An Inferential Graph Convolution Network for Explaining Traffic Congestion

Due to the growth of vehicles, traffic congestion is becoming increasingly serious. However, existing methods are used for predicting traffic congestion, which cannot been applied for evaluating traffic congestion. In this paper, we propose an Interpretable Graph Convolution Network called ShapGCN for explaining the reason of traffic congestion by considering its physical and semantic neighbors. Specifically, we first design the physical neighbor embedding and semantic neighbor embedding to collectively encode complex extern factors as well as the complex traffic cascade pattern. To interprete traffic congestion in a complex traffic cascade environment, we use the approximation of shapley value to comprehensively quantify the discovered regions and their importance score. We conduct extensive experiments on the real traffic dataset. The experiment results show our ShapGCN can well explain the reason of traffic congestion.