Beyond Noise: Incorporating Pre-Trained Contractive Autoencoders for Enhanced GAN-based Tabular Data Creation

Generative Adversarial Networks (GANs) excel at creating imagery data, but despite their success with tabular data, they struggle to accurately reproduce complex interactions, especially between categorical and numerical columns. This paper highlights how the noise quality in GANs is essential for enhancing the data it generates and stabilizing and speeding up convergence. We introduce an approach integrating a pre-trained Contractive Autoencoder (CAE) for generating semantically-consistent noise. Leveraging the CTGAN model, our methodology uses transfer learning from the CAE to optimize the generator input noise. Experiments on three benchmark datasets show our enhanced architecture outperforms traditional approach using random noise, as confirmed by several statistical metrics. Moreover, our synthetic data reveals improved outcomes in machine learning prediction tasks and enhanced inter-column semantic relationships.