Balanced Mixed-Type Tabular Data Synthesis with Diffusion Models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Diffusion models are powerful for generating synthetic tabular data.
• Existing models often inherit and amplify biases present in training data.
• Biased synthetic data can lead to discriminatory outcomes.

Purpose of the Study:

• To develop a novel tabular diffusion model that generates fair synthetic data.
• To mitigate bias by balancing joint distributions of target labels and sensitive attributes.
• To ensure high-quality synthetic data generation while promoting fairness.

Main Methods:

• Introduction of a novel tabular diffusion model incorporating sensitive guidance.
• Balancing joint distributions of target labels and sensitive attributes (e.g., sex, race).
• Empirical evaluation using fairness metrics like demographic parity ratio and equalized odds ratio.

Main Results:

• The proposed method effectively mitigates bias present in the training data.
• High-quality synthetic tabular data generation is maintained.
• Outperformed existing methods on fairness metrics, achieving over 10% improvement.

Conclusions:

• The novel diffusion model successfully generates fair synthetic tabular data.
• The approach balances sensitive attributes and target labels, reducing bias.
• This method offers a significant advancement in fair and high-quality synthetic data generation.