Multi-graph fusion based graph convolutional networks for traffic prediction

Abstract

Traffic prediction is significant for transportation management and travel route planning, and it is challenging as the spatial dependencies are complex and temporal patterns are dynamic. Local spatial dependencies exist between nodes nearby, and global spatial dependencies are between distant nodes with similar traffic patterns. One direct method to capture multiple spatial dependencies is to design a prediction model with multiple graph convolutional networks. However, it introduces high memory and parameter costs. The same with the learnable adjacency matrix-based approaches. Furthermore, existing methods are inefficient for temporal dependency modeling. To overcome such limitations, we propose a multi-Graph Fusion-based Graph Convolutional Network (GFGCN) for traffic prediction, where a multi-graph fused graph convolutional module is proposed without building multiple graph convolutional networks. The adjacency matrix in one graph convolutional network can reflect multiple spatial relationships through subspace merging on the Grassmann manifold. Moreover, a temporal module combined with the attention mechanism and a dilated convolutional network to model the temporal dynamic efficiently is designed. For validation and analysis, extensive experiments on three real-world datasets are performed. Experimental results show that the proposed GFGCN outperforms the baselines with better prediction accuracy.