Abstract:Multivariate time series (MTS) forecasting has shown great importance in numerous industries. Current state-of-the-art graph neural network (GNN)-based forecasting methods usually require both graph networks (e.g., GCN) and temporal networks (e.g., LSTM) to capture inter-series (spatial) dynamics and intra-series (temporal) dependencies, respectively. However, the uncertain compatibility of the two networks puts an extra burden on handcrafted model designs. Moreover, the separate spatial and temporal modeling naturally violates the unified spatiotemporal inter-dependencies in real world, which largely hinders the forecasting performance. To overcome these problems, we explore an interesting direction of directly applying graph networks and rethink MTS forecasting from a pure graph perspective. We first define a novel data structure, hypervariate graph, which regards each series value (regardless of variates or timestamps) as a graph node, and represents sliding windows as space-time fully-connected graphs. This perspective considers spatiotemporal dynamics unitedly and reformulates classic MTS forecasting into the predictions on hypervariate graphs. Then, we propose a novel architecture Fourier Graph Neural Network (FourierGNN) by stacking our proposed Fourier Graph Operator (FGO) to perform matrix multiplications in Fourier space. FourierGNN accommodates adequate expressiveness and achieves much lower complexity, which can effectively and efficiently accomplish the forecasting. Besides, our theoretical analysis reveals FGO's equivalence to graph convolutions in the time domain, which further verifies the validity of FourierGNN. Extensive experiments on seven datasets have demonstrated our superior performance with higher efficiency and fewer parameters compared with state-of-the-art methods.

What problem does this paper attempt to address?

This paper attempts to solve two core problems in multivariate time - series (MTS) prediction: 1. **Unified Modeling of Spatial and Temporal Dependencies**: Existing prediction methods based on graph neural networks (GNNs) usually need to use graph networks (such as GCN) simultaneously to capture the spatial dynamics between series and time networks (such as LSTM) to capture the temporal dependencies within series. However, this separately - modeled approach violates the unity of spatio - temporal dependencies in the real world and increases the manual burden of model design, thus limiting the prediction performance. 2. **Efficient Processing of Large - Scale Graph Structures**: When using classical graph neural networks (such as GCN and GAT), as the number of variables \( N \) and the window length \( T \) in multivariate time - series increase, the number of nodes in the constructed hypervariate graph will increase significantly, resulting in high computational complexity (usually quadratic complexity) and difficult optimization. To solve these problems, the paper proposes a brand - new perspective - **Multivariate Time - Series Prediction from a Pure Graph Perspective**. Specifically: - **Introducing the Hypervariate Graph**: Each time - series value (whether it is a variable or a timestamp) is regarded as a node of the graph, and the sliding window is represented as a space - time fully - connected graph. This enables the classical time - series prediction task to be reformulated as the prediction of the hypervariate graph. - **Proposing the Fourier Graph Neural Network (FourierGNN)**: By performing matrix multiplication in the Fourier space, the Fourier Graph Operator (FGO) is proposed. FourierGNN stacks FGO layers to achieve lower complexity (log - linear complexity) while maintaining sufficient expressive power, thus efficiently completing the prediction task. In addition, theoretical analysis shows that FGO is equivalent to graph convolution in the time domain, further verifying the effectiveness of FourierGNN. Experimental results show that FourierGNN performs excellently on seven real - world datasets. Compared with the existing state - of - the - art methods, its prediction accuracy is improved by more than 10%, and it also has less training time and number of parameters. In summary, by re - thinking the multivariate time - series prediction task, this paper proposes a brand - new pure - graph - perspective method, which effectively solves the limitations of existing methods in spatio - temporal dependencies and computational efficiency.

FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective

Edge-Varying Fourier Graph Networks for Multivariate Time Series Forecasting

Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks

ForecastGrapher: Redefining Multivariate Time Series Forecasting with Graph Neural Networks

MTHetGNN: A Heterogeneous Graph Embedding Framework for Multivariate Time Series Forecasting

Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting

Dynamic Personalized Graph Neural Network with Linear Complexity for Multivariate Time Series Forecasting

TDG4MSF: A temporal decomposition enhanced graph neural network for multivariate time series forecasting

Multi-channel Fusion Graph Neural Network for Multivariate Time Series Forecasting.

Multivariate Time Series Forecasting with Transfer Entropy Graph

Parallel Multi-Scale Dynamic Graph Neural Network for Multivariate Time Series Forecasting

It is all Connected: A New Graph Formulation for Spatio-Temporal Forecasting

Multivariate Time Series Forecasting Based on Causal Inference with Transfer Entropy and Graph Neural Network

Multiple Time Series Forecasting with Dynamic Graph Modeling

Time-Decay Dynamic Graph Neural Network for Multivariate Time Series Forecasting

Learning Dynamic Temporal Relations with Continuous Graph for Multivariate Time Series Forecasting (student Abstract).

TimeGNN: Temporal Dynamic Graph Learning for Time Series Forecasting

A Study of Joint Graph Inference and Forecasting

Time-aware Personalized Graph Convolutional Network for Multivariate Time Series Forecasting

Multivariate Time Series Forecasting with Dynamic Graph Neural ODEs

MDG: A Multi-Task Dynamic Graph Generation Framework for Multivariate Time Series Forecasting