A refined reweighing technique for nondiscriminatory classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning systems can perpetuate and exacerbate socioeconomic disparities.
• Existing discrimination-aware classification methods often focus solely on sensitive attributes.

Purpose of the Study:

• To propose a novel data pre-processing technique for reducing discrimination in machine learning.
• To refine instance weighting by incorporating both sensitive and insensitive attributes.

Main Methods:

• Developed a data pre-processing technique assigning weights to training instances.
• Formulated weight assignment as a linear programming problem.
• Incorporated both sensitive and insensitive attributes for weight refinement.

Main Results:

• Achieved significant discrimination reduction with a minimal impact on classification accuracy.
• Demonstrated superior scalability compared to existing pre-processing methods.
• Enabled explicit monitoring of the trade-off between fairness and accuracy.

Conclusions:

• The refined reweighing method effectively reduces discrimination without altering input data or labels.
• This approach offers a scalable and transparent solution for enhancing fairness in machine learning models.
• The method provides users with control over the fairness-accuracy balance.