Diffusion Convolutional Recurrent Neural Network with Rank Influence Learning for Traffic Forecasting

With the rapid development of urban road traffic, accurate and timely road traffic forecasting becomes a critical problem, which is significant for traffic safety and urban transport efficiency. Many methods based on graph convolutional network (GCNs) are proposed to deal with the graph-structured spatio-temporal forecasting problem, since GCNs can model spatial dependency with high efficiency. In order to better capture the complicated dependencies of traffic flow, we introduce rank influence factor to the Diffusion Convolutional Recurrent Neural Network model. The rank influence factor could adjust the importance of neighboring sensor nodes at different proximity ranks with the target node when aggregating neighborhood information. Experiments show a considerable improvement when rank influence factor is used in GCNs with a tolerable time consumption.