Graph Wavenet+TimeGrad Probabilistic Model for High-Dimensional Time Series Prediction
Document Type
Conference Proceeding
Publication Date
1-1-2023
Abstract
Time series prediction has been widely utilized across various domains, including financial markets, weather forecasting, transportation, and Internet of Things (IoT). High-dimensional time-series data poses challenges for traditional time series models in effectively managing its complexity and non-linearity. In this study, we propose a novel approach by combining the truncated GraphWaveNet architecture with the TimeGrad process. This combination allows for probabilistic multi-step prediction of high-dimensional time series data. The primary components of the GraphWaveNet architecture are the gated dilated causal convolutional neural network (GDC-CNN), graph convolutional neural network (GCNN), and residual networks. GDC-CNN and residual networks are employed to extract trend and short-term movement information from the time series data while addressing the gradient degradation/explosion issues. The function of GCNN is to leverage the graph structure and learn effective representations of features. The TimeGrad process generates probabilistic output. Experimental results on five real-world multi-variate time series datasets demonstrate that the truncated GraphWaveNet+TimeGrad model outperforms benchmark models in terms of CRPSsumcriteria. Particularly, GraphWaveNet+TimeGrad probabilistic model exhibits superior performance for traffic prediction.
Publication Source (Journal or Book title)
Proceedings - 22nd IEEE International Conference on Machine Learning and Applications, ICMLA 2023
First Page
1833
Last Page
1838
Recommended Citation
Sun, X., & Chen, J. (2023). Graph Wavenet+TimeGrad Probabilistic Model for High-Dimensional Time Series Prediction. Proceedings - 22nd IEEE International Conference on Machine Learning and Applications, ICMLA 2023, 1833-1838. https://doi.org/10.1109/ICMLA58977.2023.00278
