SASTDGCN: Self-Attention Based Spatial-Temporal Double Graph Convolutional Networks for Traffic Flow Forecasting

Abstract

As a core technology of Intelligent Transportation System, traffic flow forecasting has a wide range of applications. Existing methods typically utilize graph neural network (GNNs) and temporal neural networks (TNNs) to model spatial and temporal dependencies. However, these works still have following limitations: 1) Most methods model spatial and temporal dependencies in a static manner, which limits the ability to learn dynamic traffic patterns; 2) TNNs have difficulty in extracting the temporal dependencies in global fields. To this end, in this paper, we propose a Self-Attention Based Spatial-Temporal Double Graph Convolutional Networks (SASTDGCN) for traffic flow forecasting. Specifically, we construct temporal correlation graph to represent the dependencies among timestamps. Then, we design a spatial-temporal self-attention module to generate dynamic spatial and temporal adjacency matrices for capturing the dynamic spatial-temporal correlations. Furthermore, graph convolution module is proposed to extract the spatial patterns using Graph Convolutional Network (GCN) and capture temporal patterns using Relational Graph Convolutional Network (R-GCN). To the best of our knowledge, this is the first time RGCN has been used to model temporal dynamics in traffic flow forecasting. Finally, to validate the performance of the proposed SASTDGCN, we conduct extensive experiments on three realworld traffic datasets. Experimental results show our model outperforms the baseline methods.