Introducing Positive-Sum Fairness: A New Way to Balance Performance and Equity in Medical AI

Table of Links

1. Abstract and Introduction

2. Related work

3. Methods

   3.1 Positive-sum fairness

   3.2 Application

4. Experiments

   4.1 Initial results

   4.2 Positive-sum fairness

5. Conclusion and References

3 Methods

3.1 Positive-sum fairness

We introduce the principle of positive-sum fairness, which analyzes fairness from the prism of harmful and non harmful disparities. When looking at changes in model performance and disparities between protected subgroups, there are several explanations for a gap in performance between the most and least advantaged subgroups:

– The most advantaged group’s performance improved while others’ stayed the same,

– All subgroups’ performance improved but one of them increased more than others,

– The most discriminated group’s performance decreased while others’ stayed the same,

– All subgroups’ performance decreased, but one of them decreased more than the others, etc.

The first two would not be considered harmful as they allow to improve the general performance without harming any of the subgroups, thus achieving a collective benefit.

Definition Positive-sum fairness is a fairness evaluation framework where the goal is to find solutions that increase the overall benefit for all parties together while trying to ensure no one is worse off and ideally, everyone is better off. It looks at the situation where we have an initial model and are looking at the trade-off between fairness and performance when trying to improve the model. Unlike other fairness definitions which aim to minimize the disparity between subgroups or maximize the worst performance among subgroups, positive-sum fairness tries to avoid gains to a group which come at the expense of another group while maintaining the overall performance.