Ensuring generalized fairness in batch classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Algorithmic Fairness

Background:

• Batch classification selects groups, unlike individual classification, with unique fairness needs.
• Existing fairness methods fail when different acceptance rates per group are required.
• Sensitive attributes like gender or race necessitate group-specific fairness considerations.

Purpose of the Study:

• To propose a novel framework for regulated fairness in batch classification.
• To address limitations of existing methods in scenarios requiring differential group acceptance rates.
• To introduce a flexible and efficient post-processing approach for fairness.

Main Methods:

• Developed a configuration model to regulate group acceptance rates.
• Introduced a batch-wise fairness post-processing framework using classifier confidence scores.
• Tested the framework on four real-world datasets with demographic parity and equalized odds.

Main Results:

• Achieved consistent performance improvements over baseline methods.
• Demonstrated flexibility in handling multiple overlapping sensitive attributes.
• Showcased significant speed-up compared to existing approaches.
• Applied successfully to fair gerrymandering, improving the fairness-accuracy trade-off.

Conclusions:

• The proposed framework offers a novel and effective solution for fairness in batch classification.
• It provides regulatory flexibility and computational efficiency, outperforming existing methods.
• The framework's generalizability is shown through applications beyond standard classification tasks.