Multi-range bidirectional mask graph convolution based GRU networks for traffic prediction

Traffic prediction is a critical component in intelligent traffic systems, as the complicated spatial relationships and temporal dynamics in the traffic data make it a challenging task. Existing traffic prediction approaches mostly use fixed adjacency matrices that cannot thoroughly learn the complex spatial relationships, and the learnable adjacency matrix-based methods ignore the noise spatial dependencies. Furthermore, existing approaches usually ignore periodic temporal dependencies and node-specific traffic patterns. To overcome these limitations, we design a multi-range mask graph convolution-based bidirectional Gated Recurrent Unit (GRU) network model for traffic prediction, where three sub-modules are adopted to learn multi-range spatial–temporal features, including recent spatial–temporal features, daily periodic spatial–temporal features, and weekly periodic spatial–temporal features. To learn node-specific forward and backward spatial–temporal features, we propose a bidirectional mask graph convolutional GRU layer, where a mask adaptive adjacency matrix generation algorithm is designed to learn spatial relationships adaptively in the traffic data, and implemented a mask matrix to filter the noise spatial relationships during the adaptive graph learning for more accurate traffic prediction. Extensive experiments to validate the proposed model on four real-world datasets demonstrate that the proposed method has better prediction accuracy than existing state-of-art contrast methods.