Fair classification via domain adaptation: A dual adversarial learning approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning (ML) models are increasingly used in decision-making systems.
• ML models face challenges with discrimination and unfairness, limiting their use in critical applications.
• Existing fair ML models often require sensitive attributes, which are frequently unavailable or restricted.

Purpose of the Study:

• To address the challenge of unfairness in ML when sensitive attributes are missing.
• To explore domain adaptation techniques for achieving fair classification.
• To develop a framework that leverages auxiliary information from similar domains to improve fairness.

Main Methods:

• Propose a novel framework for domain adaptation in fair classification.
• Develop methods to learn and adapt sensitive attributes from a source domain to a target domain.
• Utilize extensive experiments on real-world datasets to validate the approach.

Main Results:

• The proposed framework effectively achieves fair classification without requiring sensitive attributes in the target domain.
• Demonstrated the ability to adapt sensitive information across domains for improved fairness.
• Experimental results confirm the model's effectiveness on real-world data.

Conclusions:

• Domain adaptation is a viable strategy for achieving fair ML when sensitive attributes are unavailable.
• The proposed framework offers a practical solution for deploying fair ML in sensitive applications.
• This research advances the field of fair machine learning by enabling fairness without direct access to sensitive data.