Henry Martin, Dominik Bucher, Ye Hong, René Buffat, Christian Rupprecht, and Martin Raubal
The graph-based traffic forecasting workflow.
Short-term traffic forecasting is important for managing the growing traffic volume to reduce its negative impact. Traffic4cast 2019 competition presented publicly available traffic forecasting benchmark dataset, with traffic data encoded as images. The data comprise normalized information of traffic volume, speed, and average direction. The data cover one year in 5-minute intervals in three big cities (Berlin, Istanbul, and Moscow). The street network information is not explicitly provided.
Traffic data encoded as images facilitate using convolutional neural networks (CNNs). Graph convolutional neural networks (GCNs) target irregularly structured domains that can be described as graphs. We develop Graph-ResNet, a GCN approach inspired by residual learning network (ResNet). Graph-ResNet explicitly incorporates the underlying street network information by applying a city-specific mask. The mask filters out pixels below a certain activity threshold; the remaining pixels are extracted and transformed into a graph. To create a graph, the image is defined as a regular grid with pixels as nodes. The nodes are connected to all the adjacent nodes; the traffic information is stored as a vector of node features.
We evaluate Graph-ResNet and compare it to the baseline GCN models and the state-of-the-art CNN models successfully applied in Traffic4cast 2019. The models are trained on Moscow traffic data and are used to predict traffic in Moscow, Berlin, and Istanbul. Our results suggest that while CNN models perform better in the known cities, GCN models show superior performance in the unknown cities, which facilitates transfer learning.
The code is available on GitHub.
Proceedings of the NeurIPS 2019 Competition and Demonstration Track, PMLR 123, 153-163, 2020-08-19.