Loss-Aversively Fair Classification

People judge fairness based on change, not just outcomes: The paper highlights that how an algorithmic decision affects someone compared to the current system matters as much as the final decision itself. This means replacing an existing system requires considering whether people experience improvements or losses, not just whether the new system is fair in isolation.

Loss aversion from behavioral economics applies to AI fairness: Drawing on prospect theory, the authors argue that people are more sensitive to losses than equivalent gains. When deploying new algorithms, this suggests decision-makers should prioritize avoiding negative changes for individuals, even if the overall system becomes more fair.

Fairness updates can be built into algorithm training: The paper proposes practical methods to incorporate "loss-averse updates" into machine learning models, meaning developers can train classifiers to improve outcomes for people compared to the status quo while also reducing discrimination based on sensitive features like race or gender.

The approach works across different types of algorithms: The authors demonstrate their fairness measures can be combined with both simple linear classifiers and more complex non-linear models, suggesting the approach has broad applicability across different algorithmic decision-making systems.

Fairness requires considering transition impacts: The research suggests that responsible AI deployment should account for how algorithmic changes affect real people during the transition from old to new systems, not just evaluate the new system's fairness in a vacuum.