(A) study of global-local decomposition for multivariate time series modeling

다국어 초록 (Multilingual Abstract)

Since the fundamentals of decomposing time series into non-observable components, the decomposition approaches have been considered as a starting point for time series modeling. The conventional time series decomposition methods have been developed for a univariate or small group of time series. In modern applications, however, big data systems generate numerous time series that cannot utilize these models. Instead, a new type of modeling approach, referred to as global-local models, has arisen. The term global is defined as common behaviors among a substantial amount of time series, and local denotes idiosyncratic features of individual series.
This study is focused on a novel design to decompose time series into global and local series. Although recent studies have shown that global and local components convincingly help improve forecasting performance, existing global-local approaches typically treat the global components as additional hidden states inside the model without providing the decomposed series for further analysis.
To alleviate this fundamental flaw, we introduce a simple but effective global-local decomposition framework, named DeepGate, which empowers disentangling and understanding the underlying global and local series beneath multivariate time series. We extend the linear global factors in Dynamic Factor Models to non-linear factors based on deep learning architectures while retaining the additive properties between components. In this study, we confirmed that DeepGate can disentangle the interpretable global series and local components that can be utilized in further analysis, rather than forms of hidden features.
In addition, we propose a novel multi-step forecasting framework that imposes global-local decomposition in DeepGate. Since global and local series can have different aspects of dynamics, we employ similar but not shared neural architectures on them. The experiments on the synthetic and real-world datasets demonstrated the effectiveness of the proposed forecasting framework.
Finally, we present an anomaly detection method based on the decomposition. While we use disentangled global series explicitly, we build a reconstruction module to detect abnormal behaviors in the multivariate time series. We proved that the proposed method outperforms several existing methods on the benchmark datasets.
Through a profound analysis of experimental results, we are convinced that the proposed global-local decomposition can enhance further modeling applications, beyond disentangling the multivariate time series itself.

참고문헌 (Reference)

1. Comparison of periodogram tests, L. A. McSweeney, vol . 76 , no . 4 , pp . 357-369, , 2006

2. Seasonal Adjustment by Signal Extraction, J. P. Burman, vol . 143 , no . 3, , 1980

3. Financial time series forecasting using support vector machines, K. J. Kim, vol . 55 , no . 1 ? 2 , pp . 307 ? 319, , 2003

4. The X-11 variant of the census method II seasonal adjustment program, J. Shiskin, 15th ed, , 1967

5. `` Freeway performance measurement system : mining loop detector data, Chen , Chao , et al, vol . 1748 , no . 1 , pp.96- 102, , 2001

6. Deep autoencoding Gaussian mixture model for unsupervised anomaly detection, B. Zong, , 2018

7. Unsupervised Anomaly Detection via Variational Auto- Encoder for Seasonal KPIs in Web Applications, H. Xu, pp . 187-196, , 2018

8. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting, S. Li, , 2019

9. The dynamic factor analysis of economic time-series models , ” in Latent Variable in Socioeconomic Models, J. Geweke, pp . 365-383, , 1977