Spatial-Temporal Hierarchical Graph Convolutional Networks for Traffic Forecasting

Abstract

Traffic forecasting is a critical task in transportation planning and management, which requires modeling the complex spatial and temporal dependencies in traffic data. Most current methods employ Graph Convolutional Networks (GCN) to model spatial dependencies, and Recurrent Neural Networks (RNN) or Temporal Convolutional Networks (TCN) to model temporal dependencies. However, the representation ability of such methods is limited due to: 1) the conventional temporal extract models, such as RNN and TCN, suffer from limited flexibility, specifically, RNN can only capture temporal dependencies sequentially, while TCN is constrained by its multi-layer dilated convolution structure; 2) spatial and temporal dependencies are intricately intertwined in the real world, but most methods fail to capture this spatial-temporal correlation resulting in sub-optimal performance. To this end, we propose the Spatial-Temporal Hierarchical Graph Convolutional Networks (STHGCN), in which we design Spatial-Temporal Hierarchical Graph (STHG) to simultaneously model spatial and temporal dependencies. Specifically, to model temporal dependencies more flexibly, we introduce two crucial components: the Local Temporal Transmission Matrix (LTTM) and the Multi-hop Temporal Similarity Matrix (MTSM). The LTTM captures adjacent temporal dependencies, while the MTSM captures multi-hop temporal dependencies. We further propose a Temporal Neighbor Fusion model that combines the LTTM and MTSM to obtain the adjacency matrix of STHG. Additionally, accounting for spatial-temporal correlation, we exploit the spatial GCN results as the STHG nodes which allows us to learn spatial and temporal dependencies simultaneously via the temporal GCN. Our experiments on four real-world datasets demonstrate that STHGCN outperforms the state-of-the-art methods for traffic forecasting. The code is available at https://github.com/sqy123qwer/STHGCN