Lanczos method for spatio‐temporal graph convolutional networks to forecast expressway flow

Abstract

Abstract Traffic forecasting has made pronounced progress with the development of graph convolution networks and the use of the topology of road networks. However, existing works face some limitations when it comes to modelling spatial dependencies. For example, pre‐defined graphs rely on global information to establish spatial relationships, and the spatial receptive field is limited by the polynomial convolutional method. To address these limitations, the authors propose the Lanczos method for Spatio‐Temporal Graph Convolutional Networks (LSTGCN); this approach uses the low‐rank approximation theory to drive pre‐defined graphs to collect important information and eliminate spatial redundancy. Additionally, a learnable dynamic graph feature (LDGF) module generates adaptive graphs and perceives the latent invariance‐variability between graph nodes. To further improve the model's ability to capture spatial dependencies and temporal correlations, multi‐span spatial learning is employed for enlarged receptive fields, which can be well integrated into gated recurrent units. The authors conducted baselines comparison and ablation experiments on real‐world datasets, and the findings show that the LSTGCN model outperforms the baselines and improves prediction accuracy. Notably, this work is the first attempt to use graph low‐rank theory for traffic prediction.

Keywords

• big data
• convolutional neural nets
• intelligent transportation systems
• management and control
• network theory (graphs)
• spatiotemporal phenomena