Time series data augmentation for deep learning: paper reading of a survey

Original title ：Time Series Data Augmentation for Deep Learning: A Survey

Chinese translation ： Time series data enhancement for deep learning : review

Time of publication ：2022 year 3 month 31 Japan

platform ：IJCAI 2022-02

source ：DAMO Academy, Alibaba Group, Bellevue, WA, USA

The article links ：https://arxiv.org/pdf/2002.12478.pdf

Open source code ：

Abstract

        Deep learning has recently performed very well in many time series analysis tasks . The superior performance of deep neural network largely depends on a large number of training data to avoid over fitting . However , Many real-world time series applications may have limited labeled data , Such as the classification of medical time series and AIOps Abnormal detection of . As an effective method to improve the scale and quality of training data , Data enhancement is the key to the successful application of deep learning model to time series data . This paper systematically reviews different time series data enhancement methods . We suggest classifying these methods , Then by emphasizing their advantages and limitations , A structured review of these methods . We also empirically compare different tasks of different data enhancement methods , Including time series classification , Anomaly detection and prediction . Last , We discussed and highlighted five directions for the future , To provide useful research guidance .

6 Conclusion

         With the popularity of deep learning model in time series data , Limited label data requires effective data enhancement methods . This paper summarizes the application of time series data enhancement methods in different tasks . We divide the methods reviewed into basic methods and advanced methods , Summarize the representative methods for each category , And make empirical comparison in typical tasks , And point out the future research direction .

1. introduction

  chart 1: Classification of time series data enhancement technology （taxonomy）

         This paper aims to fill the above gaps , Summarize the application of existing time series data enhancement methods in common tasks , Including time series prediction 、 Anomaly detection 、 Classification, etc. , And provide far sighted future development direction . So , We propose a classification of time series data enhancement methods , Pictured 1 Shown . On the basis of classification , These data enhancement methods are systematically reviewed . Let's start with a simple transformation in the time domain . Then more time series transformations are discussed in the transformed frequency domain and time-frequency domain .

         Except for the transformation of time series in different domains , We also summarized more advanced methods , Including decomposition based methods 、 Model based approach and learning based approach . For a learning based approach , We further divide it into embedded spaces 、 Depth generation model (DGMs) And automated data enhancement methods . To demonstrate the effectiveness of data enhancement , We preliminarily evaluated the enhancement methods in three typical time series tasks , Including time series classification 、 Anomaly detection and prediction . Last , We discussed and highlighted five future directions : enhance （ Time frequency domain 、 Nonequilibrium class 、 Selection and combination 、 Gaussian process and depth generation model ）.

4.2 Time series anomaly detection

         Considering the problem of data scarcity and data imbalance in time series anomaly detection , It is beneficial to use the method of data enhancement to generate more annotation data . We are [Gao wait forsomeone ,2020] The results are briefly summarized in , Based on U-Net The network is designed and made public Yahoo! Data sets [Laptev wait forsomeone ,2015] Used for time series anomaly detection . The performance comparison under different settings is shown in the table 2 Shown , They are the original data (U-Net-Raw)、 Decomposition residuals (U-Net-DeW) And data enhancement residuals (U-Net-DeWA) Apply the model （U-Net） Performance comparison . Applied data enhancement methods include flipping 、 tailoring 、 Label expansion （expansion） And based on APP Frequency domain enhancement . It can be seen that , Decomposition helps F1 The improvement of scores , Data expansion further improves performance .

surface 2: Data enhanced time series anomaly detection precision, recall, and F1 score The promotion of .