Interventional Fairness with Indirect Knowledge of Unobserved Protected Attributes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning (ML) systems deployed in societal applications require fairness for marginalized groups.
• Protected attributes are often absent from training data, but proxy attributes can still introduce unfairness.
• Feedback mechanisms allow users to flag samples, providing indirect knowledge about fairness issues.

Purpose of the Study:

• To develop a framework for identifying proxy attributes in ML systems when the protected attribute is unknown.
• To leverage flagged samples as indirect knowledge within a causal interventional fairness paradigm.
• To ensure fairness in ML models despite the absence of explicit protected attribute data.

Main Methods:

• Utilizing a feedback-based framework where flagged samples guide the identification of proxy attributes.
• Applying causal inference techniques to determine causal dependencies between proxy attributes and the unknown protected attribute.
• Performing conditional independence tests on observed data to identify relevant proxy attributes without prior knowledge of the structural causal model.

Main Results:

• An algorithm is proposed that effectively identifies proxy attributes causally linked to unknown protected attributes.
• Theoretical proofs establish the optimality of the developed algorithm.
• Empirical evaluations on diverse datasets demonstrate the algorithm's efficacy in promoting ML fairness.

Conclusions:

• The proposed approach offers a robust method for detecting and mitigating unfairness in ML systems with missing protected attributes.
• This work advances the field of causal interventional fairness by providing practical tools for real-world applications.
• The framework enables the development of more equitable and trustworthy AI systems by addressing hidden biases.