Estimating and Controlling for Equalized Odds via Sensitive Attribute Predictors

(EOD) definition of fairness. In a setting without sensitive attributes, we first provide tight and computable upper bounds for the EOD violation of a predictor. These bounds precisely reflect the worst possible EOD violation. Second, we demonstrate how one can provably control the worst-case EOD by a new post-processing correction method. Our results characterize when directly controlling for EOD with respect to the predicted sensitive attributes is - and when is not - optimal when it comes to controlling worst-case EOD. Our results hold under assumptions that are milder than previous works, and we illustrate these results with experiments on synthetic and real datasets.

Paper

References (44)

Scroll for more · 32 remaining

Similar papers

Peer review

Reviewer ntnH7/10 · confidence 4/52023-06-09

Summary

This work provides tight and computable bounds on the EOD violation of a classifier in a setting where the sensitive attributes are unknown. In addition, a post-processing technique is proposed to provably yield classifiers that maximize prediction power while achieving minimal worst-case EOD violations with respect to unobserved sensitive attributes. Experiments are done on both synthetic and real data.

Strengths

Significance: The problem under study is interesting and important. It would be useful to be able reduce the upper bound of certain fairness violation in models when sensitive attributes are unknown. Originality and quality: the proposed solution is novel and technically sound. Clarity: the presentation is clear with limitations analyzed in details.

Weaknesses

1. One assumption of the proposed approach is that there exists a dataset of (X,A) to learn $\hat{A} = h(X)$ from. This could limit the use of the approach and should be discussed. 2. It would be beneficial to demonstrate the verification of Assumption 1 on the two datasets.

Questions

Q1: can you elaborate on the potential bad outcome of applying $f_{opt}$ yields a less accurate and more biased (in terms of $\Delta_{FPR}$) model as in Figure 3? Q2: Is that possible to use a different dataset of similar features (but not the same features as in the experiments) to learn $\hat{A} = h(X)$ (in some form of transfer learning)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Results on the CheXpert data shows that, although the proposed method successfully reduces the worst case scenario TPR and FPR, it may actually increase the true TPR or FPR. Figure 3b shows an increase of $\Delta_{FPR}$ of $f_{opt}$ over $f$. This should be discussed as the potential outcome of applying $f_{opt}$ can be a less accurate and more biased model.

Reviewer RpWQ7/10 · confidence 4/52023-07-01

Summary

The paper explores the fairness problem in machine learning when sensitive attributes (e.g., demographics and gender) are unavailable for practitioners. The authors focus on the Equalized-Odds (EOD), a well-known definition of fairness for classifiers. The authors provide bounds for EOD violation in a setting where sensitive attributes are unavailable. Additionally, they propose a post-processing method to control the worst-case EOD. Finally, the authors illustrate their results via synthetic and real-world datasets experiments.

Strengths

* The paper is very well written. Particularly the related work section is very informative. * The authors explore an interesting problem for the community and propose a different view of the problem, i.e., a worst-case optimization. * The theoretical guarantees provide a method for practitioners to access the quality of the method (assuming a worst-case scenario). * The proposed method is simple, intuitive, and computationally inexpensive.

Weaknesses

* The evaluation in the main paper is limited. The authors only apply their method to one real-world example. I suggest adding at least two more examples and as many examples as possible in the appendix. These examples can convince us that: (i) Assumption 1 is followed in multiple scenarios and (ii) your method helps to improve fairness. * The fact that the Naïve approach had a better "true" EOD is interesting, and it might indicate that your method is not very efficient, i.e., using the sharp upper bound as a proxy for the true value is too pessimistic. Performing more experiments in real-world datasets can help the authors and readers understand when minimizing the upper bound is better than the naïve approach. * It needs to be clarified how to extend the approach to a setting where the sensitive attribute is not binary. In the introduction, the authors list examples of sensitive attributes as "demographics and gender," which are not binary. A subsection is recommended to discuss how your results would change when $A$ is not binary.

Questions

* Is it possible to extend this worst-case analysis to other post-processing methods? Combining it with the optimized method proposed in [1] would be very interesting. * In Theorem 2 the authors assume that $\mathcal{F}$ is such that $ \forall f \in \mathcal{F}$ the Assumption 1 is followed. After that, they conclude that then, there exists a fair model in $\mathcal{F}$. This doesn’t seem correct. For example, take $\mathcal{F} = \{f_0\}$ where $f_0$ follows Assumption 1 but is not worst-case optimal. Then, there is no fair model in $\mathcal{F}$. Perhaps, the authors mean $\mathcal{F}$ to be the set of *all* models that follow Assumption 1. However, in this case, assuming knowledge of $\mathcal{F}$ does not sound reasonable. Could the authors clarify this point? * Could you describe (in the main paper) when the worst-case bound is achieved?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors discuss their work limitations. I also suggest discussing the ethics of predicting users' sensitive attributes.

Reviewer XmGE7/10 · confidence 3/52023-07-05

Summary

The paper proposes a tight upper bounds for the equalized odds violation of a predictor in a setting without sensitive attributes. It also presents a post-processing correction method to control the worst-case equalized odds violation and presents results on a variety of synthetic and real datasets.

Strengths

1. The paper addresses a critical issue in fair ML. 2. The tight and computable upper bounds for equalized odds (EOD) violation is a valuable contribution as it allows for a precise understanding of the worst-case EOD violation of a predictor. 3. The proposed post-processing correction method for controlling worst-case EOD is interesting. 4. The paper is backed by experiments on both synthetic and real datasets, strengthening the validity of the results.

Weaknesses

1. Assumption 1 may limit its applicability in scenarios where the proxy sensitive attributes are not accurate. 2. The paper's results are limited to EOD and its relaxations as definitions of fairness. Extending the results to other notions of fairness may be non-trivial and deserves further consideration. 3. The paper does not consider how one could train a classifier from scratch to have minimal violations, which could be an important

Questions

1. Could you provide more insight into the limitations of Assumption 1 and how these might be addressed? 2. Could you elaborate on how the proposed post-processing correction method could be implemented in real-world scenarios? 3. How could your results be extended to other definitions of fairness, and what challenges might be encountered in this process?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper relies on Assumption 1, which may limit its applicability in scenarios where the proxy sensitive attributes are not accurate. Despite this limitation, this work is a nice contribution.

Reviewer XmGE2023-08-11

Thanks for the detailed rebuttal. All my questions have been addressed.

Reviewer RpWQ2023-08-14

I want to thank the authors for their careful answers. The paper substantially improved after the revision. I am suggesting that the papers get accepted, increasing the score to 7 and the contribution score to 3. Sorry for not sharing the citation. I probably had a problem with the format I copied. However, I still think it could be interesting for the authors to explore the connections between [1] and their work. [1] Alghamdi, W. et al. Beyond Adult and COMPAS: Fair Multi-Class Prediction via Information Projection. NeurIPS 2022.

Reviewer ntnH2023-08-16

Thanks for the authors' revision. I would recommend accepting this paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC