VRNN’s got a GAN: Generating Time Series using Variational Recurrent Neural Models with Adversarial Training

TMLR Paper562 Authors

Abstract:

Time-series data generation is a machine learning task growing in popularity, and
has been a focus of deep generative methods. The task is especially important
in fields where large amounts of training data are not available, and in applica-
tions where privacy preservation using synthetic data is preferred. In the past,
generative adversarial models (GANs) were combined with recurrent neural net-
works (RNNs) to produce realistic time-series data. Moreover, RNNs with time-
step variational autoencoders were shown to have the ability to produce diverse
temporal realizations. In this paper, we propose a novel data generating model,
dubbed VRNN-GAN, that employs an adversarial framework with an RNN-based
Variational Autoencoder (VAE) serving as the generator and a bidirectional RNN
serving as the discriminator. The recurrent VAE captures temporal dynamics into
a learned time-varying latent space while the adversarial training encourages the
generation of realistic time-series data. We compared the performance of VRNN-
GAN to state-of-the-art deep generative methods on the task of generating syn-
thetic time-series data. We show that VRNN-GAN achieves the best predictive
score across all methods and yields competitive results in other well-established
performance measures compared to the state-of-the-art.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Antoni_B._Chan1

Submission Number: 562