A Decomposition Based Dual Projection Model for Multivariate Time Series Forecasting and Anomaly Detection

Peng Zhang, Qin Xie, Jaesik Choi

Keywords: Multivariate Time Series, Decomposition, Projection, Forecasting, Anomaly Detection

TL;DR: A seasonal-trend decomposition based model with the channel-wise and sequence-wise dual projection is developed for efficient and accurate multivariate time series forecasting and anomaly detection.

Abstract: Efficient anomaly detection and diagnosis in multivariate time series data is of great importance for various application areas. Forecasting of long-sequence time series is an important problem to prepare for future changes. An accurate prediction can help to detect anomaly events beforehand and make better decisions. It seems that one has to use more complex structures for deep learning models to get better performance, e.g., the recent surge of Transformer variants for time series modeling. However, such complex architectures require a large amount of training data and extensive computing resources. In addition, many of the considerations behind such architectures do not hold for time series applications. The objective of this study is to re-consider the effectiveness of deep learning architectures for efficient and accurate time series forecasting and anomaly detection. A model with direct projections is proposed, and it outperforms existing Transformer based models in most cases by a significant margin. The new decomposition based dual projection (DBDP) model consists of an anchored global profile and a varied number of decomposed seasonal local profiles of the time series for better forecasting performance. In addition to forecasting, a non-contrastive self-supervised learning approach, we propose to include a contrastive learning module in the DBDPC model for better forecasting performance and robustness. Finally, we apply the DBDP and DBDPC models to forecasting based time series anomaly detection and achieve superior performance over the latest SoTA models. These results demonstrate the effectiveness of the several key considerations behind the DBDP and DBDPC models, which also encourages the development of new architectures for time series applications.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning