Class Conditional Time Series Generation with Structured Noise Space GAN

Hamidreza Gholamrezaei, Alireza Koochali, Andreas Dengel, Sheraz AhmedarXiv 2023 · DOI · arXiv

Abstract

Generative Adversarial Networks (GANs) often struggle with class-conditional generation, particularly in the time series domain where standard techniques like conditional batch normalization can disrupt temporal correlations. We present Structured Noise Space GAN (SNS-GAN), an architecture-agnostic approach that embeds class information directly into the input noise space. By shifting the means of specific dimensions in the noise vector, SNS-GAN allows the generator to learn class-specific mappings without internal architectural modifications. Our results demonstrate that SNS-GAN achieves state-of-the-art performance on both image benchmarks and complex time series datasets from the UCR archive.

Introduction

Generative Adversarial Networks (GANs) have revolutionized synthetic data generation, yet conditioning these models to produce specific classes remains a significant challenge. Traditional Conditional GANs (CGANs) typically rely on concatenating labels to noise vectors or using Conditional Batch Normalization. However, these methods often fail in sensitive domains like time series analysis, where architectural modifications can inadvertently destroy delicate temporal dependencies. This paper introduces Structured Noise Space GAN (SNS-GAN), a "plug-and-play" method that shifts the conditioning logic from the network architecture to the input distribution itself.

The Challenge of Conditional Generation

Standard conditioning techniques face two primary hurdles:

  1. Label Neglect: In simple concatenation-based CGANs, generators often learn to ignore the label input if the noise vector dominates the signal.
  2. Temporal Disruption: Conditional Batch Normalization, while effective for images, is often unsuitable for time series. It can normalize away the very temporal correlations that define the data's authenticity.

SNS-GAN addresses these issues by being architecture-agnostic. It requires no changes to the internal wiring of the generator, making it inherently safe for time series data.

Methodology: Structured Noise Space

The core innovation of SNS-GAN is the Structured Noise Space. Instead of sampling from a standard unimodal Gaussian distribution, the noise vector $z_c$ is transformed into a multimodal distribution. For a dataset with $N$ classes, an $N$-dimensional noise space is used. To generate a sample for class $c$, the noise is reparameterized:

$$z_c = z + \text{one\_hot\_encode}(c)$$

By shifting the mean of the dimension corresponding to the target class, the noise vector itself carries the class information. The generator then learns to map these specific regions of the noise space to their respective data classes.

Overview of the SNS-GAN generation pipeline, illustrating how class conditions are integrated into the noise space via reparameterization.

Architecture and Implementation

The SNS-GAN framework can be applied to various generator architectures. In the image domain, the model utilizes fully connected (FC) layers followed by transposed convolutional (ConvT) blocks. For time series, the researchers explored Linear, RNN (GRU-based), and TCN (Temporal Convolutional Network) variants. Notably, the SNS-GAN-Linear variant—the simplest architecture—performed best for time series, as it could capture global temporal dependencies across all time steps simultaneously without the drift issues common in recurrent models.

Detailed architecture of the SNS-GAN generator and discriminator, showing the flow from structured noise $Z_c$ to synthetic data and the discriminator's evaluation.

Experimental Results

SNS-GAN was validated across both image and time series domains:

  • Image Generation: On the CIFAR-10 dataset, SNS-GAN achieved a Fréchet Inception Distance (FID) of 14.46, outperforming established baselines like SN-GAN (18.8) and WGAN-GP (21.5).
  • Time Series Generation: Using the UCR benchmark, SNS-GAN-Linear consistently outperformed the Recurrent Conditional GAN (RCGAN) baseline. On the complex "Fifty Words" dataset (50 classes), it achieved a Fréchet Inception Time Distance (FITD) of 1.65, significantly better than the best RCGAN variant (2.51).

Conclusion

SNS-GAN provides a simple yet powerful alternative to traditional conditional GANs. By structuring the input noise space, it enables effective class-conditional generation while preserving the structural integrity of the data. This approach is particularly valuable for time series generation, where maintaining temporal correlations is paramount. Future work may extend this "structured noise" philosophy to other complex modalities like audio and music synthesis.

Summary updated

Abstract

This paper introduces Structured Noise Space GAN (SNS-GAN), a novel approach in the field of generative modeling specifically tailored for class-conditional generation in both image and time series data. It addresses the challenge of effectively integrating class labels into generative models without requiring structural modifications to the network. The SNS-GAN method embeds class conditions within the generator's noise space, simplifying the training process and enhancing model versatility. The model's efficacy is demonstrated through qualitative validations in the image domain and superior performance in time series generation compared to baseline models. This research opens new avenues for the application of GANs in various domains, including but not limited to time series and image data generation.
BibTeX
@misc{gholamrezaei2023class,
  author = {Hamidreza Gholamrezaei and Alireza Koochali and Andreas Dengel and Sheraz Ahmed},
  title = {{Class Conditional Time Series Generation with Structured Noise Space GAN}},
  year = {2023},
  eprint = {2312.12946},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2312.12946},
}
Figure