Region-Aware Graph Convolutional Network for Traffic Flow Forecasting

Abstract

Urban traffic flow prediction is a crucial service in intelligent transportation systems. It is very challenging due to the complex spatiotemporal dependencies and inherent uncertainty caused by dynamic urban traffic conditions. Recent work has focused on designing complex Graph Convolutional Network (GCN) architectures to capture spatial dependencies among segment-level traffic status and achieves state-of-the-art performance. But these GCN based methods has two shortcomings. One on hand, they ignore cross-region movement which reflects traffic flow transfer patterns at the regional level. On the other hand, they fail to capture the long-term temporal dependencies of traffic flows due to its non-linearity and dynamics. In order to address the above-mentioned deficiencies, we propose a novel Region-aware Graph Convolution Networks (RGCN) for traffic forecasting. Specially, a DTW-based pooling layer is introduced to capture the cross-regional spatial correlation, based on which a traffic region graph is constructed from the original traffic network and is employed to model cross-region traffic flow. Besides, a transformer-based temporal module is proposed to model long-term and dynamic temporal dependencies across multiple time steps. The proposed model is evaluated on two public traffic network datasets and the experimental results show that RGCN outperforms the state-of-the-art baselines.