Abstract:Time series forecasting lies at the core of important real-world applications in many fields of science and engineering. The abundance of large time series datasets that consist of complex patterns and long-term dependencies has led to the development of various neural network architectures. Graph neural network approaches, which jointly learn a graph structure based on the correlation of raw values of multivariate time series while forecasting, have recently seen great success. However, such solutions are often costly to train and difficult to scale. In this paper, we propose TimeGNN, a method that learns dynamic temporal graph representations that can capture the evolution of inter-series patterns along with the correlations of multiple series. TimeGNN achieves inference times 4 to 80 times faster than other state-of-the-art graph-based methods while achieving comparable forecasting performance

What problem does this paper attempt to address?

This paper attempts to solve several key problems in time - series prediction: 1. **Complex Patterns and Long - Term Dependencies**: Time - series data in the real world often contains complex patterns and long - term dependency relationships, which pose challenges to the predictive ability of models. Traditional statistical methods and simple neural network models perform poorly when dealing with these characteristics. 2. **Computational Cost and Scalability**: Existing time - series prediction methods based on graph neural networks (GNNs) can capture the correlations between time - series, but they are usually costly to train and difficult to scale to large - scale datasets. 3. **Learning of Dynamic Graph Structures**: Most existing methods rely on static graph structures to represent the relationships between time - series, which cannot effectively capture the dynamically changing patterns between time - series. Therefore, a method that can dynamically generate graph structures is needed to better adapt to the characteristics of time - series data. ### Solutions To solve the above problems, the author proposes a new method - TimeGNN. The main contributions of TimeGNN are as follows: - **Dynamic Time - Graph Learning**: TimeGNN extracts the features of time - series through a sliding window and constructs a dynamic time - graph structure based on these features. Each time step is regarded as a node, and the edges of the graph represent the similarities between time steps. - **Efficient Predictive Performance**: TimeGNN significantly reduces the inference and training time while maintaining the predictive performance. Compared with existing graph - based methods, the inference speed of TimeGNN is increased by 4 to 80 times. - **End - to - End Learning**: TimeGNN adopts the Gumbel softmax technique, making the learning process of the graph structure differentiable, thereby achieving end - to - end optimization. ### Method Overview 1. **Time - Series Feature Extraction**: - Use stacked dilated convolutions to extract features from the sliding window. - The feature representation of each window \( S_k \) is \( z_k \), and multiple feature maps are spliced together through a fully - connected layer. 2. **Graph Structure Learning**: - Calculate the similarity of the feature vectors of each time step, and use the Gumbel softmax technique to generate a dynamic adjacency matrix \( A_k \). - Only learn the upper - triangular part of the adjacency matrix to ensure that the graph structure is forward (i.e., does not backtrack to historical time steps). 3. **Graph Neural Network Prediction**: - Use GraphSAGE as the basic GNN module, and generate node embeddings by aggregating the neighbor features of nodes. - The node embedding of the last time step is passed to the output module to generate the final predicted value. ### Experimental Results - **Scalability**: TimeGNN performs excellently in terms of inference and training time on different datasets, especially when the number of variables and the window size increase. - **Predictive Performance**: On multiple datasets, the predictive performance of TimeGNN is comparable to or better than that of existing graph - based methods, especially in long - term prediction. ### Conclusion TimeGNN provides an efficient and scalable method for dynamically generating graph structures from raw time - series data and performing time - series prediction. This method significantly reduces the computational cost while maintaining the predictive performance, and is suitable for processing large - scale time - series data.

TimeGNN: Temporal Dynamic Graph Learning for Time Series Forecasting

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

Dynamic Spatio-Temporal Graph Network with Adaptive Propagation Mechanism for Multivariate Time Series Forecasting

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

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

Dynamic Graph Learning-Neural Network for Multivariate Time Series Modeling

Temporal Graph Networks for Deep Learning on Dynamic Graphs

Graph Deep Learning for Time Series Forecasting

Multiple Time Series Forecasting with Dynamic Graph Modeling

Scalable Spatiotemporal Graph Neural Networks

TimeSGN: Scalable and Effective Temporal Graph Neural Network

Learning Time-aware Graph Structures for Spatially Correlated Time Series Forecasting

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

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

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

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

A Survey on Graph Neural Networks for Time Series: Forecasting, Classification, Imputation, and Anomaly Detection

Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting

Dynamic Graph Structure Learning for Multivariate Time Series Forecasting

Discrete Graph Structure Learning for Forecasting Multiple Time Series

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