DeepGate: Global-local decomposition for multivariate time series modeling

• Published in: Information sciences Vol. 590; pp. 158 - 178
• Main Authors: Park, Jinuk, Park, Chanhee, Choi, Jonghwan, Park, Sanghyun
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.04.2022
• Subjects: Global-local decomposition; Interpretable neural networks; Multi-step forecasting; Multivariate time series; Time series analysis; Interpretable neural networks; Multivariate time series; Time series analysis; Global-local decomposition; Multi-step forecasting
• DOI: 10.1016/j.ins.2022.01.015
• ISSN: 0020-0255

•A global–local decomposition is introduced for interpretable time series modeling.•A time series model that sustains the global–local decomposition is proposed.•The denoising training method is applied to make the forecasting model robust.•The proposed method outperforms the baselines on several datasets for forecasting.•Explicit decomposition empowers the model to disentangle underlying shared effects. In multivariate time series, a substantial amount of variables exhibit common dynamics stemming from a small number of global factors. Recent studies have shown that the shared information from global components can enhance the forecasting performance of time series. However, existing global–local approaches treat the global factors as additional hidden states inside the model without providing global series for downstream analysis. In this study, we propose DeepGate, a novel time series forecasting framework based on the explicit global–local decomposition. To retain the global and local series property, we have built decomposition and prediction modules separately. In this way, DeepGate can produce interpretable global series for further tasks while improving forecasting performance with the aid of global and local series. In addition, to alleviate the discrepancy between the training and testing steps, we employ a denoising training technique for multi-step forecasting problems. In numerous experiments on real-world benchmarks for time series forecasting, DeepGate outperforms the baselines including existing global–local models. In particular, the experimental results on synthetic tasks demonstrate that our model can effectively extract underlying global series.