable to representative models, including STAEformer and AdapGL, with the non-overlapping neighbourhood partition and both types of weights contributing to predictive accuracy. While maintaining predictive accuracy, the method enables neighbour differences, scale allocation and their variation with traffic state to be identified and tested in urban road-network spatial dependence.