Fair Generalized Linear Models with a Convex Penalty.

Area of Science:

• Machine Learning
• Algorithmic Fairness
• Statistical Modeling

Background:

• Generalized linear models (GLMs) are widely used but lack established fairness methodologies.
• Algorithmic fairness research has advanced, yet GLMs remain underexplored in this domain.

Purpose of the Study:

• To introduce and formalize fairness criteria specifically for GLMs.
• To develop an efficient optimization method for achieving fairness in GLMs.

Main Methods:

• Proposed two fairness criteria: equalizing expected outcomes and log-likelihoods.
• Developed a convex penalty term based on GLM linear components for efficient optimization.
• Derived theoretical properties of the fair GLM estimator.

Main Results:

• Demonstrated that fairness criteria can be achieved via convex penalty optimization.
• Empirically validated the proposed fair GLM against existing methods on benchmark datasets.
• Showcased the fair GLM's ability to produce fair predictions for diverse response variables.

Conclusions:

• The proposed fair GLM offers a practical and efficient solution for achieving algorithmic fairness in a widely used statistical framework.
• The methodology extends fairness considerations beyond binary and continuous outcomes.