CPNet: Conditionally parameterized graph convolutional network for traffic forecasting

Traffic forecasting has attracted increasing attention to its vital role in intelligent transportation systems (ITS), many approaches have been proposed for improving the performance of traffic forecasting. Aiming at capturing complex spatial and temporal dependencies while maintaining efficient inference, we propose a novel model, named Conditionally Parameterized Graph Convolutional Network (CPNet), to model the dynamics of traffic data from spatial and temporal dimensions for traffic forecasting. Specifically, in the temporal dimension, we design a novel multi-scale temporal convolution module, which captures the temporal dynamics of traffic data from different scales. It is beneficial for both long and short-term forecasting. In the spatial dimension, we develop a conditional parameterized graph convolution module to exploit the spatial dependencies in different ranges. In addition, an attention layer is designed to model the nonlinear and dynamic features. Extensive experiments on four real-world datasets demonstrate the superiority of CPNet against the state-of-the-art baseline models.