Metrics and Algorithms for Locally Fair and Accurate Classifications using Ensembles.

Area of Science:

• Machine Learning and Artificial Intelligence
• Algorithmic Fairness and Bias Mitigation

Background:

• Dynamic model ensembles enhance classifier prediction accuracy by selecting models based on data similarity.
• Existing dynamic model ensembles and global fair model ensembles can suffer from local unfairness, persisting bias in specific data regions.
• Addressing local bias is crucial for equitable treatment across diverse population groups.

Purpose of the Study:

• To develop a framework for creating dynamic model ensembles that are both locally fair and accurate.
• To optimize for equal opportunity for similar subjects within machine learning models.
• To bridge the gap between dynamic model ensembles and global fair model ensembles.

Main Methods:

• Proposed a general framework for devising locally fair and accurate dynamic model ensembles.
• Developed several algorithms to implement the framework's components.
• Introduced a runtime-efficient adaptation of the framework to maintain result quality.

Main Results:

• The proposed framework and algorithms demonstrated superior performance in mitigating local and global bias compared to state-of-the-art methods.
• The approach achieved comparable accuracy to existing methods while significantly improving fairness metrics.
• Evaluation showed effectiveness across various types and degrees of bias in training data.

Conclusions:

• The developed framework successfully achieves locally fair and accurate dynamic model ensembles, optimizing for equal opportunity.
• The runtime-efficient adaptation ensures practical applicability without compromising fairness or accuracy.
• The study presents novel fairness metrics and data preparation insights, advancing the field of algorithmic fairness.