Special TAPIR Seminar

Real-world datasets are often imbalance which is a problem for training conventional machine learning algorithms. To address the imbalance problem, many data augmentation techniques have been proposed for image recognition tasks, but only a few have been developed for time series. In this talk, I will describe a conditional Wasserstein GAN. Our model can learn the implicit probability distribution of a dataset conditioned on the irregular sampling times, amplitudes and class of the time series and generate a variety of realistic samples to complement the original dataset. We trained and evaluated our model using a pair of toy datasets and a real-world astronomical survey. We then generated realistic samples to augment the original datasets and compared the performance of a classifier trained on the GAN augmented datasets against oversampled datasets and noise-augmented datasets. The resulting generator can be used as any other generative model, allowing interpolations and extrapolations in the parameter space. [NOTE: Unusual venue]