Learnable Gated Graph Convolutional Residual Network for Traffic Prediction

Abstract

In Intelligent Transportation Systems (ITS), traffic data prediction is a crucial component. Accurate traffic state prediction depends on appropriate modeling of complex spatio-temporal correlations of traffic data. The traffic data contains nonlinear and intricate correlations, which poses a huge challenge for accurate prediction. To completely capture spatio-temporal correlations, a traffic data prediction model based on a learnable gated graph convolution residual network is proposed. This model uses multi-receptive field dilated causal convolution (MRDCC) and learnable graph convolution to capture the spatio-temporal correlations respectively. Furthermore, the proposed model also designs a gating mechanism between different graph convolutional layers to alleviate the over-smoothing problem which is caused by multi-layer graph convolution stacking. To further capture temporal trends across different periods, a multi-branch residual network strategy is also introduced in this paper. The experimental results on multiple traffic datasets demonstrate that the predictive performance of our proposed model exceeds existing models.