Data augmentation via diffusion model to enhance AI fairness.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• AI fairness aims to ensure AI systems are transparent, explainable, and serve user interests.
• Data augmentation using synthetic data generation is a key strategy for addressing data scarcity.
• Diffusion models are advanced generative techniques, particularly effective in computer vision and increasingly explored for other data types.

Purpose of the Study:

• To investigate the efficacy of diffusion models in generating synthetic tabular data for improving AI fairness.
• To evaluate the Tabular Denoising Diffusion Probabilistic Model (Tab-DDPM) for tabular data augmentation in the context of fairness.

Main Methods:

• Utilized the Tabular Denoising Diffusion Probabilistic Model (Tab-DDPM) to generate synthetic tabular data.
• Applied data augmentation with varying amounts of generated data.
• Employed sample reweighting from AIF360 to further enhance fairness.
• Validated the approach using five traditional machine learning models: Decision Tree, Gaussian Naive Bayes, K-Nearest Neighbors, Logistic Regression, and Random Forest.

Main Results:

• The synthetic data generated by Tab-DDPM demonstrably improved fairness metrics in binary classification tasks.
• The effectiveness of Tab-DDPM was validated across multiple traditional machine learning algorithms.

Conclusions:

• Diffusion models, exemplified by Tab-DDPM, show significant potential for enhancing AI fairness through synthetic tabular data generation.
• This approach offers a viable method for improving the ethical performance of AI systems dealing with tabular data.