Graph-ResNets for short-term traffic forecasts in almost unknown cities

Abstract

The 2019 IARAI traffic4cast competition is a traffic forecasting problem based on traffic data from three cities that are encoded as images. We developed a ResNet-inspired graph convolutional neural network (GCN) approach that uses street network-based subgraphs of the image lattice graphs as a prior. We train the Graph-ResNet together with GCN and convolutional neural network (CNN) benchmark models on Moscow traffic data and use them to first predict the traffic in Moscow and then to predict the traffic in Berlin and Istanbul. The results suggest that the graph-based models have superior generalization properties than CNN-based models for this application. We argue that in contrast to purely image-based approaches, formulating the prediction problem on a graph allows the neural network to learn properties given by the underlying street network. This facilitates the transfer of a learned network to predict the traffic status at unknown locations.