Equalised Odds is not Equal Individual Odds: Post-processing for Group and Individual Fairness

Overview

• Achieving group fairness and individual fairness in machine learning models can be incompatible when the models use discontinuous probability functions.
• Discontinuous probability functions can randomly assign different outcomes to similar individuals within the same protected group, violating individual fairness.
• Assigning unique odds to each protected subpopulation can also prevent members of one subpopulation from ever receiving equal chances of a positive outcome compared to another, which is another type of unfairness called "individual odds."
• The paper proposes a solution that uses continuous probability functions constrained by their Lipschitz constant to preserve the model's predictive power, individual fairness, and robustness while ensuring group fairness.

Plain English Explanation

Machine learning models are often designed to be "fair" - that is, to treat different groups of people equally. There are two main types of fairness that are often considered: group fairness and individual fairness.

Group fairness means that the model's predictions are equally distributed across different protected subgroups (e.g., race, gender). Individual fairness means that the model treats similar individuals alike, regardless of which subgroup they belong to.

However, the paper explains that these two fairness objectives can be incompatible when the model uses certain types of probability functions. Specifically, if the model uses "discontinuous" probability functions, it can randomly assign different outcomes to similar individuals within the same protected group. This violates individual fairness.

The paper also notes that assigning unique odds to each protected subpopulation can prevent members of one subpopulation from ever having an equal chance of a positive outcome compared to another subpopulation. This is another type of unfairness called "individual odds."

To address these issues, the paper proposes a solution that uses "continuous" probability functions that are constrained by their Lipschitz constant. This approach preserves the model's predictive power, ensures individual fairness, and maintains robustness while also achieving group fairness.

Technical Explanation

The paper examines the tension between group fairness and individual fairness in machine learning models. Group fairness is achieved by equalizing prediction distributions between protected subgroups, while individual fairness requires treating similar individuals alike.

The authors show that these objectives can be incompatible when the scoring model is calibrated through discontinuous probability functions. In this case, individuals can be randomly assigned an outcome determined by a fixed probability, leading to disparate classification odds for similar individuals within the same protected group - a violation of individual fairness.

Additionally, assigning unique odds to each protected subpopulation can prevent members of one subpopulation from ever receiving equal chances of a positive outcome compared to another, which the authors refer to as "individual odds" unfairness.

To address these issues, the paper proposes constructing continuous probability functions between group thresholds that are constrained by their Lipschitz constant. This solution preserves the model's predictive power, individual fairness, and robustness while ensuring group fairness.

Critical Analysis

The paper identifies an important challenge in achieving both group fairness and individual fairness in machine learning models, especially when using discontinuous probability functions. The proposed solution of using continuous probability functions constrained by the Lipschitz constant seems promising, as it aims to balance the competing objectives of fairness.

However, the paper does not provide a comprehensive evaluation of the approach or compare it to other fairness-aware machine learning techniques. It would be useful to see how the proposed method performs relative to other state-of-the-art fairness interventions, both in terms of fairness metrics and model performance.

Additionally, the paper does not discuss the computational complexity or training time of the proposed solution, which could be important practical considerations. It would be valuable to understand the trade-offs involved in implementing this approach.

Finally, the paper could benefit from a more in-depth discussion of the potential limitations and areas for future research. For example, the approach may be sensitive to the choice of the Lipschitz constant, and it's unclear how it would scale to high-dimensional or highly complex data distributions.

Conclusion

This paper tackles the challenging problem of reconciling group fairness and individual fairness in machine learning models. By constructing continuous probability functions constrained by the Lipschitz constant, the proposed solution aims to preserve the model's predictive power, individual fairness, and robustness while ensuring group fairness.

The key insight is that discontinuous probability functions can lead to violations of individual fairness, even when group fairness is achieved. The paper's approach addresses this issue, offering a potential path forward for developing fair and effective machine learning models.

While the paper does not provide a comprehensive evaluation or discuss all potential limitations, it highlights an important problem and presents a promising direction for further research and development in the field of algorithmic fairness.