Deterministic Autoencoder using Wasserstein loss for tabular data generation.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Tabular data synthesis is challenging due to data complexity.
• Variational Autoencoders (VAEs) adapted for tabular data face limitations like unstable latent spaces and constrained interpolation.
• These limitations hinder effective synthetic data generation and control.

Purpose of the Study:

• To introduce the Tabular Wasserstein Autoencoder (TWAE) as a novel deep learning approach for tabular data synthesis.
• To address the limitations of VAEs in tabular data generation by employing a deterministic latent space.
• To enable stable latent space interpolation for generating high-quality synthetic tabular data.

Main Methods:

• Developed Tabular Wasserstein Autoencoder (TWAE) utilizing Wasserstein Autoencoders' deterministic encoding.
• Integrated TWAE with shallow interpolation techniques, such as Synthetic Minority Over-sampling Technique (SMOTE), for data generation.
• Trained TWAE once to create a low-dimensional data representation, followed by latent space interpolation for synthetic point generation.

Main Results:

• TWAE demonstrates superior performance in tabular data synthesis compared to existing methods.
• The model exhibits versatility across various feature types and dataset sizes.
• Achieved a balance between accuracy and efficiency in generating synthetic tabular data.

Conclusions:

• TWAE offers a robust and stable solution for complex tabular data synthesis.
• The deterministic latent space of TWAE enhances control and expressiveness.
• Combining WAE principles with SMOTE provides an effective deep learning framework for generating high-quality synthetic tabular data.