Few-Shot Learning for Time Series Data Generation Based on Distribution Calibration

Abstract

Insufficient training data often makes the learning model prone to overfitting and bias in the selection of the sample leads to obtaining the wrong distribution. For this reason, few-shot learning has gained widespread attention as a challenging endeavor. Current work in few-shot learning is focused on developing stronger models, but these models does not have good generalization capabilities. In this paper, Our approach is find a similar base class with sufficient data for class with few-shot samples, then use statistical information to calibrate the distribution of class with few-shot samples. Time series are characterized by variability within the variance at each point in time and by overall statistical regularity and periodicity. So time series are extremely suitable for our approach. This approach do not require complex models and additional parameters. Our approach generate data that better match the actual distribution of the data. Validated with 9 time series data sets, the data generation for five samples led to some improvement in the classification accuracy. Moreover, it is found that this approach is not only applicable to the case of small data size, but also the classification effect is improved if the method of this paper is applied on the basis of sufficient data size.