Multivariate Time Series Forecasting with Dynamic Graph Learning Layer Reflecting Temporal and Spatial Features

Abstract

Recently, graph neural networks (GNNs) have been successful in processing graph-represented data by incorporating the node relations. One of its critical elements is the availability of node relations, that is to find an effective node relationship when it is unknown. Recently, multivariate time series forecasting with graph neural network (MTGNN) overcomes this by learning node relationships using unsupervised methods. However, the model assumes static relationships that may not fit real-world situations where the node relationship can change over time. This paper addresses the limitations by adopting relational dependencies, dynamically and gradually changing over time, aiming at good forecasting in multivariate time series. We propose a modified version of MTGNN that incorporates a dynamic graph learning layer. Our experimental results demonstrate that the proposed model also applies well to domain, where inter-variable relationships change over time, while maintaining the performance of original MTGNN and several state-of-the-art methods in terms of accuracy.