Time-series data augmentation based on interpolation

Abstract

In machine learning, data augmentation is the process of generating synthetic data samples that will be used to train the model to improve the performance of the machine learning model. Data augmentation has been shown to improve the generalization capabilities of models and is particularly popular in the field of computer vision, which is to deal with image data. In contrast, data augmentation is less widely used in the field of time-series data, such as time-series classification, than in computer vision. This is because time-series data is particularly vulnerable to the transformation of data that occurs while performing data augmentation. For example, flipping or rotating images to augment image data does not significantly undermine the meaning of the original data. However, in the case of time-series data, it is likely to distort the meaning of the data, and it is difficult to identify whether or not it has altered. Previously proposed time-series data augmentation methods performed well in many fields, but often did not consider trend information of time-series data such as slicing or reordering the original time-series. In this paper, we propose a time-series data augmentation method based on interpolation. The proposed method is robust against the impairment of trend information of the original time-series and has the advantage of not high complexity. To evaluate the performance of the proposed method, we experimented with time-series datasets from the UCR archive and showed that the performance of the model could be improved. © 2020 The Authors.