Measuring Model-Induced Discrimination via Efficient Fairness Approximation

Area of Science:

• Artificial Intelligence
• Machine Learning Ethics
• Algorithmic Fairness

Background:

• Machine learning (ML) models are increasingly used in high-stakes domains, raising concerns about hidden discrimination.
• Existing fairness measures (group and individual) often lack compatibility, complicating bias assessment.
• The incompatibility of current fairness metrics hinders comprehensive discrimination evaluation in ML.

Purpose of the Study:

• To develop a novel fairness measure for ML models that reconciles individual and group fairness aspects.
• To address the computational challenges in calculating fairness metrics by proposing an efficient approximation algorithm.
• To evaluate the proposed methods' effectiveness in reflecting and quantifying discrimination in ML classifiers.

Main Methods:

• Investigated ML classifier discrimination using a manifold perspective.
• Proposed a new fairness measure: harmonic fairness via manifolds (HFM), based on distances between sets.
• Devised an approximation algorithm, ApproxDist, to efficiently estimate distances between sets for practical fairness assessment.

Main Results:

• The proposed HFM measure effectively captures bias from both individual and group fairness perspectives.
• The ApproxDist algorithm demonstrates effectiveness and efficiency in approximating set distances.
• Empirical results validate the practical applicability and accuracy of the proposed fairness assessment methods.

Conclusions:

• Harmonic fairness via manifolds (HFM) offers a unified approach to measuring ML model bias.
• The approximation algorithm (ApproxDist) significantly enhances the practical utility of manifold-based fairness measures.
• The developed methods provide a more robust and efficient framework for assessing and mitigating discrimination in machine learning.