Sampling Spatial-Temporal Attention Network for Traffic Forecasting.

Spatial-temporal graph learning has been a critical approach to modeling complex dependencies between variables in multivariate time series such as traffic series. However, when modeling the spatial dependencies between traffic nodes, most existing approaches regard the predefined or adaptive graphs as static, overlooking the dynamic nature of realistic graphs that change over time. In addition, due to limitations in model complexity and information density, most models only consider short-term historical series for future series forecasting, failing to account for the periodicity of long-term series. Furthermore, spatial-temporal indistinguishability is also a challenge for many approaches. Aiming to address these problems, we propose a novel neural network framework S ampling S patial- T emporal A ttention N etwork ( SSTAN ) to effectively capture latent spatial and temporal dependencies. Firstly, a spatial encoder is proposed to learn multi-level dynamic graph structures. Secondly, a temporal encoding framework with long-term sampling and temporal encoders is designed to capture long-term periodic features that contain high-density information. Thirdly, with the global information of the entire graph and long-term series, our model overcomes the challenge of spatial-temporal indistinguishability to distinguish similar series with different latent patterns. Finally, experimental results on three real-world datasets not only demonstrate the superiority of our model over baselines on traffic forecasting but also illustrate our model’s effectiveness in spatial-temporal dependencies learning.