Fair Representation: Guaranteeing Approximate Multiple Group Fairness for Unknown Tasks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Diverse prediction tasks necessitate fair data representations.
• Existing fairness impossibility results highlight challenges in achieving universal fairness.
• Multiple group fairness notions (independence, separation, calibration) require consideration.

Purpose of the Study:

• Investigate fair representation for guaranteeing fairness across unknown tasks and multiple fairness notions.
• Explore approximate fairness in light of theoretical impossibility results.
• Develop methods for learning representations that are both fair and discriminative.

Main Methods:

• Theoretical analysis of fair representation properties for prediction tasks.
• Proposing a pretext loss for self-supervised learning of representations.
• Utilizing Maximum Mean Discrepancy (MMD) as a fairness regularizer.
• Empirical validation on tabular, image, and face datasets.

Main Results:

• Fair representation guarantees fairness for discriminative tasks, where fairness and discriminativeness are linearly controlled.
• Fair and discriminative representations achieve approximate fairness for incompatible fairness notions.
• Learned representations improve fairness in downstream, previously unknown prediction tasks.
• Theoretical fairness guarantees are empirically validated.

Conclusions:

• Fair representation learning is a viable strategy for achieving approximate fairness across diverse and unknown prediction tasks.
• The proposed method effectively balances multiple fairness notions, even when they are in conflict.
• Empirical results confirm the practical utility of fair and discriminative representations.