Task-tailored Pre-processing: Fair Downstream Supervised Learning

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Fairness-aware machine learning aims to mitigate discrimination in data-driven tasks.
• Existing pre-processing methods fall into data fairness (group-independent) and task-tailored fairness (group-dependent).
• Data fairness methods may impose overly strong regularization, impacting utility.

Purpose of the Study:

• To develop a novel pre-processing approach for supervised learning that is tailored to the specific task.
• To theoretically investigate the trade-offs between fairness and utility.
• To provide downstream guarantees for fairness improvement and utility preservation.

Main Methods:

• Devised a novel pre-processing approach explicitly considering the supervised learning task.
• Analyzed the trade-off between fairness and utility in the pre-processing map.
• Studied downstream model behavior on transformed data to establish fairness and utility guarantees.
• Evaluated the framework on tabular and image datasets.

Main Results:

• The proposed task-tailored fairness approach outperforms existing methods in preserving fairness-utility trade-offs.
• Demonstrated superiority on both tabular and image datasets.
• Showed that the method alters only necessary semantic features for fairness, particularly in computer vision tasks.
• Provided theoretical guarantees for downstream model fairness and utility.

Conclusions:

• The novel task-tailored pre-processing framework offers a superior approach to achieving algorithmic fairness in supervised learning.
• The method effectively balances fairness and utility, outperforming existing techniques.
• Theoretical guarantees and empirical results validate the framework's effectiveness across diverse datasets.