Improving Fair Training under Correlation Shifts

Real-world fairness challenges differ from training assumptions: The paper identifies a critical gap in existing fairness techniques—most assume training and deployment data are identical, but in practice the relationship between labels and sensitive groups (like race or gender) often changes over time, causing previously fair models to become unfair.

Existing fairness methods have fundamental limits: The researchers analytically demonstrate that current in-processing fairness algorithms (techniques applied during model training) cannot simultaneously achieve both high accuracy and group fairness when correlation shifts occur, establishing theoretical boundaries to what these approaches alone can accomplish.

A two-step solution improves fairness and accuracy: The paper proposes combining data preprocessing (adjusting which training examples are used) with existing fairness techniques, showing this combination helps overcome the limitations of fairness algorithms used alone and improves performance on both synthetic and real-world datasets.

Correlation shifts are a measurable problem: By formally defining "correlation shifts"—the change in how bias between labels and sensitive groups evolves—the paper provides a concrete way to understand and address fairness degradation, moving beyond vague notions of "distribution change."

Decoupling fairness strategies is more effective: Rather than trying to solve all fairness problems in one step, separating the roles of data adjustment (preprocessing) and fairness mitigation (in-processing) creates a more modular and effective framework for building trustworthy AI systems.