MFGCN: Multi-faceted spatial and temporal specific graph convolutional network for traffic-flow forecasting

Abstract

Traffic-flow forecasting is a fundamental issue in Intelligent Transportation Systems. Owing to the natural topological structure of road networks, graph convolutional networks (GCNs) have become one of the most promising components. However, existing methods usually implement graph convolution on a static adjacent matrix to capture the spatial relations between road segments, ignoring the fact that the spatial impact varies across time. Moreover, they always learn the common temporal relations for all segments and fail to capture unique patterns for each distinct node. To address these issues, this study explores time-specific spatial dependencies and node-specific temporal relations to utilize GCN for improved traffic-flow forecasting. First, graph convolution is extended to learn the temporal relations between different time slots. The trained graphs contain unique temporal patterns for each node and share patterns among different nodes. Second, a time-specific spatial graph-learning module is designed to establish dynamic spatial dependencies between traffic nodes, which can vary at different times. Finally, an adaptive pattern-sharing mechanism is proposed to adaptively learn the layer-specific patterns and sharing-across-layer patterns. The proposed model is evaluated on four public real-world traffic datasets, and the results show that it outperforms all state-of-the-art methods on the four real-world datasets.