No abstract

Most works on graph neural networks (GNNs) for traffic speed prediction assume near-complete data and little variance of base speed levels. However, both assumptions do not necessarily hold true for network-wide probe vehicle data (PVD). Therefore, we applied two state-of-the-art GNNs to sparse PVD from a road network with highly varying speed levels and to dense motorway data for comparison. We introduce two methods to adapt preexisting GNNs for improved prediction performance: normalization of speed values with respect to the base speed levels of different roads led to significant improvements of prediction performance on both datasets. Using the number of observations supporting each speed value as an additional input feature can improve prediction performance. Furthermore, we identified characteristics of data and models encouraging the use of either method. As no fitting dataset to evaluate these approaches was found, a novel dataset derived from PVD is introduced. It features sparse speed values and underlying numbers of observations for a road network with varying speed levels.