Graph CNNs for urban traffic passenger flows prediction

Urban traffic passenger flows prediction has always been a great challenge in transportation field. Efficiently and correctly predicting the future flows of various regions can improve traffic resources scheduling and reduce the possibility of accidents. However, factors which affect the change of traffic passenger flows are complex, including interlaced lines and stations in large areas, diversified traveling demands for people, accidents and bad weathers. So the predicting algorithms or models should be more sensitive to multiply elements and their effecting patterns. Recently, deep learning performs the excellent ability to extract high dimensional spatial-temporal characters in regression and classification tasks. In this paper, we propose a new modeling method for urban traffic passenger flows. Instead of the grid matrices, we quantify the relationship between stations and represent it by a undirected graph. Then we sort the stations by their passenger flows and construct the two-channels graph flows matrices as the input of deep convolutional neural networks. To increase the temporal information of inputs, we also combine the input matrices with recent historical samples. In addition, we add date markers to correct the final prediction flows to further improve the accuracy. Finally we evaluate our model with the real Beijing subway data and compare with other traditional models on short-term passenger flows prediction tasks. Experiments show that our model including multidimensional flows graph matrices and the deep learning model can significantly improve the prediction accuracy.