Time-adaptive graph convolutional network for traffic prediction

Traffic prediction is of great significance to route planning and transportation management. Due to the complex nonlinear spatiotemporal dependence between traffic data and various unpredictable traffic conditions, traffic prediction has been considered as a challenging research topic. Graph convolutional network (GCN) has been widely used to model the spatial correlation between traffic nodes, typically with a fixed weighted graph. This paper proposes a time-adaptive graph convolutional network (TAGCN) to capture the time-varying spatial relationship in traffic information, and designs a skip temporal convolutional network (skip-TCN) to acquire multi-level temporal patterns in time series. TAGCN adapts encoder-decoder structure for multi-step prediction. The encoder encodes the input traffic information, and the decoder obtains the predicted result. Experimental results on two real-world traffic datasets demonstrate the superior performance of our model.