JointSTNet: Joint Pre-Training for Spatial-Temporal Traffic Forecasting
Danqi Cai, Kunwei Chen, Zhizhe Lin, Duo Li, Teng Zhou, Man-Fai Leung
Abstract
In recent years, there has been rapid development in autonomous vehicle techniques due to the increasing demands in traffic management and travel planning. Accurate spatiotemporal traffic forecasting is crucial for autonomous vehicles. However, most existing methods are task-specific and tailored for specific cities, making them hard to apply for more downstream applications. These methods fail to accurately simulate the common variation of traffic states at different locations based on their spatial distance and time interval. Compared with most existing spatiotemporal prediction baselines, this paper proposes a joint pre-training framework, named JointSTNet, to further enhance the traffic spatiotemporal prediction accuracy while reducing model complexity for traffic forecasting tasks. Firstly, unlike treating traffic spatiotemporal data on roads as independent patterns, the spatial graph capsule module constructs inter-class traffic patterns based on shared dynamics within the dynamic graph structure. Additionally, the temporal encoding gated module proposed in our work expands the temporal receptive field along the time axis. Finally, we maximize reconstruction losses to handle incomplete connections within data involving intra-cluster and inter-cluster regional semantic relationships that can be enhanced through pre-training. Experimental results on four benchmarking datasets from the real world demonstrate that JointSTNet outperforms various types of state-of-the-art baselines.