Multi-view Cascading Spatial-Temporal Graph Neural Network for Traffic Flow Forecasting

Zibo Liu, Kaiqun Fu, Xiaotong Liu

Abstract: Spatial-temporal patterns have been applied in many areas, such as traffic forecasting, skeleton-based recognition, and so on. In such areas, researchers can convert the prior knowledge into graphs and combine the latent graph dependencies into original features to get better representation. However, few works focus on the underlying pattern in the original feature, and they cannot capture the flexible interaction both spatially and temporally. What is more, they often ignore the heterogeneity in spatial-temporal data. In this paper, we solve this problem by designing a novel model, Multi-view Cascading Spatial-temporal Graph Neural Network. Our model has a cascading structure to enhance interaction and capture heterogeneity. Also, it takes the differencing orders of flow data into account to get a better representation and contains specific coupled graphs designed based on the sliding window technique. Extensive experiments are conducted on four real-world datasets, demonstrating that our method achieves state-of-the-art performance and outperforms other baselines.