Residual attention enhanced Time-varying Multi-Factor Graph Convolutional Network for traffic flow prediction

Precise and timely traffic flow prediction holds significant importance in alleviating traffic congestion. Despite the success of graph convolution traffic flow prediction methods, there is still room for improvement in global spatial feature extraction and external factor measurement. To address this challenge, a novel Residual attention enhanced Time-varying Multi-factor Graph Convolutional Network (RTM-GCN) is proposed. RTM-GCN includes a multi-factor graph convolution module, a spatial enhancement module, and a time-varying module. First, a multi-factor matrix is constructed to measure internal and external factors affecting traffic flow, divided into internal factor matrix containing node distance and data correlation and external factor matrix containing weather and news. Next, a multi-factor graph convolution module is constructed for extracting local spatial features of the multi-factor matrix. Then, a novel spatial enhancement module based on residual convolution operation and self-attention gated linear unit is used to enhance the global spatial features. Finally, a time-varying module based on dilation causal convolution is constructed to enhance the long-term temporal features and output the final prediction values. Compared with the state-of-the-art models, the RTM-GCN model reduces the RMSE by 6.362% on the five real-world datasets. The key source code and data are available at https://github.com/Bounger2/RTMGCN.