Dynamic spatio-temporal graph network with adaptive propagation mechanism for multivariate time series forecasting

Spatio-temporal prediction on multivariate time series has received tremendous attention for extensive applications in the real world, where the dynamic unknown spatio-temporal dependencies among variables make the task challenging. Graph neural networks (GNN) are applied to time series due to their powerful ability to model dependencies, where the current approaches either rely on pre-defined or learned fixed graphs to model the inter-node linkages. It ignores the dynamics among variables in spatio-temporal data and adheres to the same information propagation path (graph structure) in different layers of the network, leading to sub-optimal performance of networks. In this paper, we propose a novel dynamic spatio-temporal graph neural network (DSTGN), where the key components are dynamic graph estimation and adaptive guided propagation. In graph estimation, we infer dynamic associations between nodes based on both changing node-level inputs and fixed topology information, which is learned with trainable node embedding, and introduce graph loss to control the graph learning direction. To fully exploit the capabilities of the stacked network, we propose adaptive guided propagation, which automatically change the propagation and aggregation process according to the features extracted at each layer. To learn the process adaptively, we design the learnable guide matrix and incorporate it into a graph convolution framework trained in end-to-end mode. Experimental results show that our method outperforms state-of-the-art baseline methods on four datasets, with comparisons including pre-defined graph-and graph learning-based GNN methods.

Automated summary

This paper proposes a novel dynamic spatio-temporal graph neural network (DSTGN) to tackle the challenge of predicting multivariate time series. The key components of DSTGN are dynamic graph estimation and adaptive guided propagation. Experimental results demonstrate that our method outperforms state-of-the-art baseline methods on four datasets.