[Paper Review] Conditional GAN for timeseries generation

Kaleb Smith, Anthony O. Smith|arXiv (Cornell University)|Jun 30, 2020

Time Series Analysis and Forecasting34 references63 citations

TL;DR

TSGAN uses a two-stage Wasserstein GAN (one unconditional, one conditional) to generate realistic 1D time series from spectrogram-like representations, enabling few-shot generation. It shows improved Fréchet Inception Distance and classification-based evaluation over a baseline WGAN on 70 UCR time series datasets.

ABSTRACT

It is abundantly clear that time dependent data is a vital source of information in the world. The challenge has been for applications in machine learning to gain access to a considerable amount of quality data needed for algorithm development and analysis. Modeling synthetic data using a Generative Adversarial Network (GAN) has been at the heart of providing a viable solution. Our work focuses on one dimensional times series and explores the few shot approach, which is the ability of an algorithm to perform well with limited data. This work attempts to ease the frustration by proposing a new architecture, Time Series GAN (TSGAN), to model realistic time series data. We evaluate TSGAN on 70 data sets from a benchmark time series database. Our results demonstrate that TSGAN performs better than the competition both quantitatively using the Frechet Inception Score (FID) metric, and qualitatively when classification is used as the evaluation criteria.

Motivation & Objective

• Motivate synthetic time series generation to address limited real data for machine learning in domains like finance, medicine, and science.
• Propose a novel two-stage architecture (TSGAN) that leverages spectrogram-like representations to guide time series generation.
• Evaluate on a large, diverse 1D time series benchmark to demonstrate few-shot learning capabilities and generation quality.
• Compare against baseline WGAN and analyze both quantitative and qualitative results to validate realism and diversity of generated sequences.

Proposed method

• Use two Wasserstein GANs in sequence: WGAN 1 generates a 2D spectrogram image from a random latent vector z; WGAN 2 is a conditional GAN that takes the spectrogram image to produce a 1D time series.
• Adopt WGAN-GP (gradient penalty) to enforce the 1-Lipschitz constraint and stabilize training.
• Train WGAN 1 and WGAN 2 independently, with WGAN 2 conditioned on the output of WGAN 1 to enable few-shot learning and capture frequency-domain information.
• Represent time series generation as transforming latent space into spectrograms and then decoding to time-domain signals, leveraging encoder-like structure in the second generator.
• Use fully convolutional networks (FCN) and a simple 1D/2D conditioning scheme to measure similarity between real and generated samples via adapted 1D Fréchet Inception Distance (FID), plus TSTR/TRTS classification-based metrics.

Experimental results

Research questions

• RQ1 Can a two-stage GAN architecture that conditions time-domain generation on frequency-domain representations improve realism and diversity of 1D time series with limited training data?
• RQ2 Does conditioning a second GAN on spectrogram-like outputs from the first GAN yield better time-series synthesis than a single GAN approach?
• RQ3 How does the proposed method perform across a large, diverse set of univariate time series with few-shot training scenarios in terms of approximate data realism and downstream classification performance?

Key findings

• TSGAN outperforms a baseline WGAN in 1D FID scores on several datasets (e.g., Beef: 207.1 for TSGAN vs 638.8 for WGAN).
• On the Gun Point dataset, TSGAN achieves a FID of 4.778 whereas WGAN scores 39.95, indicating notable improvement in generated signal realism.
• TSGAN generally yields lower FID scores and visually sharper, more accurate time-series samples than WGAN across multiple datasets, especially with small and medium training sizes.
• In terms of classification-based evaluation (TRTS and TSTR), TSGAN outperforms WGAN on a majority of datasets, with several datasets showing clear gains in downstream FCN accuracy when trained on synthetic data.
• TSGAN demonstrates strong few-shot learning capabilities, delivering better results than WGAN on many datasets with limited training samples (small and medium subsets dominate improvements).
• The conditioning mechanism, where the second WGAN decodes from spectrogram-like representations produced by the first WGAN, appears to provide richer latent information and improves temporal and frequency characteristics of generated signals.