Learning Time-aware Graph Structures for Spatially Correlated Time Series Forecasting

Minbo Ma,Jilin Hu,Christian S. Jensen,Fei Teng,Peng Han,Zhiqiang Xu,Tianrui Li
DOI: https://doi.org/10.1109/icde60146.2024.00338
2024-01-01
Abstract:Spatio-temporal forecasting of future values of spatially correlated timeseries is important across many cyber-physical systems (CPS). Recent studiesoffer evidence that the use of graph neural networks to capture latentcorrelations between time series holds a potential for enhanced forecasting.However, most existing methods rely on pre-defined or self-learning graphs,which are either static or unintentionally dynamic, and thus cannot model thetime-varying correlations that exhibit trends and periodicities caused by theregularity of the underlying processes in CPS. To tackle such limitation, wepropose Time-aware Graph Structure Learning (TagSL), which extracts time-awarecorrelations among time series by measuring the interaction of node and timerepresentations in high-dimensional spaces. Notably, we introduce timediscrepancy learning that utilizes contrastive learning with distance-basedregularization terms to constrain learned spatial correlations to a trendsequence. Additionally, we propose a periodic discriminant function to enablethe capture of periodic changes from the state of nodes. Next, we present aGraph Convolution-based Gated Recurrent Unit (GCGRU) that jointly capturesspatial and temporal dependencies while learning time-aware and node-specificpatterns. Finally, we introduce a unified framework named Time-aware GraphConvolutional Recurrent Network (TGCRN), combining TagSL, and GCGRU in anencoder-decoder architecture for multi-step spatio-temporal forecasting. Wereport on experiments with TGCRN and popular existing approaches on fivereal-world datasets, thus providing evidence that TGCRN is capable of advancingthe state-of-the-art. We also cover a detailed ablation study and visualizationanalysis, offering detailed insight into the effectiveness of time-awarestructure learning.
What problem does this paper attempt to address?