MTGCN: Multi-graph Fusion Based Temporal-Spatial Convolution for Traffic Flow Forecasting

Abstract

Traffic flow prediction plays an important role in traffic management and urban planning. This task is challenging due to the dependence of the road network and the complexity of information. The existing forecasting methods usually consider the spatio-temporal correlation of traffic flow, which overlook the rich semantic correlation between the nodes of the road network. For example, roads that have similar functional city blocks tend to have similar traffic patterns. To make use of the semantic information contained in road network, we propose a temporal-spatial convolution model based on multi-graph fusion (namely MTGCN). Specifically, we build adjacency graph, similarity graph and reachability graph from the original traffic road network, and fuse them by a learnable parameter-based fusion method. Then, we alternately use causal convolution module and graph convolution module to fully capture the potential temporal dependencies and spatial dependence with semantic correlation in the road network. Experimental results on two real datasets show that our method achieves better performance and consistently outperforms other baselines in short, middle, and long-term forecasting task. From the ablation experiments, we also demonstrate the proposed multi-graph mechanism is effective and can effective encoding the non-Euclidean spatial correlation and semantic attributes in road network.