Dependency Learning Graph Neural Network for Multivariate Forecasting

Arth Patel, Abishek Sriramulu, Christoph Bergmeir, Nicolas Fourrier

Abstract: Multivariate time series forecasting is an important topic in various disciplines. Many deep learning architectures have been proposed for accurate multivariate forecasting. However, most existing models fail to learn the dependencies between different time series. Lately, studies have shown that implementations of Graph Neural Networks in the field of Natural Language, Computer Vision, and Time Series have achieved exceptional performance. In this paper, we propose an attention-based Multivariate Dependency Learning Graph Neural Network, which aims to better learn the dependencies among variables of a multivariate dataset. The attention scores corresponding to each variable complement the construction process of the graph adjacency matrix to model the spatial dependencies. Our experiments on benchmark datasets show that the proposed architecture improves accuracy on different benchmark datasets compared with the state-of-the-art baseline models.