Time Series Data Augmentation using Time-Warped Auto-Encoders

The recent eminent success of deep learning models has triggered a lot of interest in data augmentation studies. The main challenge of data-hungry models is that they are bound to overfitting and require massive amounts of data to be trained on to achieve optimal levels of convergence. It is of high importance for any data oversampling algorithm to produce synthetic data samples that are not only stemming from the same real data distribution, but also have enough variation to provide more learning opportunities to trained models. Early efforts on time series data augmentation either rely on generating new samples by interpolating between two close real data neighbors or using Generative Adversarial Networks. In this paper, we present Time Series Generation based on Auto-encoders (TiGA), a novel algorithm for time series generation using time-warped autoencoders. Our idea is to exploit the lossy transformation of autoencoders for the purpose of generating synthetic time series data samples. To the best of our knowledge, this is the first effort that leverages the latent features generated by autoencoders for the purpose of time series data oversampling. We evaluate our proposed approach on an open-source real-life solar flare prediction dataset. Results show that TiGA produces samples that are both quantitatively and qualitatively superior to current state-of-the-art methods.