Spatio-temporal hierarchical MLP network for traffic forecasting

Traffic forecasting is an indispensable part of intelligent transportation systems. However, existing methods suffer from limited capability in capturing hierarchical temporal characteristics of the traffic time series. To be specific, they neglect the property that the time series is composed of trend-cyclical and seasonal parts. On the other hand, prior methods ignore the natural hierarchical structure of traffic road networks and thus fail to capture the macro spatial dependence of region networks. To address these issues, we propose a novel spatio-temporal hierarchical MLP network (STHMLP) for traffic forecasting. By adopting a decomposition architecture in the STHMLP, trend-cyclical and seasonal features are gradually grasped from multi-scale local compositions of traffic time series. For each scale of traffic time series, we design a fine module and a coarse module to extract spatio-temporal information from roads and regions, respectively. Specifically, the fine module utilizes spatial filters on the frequency domain features of traffic time series to efficiently capture fine-grained spatial dependencies. The coarse module adaptively coarsens road networks to region networks and captures coarse-grained spatial dependencies from region networks. Experiments on four real-world traffic datasets demonstrate the STHMLP outperforms state-of-the-art baselines on traffic forecasting.