Bipartite Ranking Fairness Through a Model Agnostic Ordering Adjustment.

Area of Science:

• Machine Learning
• Algorithmic Fairness
• Data Science

Background:

• Algorithmic fairness is a growing concern in machine learning, particularly in ranking scenarios.
• Learned ranking functions can exhibit systematic disparities across different protected groups.
• Existing methods may face a trade-off between fairness and classification performance.

Purpose of the Study:

• To propose a model-agnostic post-processing framework, xOrder, for achieving fairness in bipartite ranking.
• To maintain algorithm classification performance while ensuring fairness.
• To address fairness concerns across binary and multiple protected groups.

Main Methods:

• xOrder optimizes a weighted sum of utility by identifying an optimal warping path across protected groups.
• The optimization is solved using a dynamic programming process.
• The framework is compatible with various classification models and fairness metrics (supervised and unsupervised).

Main Results:

• xOrder consistently achieves a better balance between algorithm utility and ranking fairness across diverse datasets and metrics.
• Visualizations show xOrder mitigates score distribution shifts between groups compared to baselines.
• Analytical results confirm robust performance with limited samples and training-testing distribution differences.

Conclusions:

• xOrder effectively enhances fairness in bipartite ranking without sacrificing performance.
• The framework offers a flexible and robust solution for mitigating algorithmic disparities.
• xOrder demonstrates practical applicability on benchmark and real-world electronic health record data.