SAMSGL: Series-Aligned Multi-Scale Graph Learning for Spatio-Temporal Forecasting

TL;DR

SAMSGL tackles two central challenges in spatio-temporal forecasting: propagation time delays and high-dimensional node interactions. It introduces a Series-Aligned Graph Convolution to mitigate delays, and a Multi-Scale Graph Structure Learning module to capture global delayed, global non-delayed, and local interactions, fused via Graph-FC blocks. Empirical results on WeatherBench and traffic datasets show superior accuracy over state-of-the-art methods, with ablations confirming the value of each component. The approach offers a principled, end-to-end framework for robust spatio-temporal prediction in weather and traffic domains, with potential for broader deployment where time delays and multi-scale interactions are critical.

Abstract

Spatio-temporal forecasting in various domains, like traffic prediction and weather forecasting, is a challenging endeavor, primarily due to the difficulties in modeling propagation dynamics and capturing high-dimensional interactions among nodes. Despite the significant strides made by graph-based networks in spatio-temporal forecasting, there remain two pivotal factors closely related to forecasting performance that need further consideration: time delays in propagation dynamics and multi-scale high-dimensional interactions. In this work, we present a Series-Aligned Multi-Scale Graph Learning (SAMSGL) framework, aiming to enhance forecasting performance. In order to handle time delays in spatial interactions, we propose a series-aligned graph convolution layer to facilitate the aggregation of non-delayed graph signals, thereby mitigating the influence of time delays for the improvement in accuracy. To understand global and local spatio-temporal interactions, we develop a spatio-temporal architecture via multi-scale graph learning, which encompasses two essential components: multi-scale graph structure learning and graph-fully connected (Graph-FC) blocks. The multi-scale graph structure learning includes a global graph structure to learn both delayed and non-delayed node embeddings, as well as a local one to learn node variations influenced by neighboring factors. The Graph-FC blocks synergistically fuse spatial and temporal information to boost prediction accuracy. To evaluate the performance of SAMSGL, we conduct experiments on meteorological and traffic forecasting datasets, which demonstrate its effectiveness and superiority.

Figures (7)

• Figure 1: The framework of Series-Aligned Multi-Scale Graph Learning (SAMSGL). (a) The overall structure of SAMSGL. The Multi-scale graph structure learning module is developed to generate spatial relations of different scales, which are denoted as $\mathcal{A}_{\rm nd}$, $\mathcal{A}_{\rm d}$, and $\mathcal{A}_{\rm l}$, respectively. These generated adjacency matrices, along with the node states $X$, are input into stacks of Graph-FC blocks. (b) The structure of Multi-Graph Convolution in the Graph-FC blocks. (c) The workflow of Series-Aligned Graph Convolution where "fft" represents fast Fourier transform and "ifft" is inversed fast Fourier transform.
• Figure 2: Performance comparison on various spatio-temporal datasets. Panels (a), (e), (i), and (m) are the ground truths. (b-d), (f-h), and (j-l) show, in turn, the errors of RGSL, Corrformer and our SAMSGL model on wind speed, temperature and humidity datasets, respectively. (n-p) display the errors of RGSL, MGCRN and our SAMSGL model on traffic flow dataset, respectively. The blue boxes in panels (b)-(d), (f)-(h), and (j)-(l) highlight regions with significant errors.
• Figure 3: Comparison of forecasting performance at a time step on wind dataset. The red boxes zoom in on typical regions to illustrate the differences in predicted spatial patterns.
• Figure 4: Comparison of forecasting performance in single node's temporal dynamics. (a), (b) are the predicted time series on Temperature dataset. (c), (d) are the predicted time series on Wind dataset. (e), (f) are the predicted time series on PeMSD8 dataset.
• Figure 5: Comparison of spatio-temporal forecasting performance among SAMSGL (ours), MGCRN and RGSL on various datasets. (a), (c), and (e) present the prediction results on the Wind dataset, the Temperature dataset, and the PeMSD8 dataset, respectively. (b), (d), and (f) display the MAE of the methods on the Wind dataset, the Temperature dataset, and the PeMSD8 dataset, respectively.