Graph Construction Method for GNN-Based Multivariate Time-Series Forecasting

Abstract

Multivariate time-series forecasting (MTSF) plays an important role in diverse real-world applications. To achieve better accuracy in MTSF, time-series patterns in each variable and interrelationship patterns between variables should be considered together. Recently, graph neural networks (GNNs) has gained much attention as they can learn both patterns using a graph. For accurate forecasting through GNN, a well-defined graph is required. However, existing GNNs have limitations in reflecting the spectral similarity and time delay between nodes, and consider all nodes with the same weight when constructing graph. In this paper, we propose a novel graph construction method that solves aforementioned limitations. We first calculate the Fourier transform-based spectral similarity and then update this similarity to reflect the time delay. Then, we weight each node according to the number of edge connections to get the final graph and utilize it to train the GNN model. Through experiments on various datasets, we demonstrated that the proposed method enhanced the performance of GNN-based MTSF models, and the proposed forecasting model achieve of up to 18.1% predictive performance improvement over the state-of-the-art model.