loading page

DCGCN: Dynamic community graph convolutional network for traffic forecasting
  • +3
  • Yi Xu ,
  • Dalin Zhang ,
  • Yunjuan Peng ,
  • Nan Wang ,
  • Lingyun Lu ,
  • Jiqiang Liu
Yi Xu
Beijing Jiaotong University

Corresponding Author:[email protected]

Author Profile
Dalin Zhang
Author Profile
Yunjuan Peng
Author Profile
Lingyun Lu
Author Profile
Jiqiang Liu
Author Profile


Traffic forecasting is one of the core issues in transportation systems. Graph convolution based spatiotemporal model can process the complex and highly nonlinear traffic data, but rely heavily on the graph construction, some creative works make innovations in dynamic graph modeling and expansion of the graph. However, these works ignored the dynamic community structure of traffic graph and the inter-community traffic flow interaction, and not fully exploited the spatiotemporal characteristics. In this paper, we propose a dynamic community graph convolution model (DCGCN), which can more precisely characterize the dynamic traffic network and capture the dynamic spatiotemporal dependencies by capturing the community properties and community interactions of traffic flow. First, we propose a noval method to capture inter-community traffic flow interactions. Second, we take advantage of the spatiotemporal properties of traffic data and introduce a community-based dynamic traffic graph learning mechanism to construct traffic networks adaptively. Finally, we fuse inter-community traffic flow interactions and dynamic traffic graph to update the traffic graph, introduce a multi-attention spatiotemporal convolution for traffic forecasting. We conduct experiments on three large-scale real datasets containing multiple traffic scenarios, prove the superiority of our model compared with SOTA models such as ASTGCN, MATGCN, and DCRNN. Meanwhile, the visualization in these datasets shows the method can effectively identify the stations that have a large impact in the forecasting process and dynamic community changes, and the extracted inter-station dependencies and traffic flow region interactions are interpretable. The source code is available  athttps://github.com/1998XuYi/DCGCN.