JointGraph: Joint Pre-Training Framework for Traffic Forecasting with Spatial-Temporal Gating Diffusion Graph Attention Network.

Accurate highway traffic forecasting is a critical task in intelligent transportation systems (ITSs), which needs to capture complex spatiotemporal dependencies from traffic sensors data. Recently, spatial-temporal graph networks have become one famous technology for various traffic forecasting tasks. Nevertheless, most of these works have assumed that the correlations between the sensors are fixed, so that it is often unable to effectively deal with the more realistic and dynamic traffic environment. To tackle this issue, we propose a joint pretraining framework for traffic flow forecasting with a gating diffusion graph attention network (JointGraph). Specifically, our proposed joint training architecture consists of two parts: network reconstructor (reconstruct a discrete graph from input data) and spatiotemporal model (forecast traffic speed with the generated network). Owing to the capabilities of the network reconstructor in generating graph structure through input node features, it is possible to apply our spatiotemporal model among multiple datasets directly. In addition, our model can expediently use information from data-rich regions and improve traffic forecasting performance on data-lacking regions in highway networks. Experiments are conducted on traffic datasets: METR-LA, PEMS-BAY, and trajectory dataset: BikeNYC. The results show that our JointGraph achieves superior performance with the state-of-the-art baselines, which further indicate that our pre-training mechanism in JointGraph provides an effective solution for multi-dataset cooperative training.