Spectral Temporal Graph Neural Network for Multivariate Time-series\n Forecasting

Abstract

Multivariate time-series forecasting plays a crucial role in many real-world\napplications. It is a challenging problem as one needs to consider both\nintra-series temporal correlations and inter-series correlations\nsimultaneously. Recently, there have been multiple works trying to capture both\ncorrelations, but most, if not all of them only capture temporal correlations\nin the time domain and resort to pre-defined priors as inter-series\nrelationships.\n In this paper, we propose Spectral Temporal Graph Neural Network (StemGNN) to\nfurther improve the accuracy of multivariate time-series forecasting. StemGNN\ncaptures inter-series correlations and temporal dependencies \\textit{jointly}\nin the \\textit{spectral domain}. It combines Graph Fourier Transform (GFT)\nwhich models inter-series correlations and Discrete Fourier Transform (DFT)\nwhich models temporal dependencies in an end-to-end framework. After passing\nthrough GFT and DFT, the spectral representations hold clear patterns and can\nbe predicted effectively by convolution and sequential learning modules.\nMoreover, StemGNN learns inter-series correlations automatically from the data\nwithout using pre-defined priors. We conduct extensive experiments on ten\nreal-world datasets to demonstrate the effectiveness of StemGNN. Code is\navailable at https://github.com/microsoft/StemGNN/\n