.png){kind=icon}
.png){kind=icon}
Return
A Multi-Scale Spatial-Temporal Interactive Attention Network for Traffic Forecasting
.png){kind=icon}
PRE
AI
DOI：10.1109/TVT.2025.3586380.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Forecasting of traffic conditions can support the decision-making processes of travelers and traffic managers. However, the complex spatial and temporal correlations in traffic data make traffic forecasting challenging. Modeling such correlations in a point-to-point manner as in most state-of-the-art approaches ends up ignoring dependencies across multiple temporal and spatial scales. Moreover, despite several attention-based mechanisms being proposed for capturing complex spatial and temporal dependencies, the quadratic complexity of these mechanisms puts their scalability to large-scale traffic networks at stake. To address the above challenges, we propose a novel Multi-Scale Spatial-Temporal Attention Network, abbreviated as MSSTAN, which aggregates the traffic data into subseries patches. Such a patch-to-patch approach can capture comprehensive spatial-temporal correlations that would not be possible in point-to-point approaches. MSSTAN embeds a novel attention mechanism designed to reduce the quadratic complexity of existing attention mechanisms while still effectively capturing temporal and spatial correlations. Extensive comparisons with more than ten state-of-the-art methods on five traffic datasets show that the proposed MSSTAN outperforms existing models in both short-term and long-term forecasting.
Keywords:
Traffic forecasting
graph convolutional network
attention mechanism
multi-scale fusion
Journal
IF:
7.1
Papers:
1.8W
Citations:
6.6W