Learning Counterfactual Fair Representation Under Covariate Shift via Reflux

Area of Science:

• Artificial Intelligence
• Machine Learning
• Fairness in AI

Background:

• Existing fairness methods often rely on group metrics or in-processing techniques.
• Data preprocessing for bias mitigation and handling real-world distribution shifts remain underexplored.
• Generalizability of fair models across domains is often compromised by distribution shifts.

Purpose of the Study:

• To propose a novel framework for counterfactual fair representation learning.
• To address limitations of existing methods by incorporating data preprocessing and handling covariate shift.
• To enable both single-domain and covariate shift prediction tasks with enhanced fairness and transferability.

Main Methods:

• Developed the Counterfactual Reflux Variational Autoencoder (CRVAE) for generating counterfactual samples and learning fair representations.
• Introduced a Reflux technique to enforce consistency between factual and counterfactual representations for fairness.
• Incorporated a domain discriminator to align fair representations across domains, enhancing transferability.

Main Results:

• CRVAE improves fairness with minimal impact on model performance.
• The framework demonstrates effective generalization across different domains, maintaining performance under distribution shifts.
• Experimental results validate the approach's ability to learn fair and transferable representations.

Conclusions:

• CRVAE offers a novel data preprocessing approach for counterfactual fairness, addressing individual fairness and covariate shift.
• The method enhances model generalizability and can be combined with existing in-processing fairness techniques.
• This work represents a significant step towards robust and transferable fair machine learning models.