acceptodds
Under review as a conference paper at ICLR 2027

Global Fairness: Learning Fair Score Functions Across Acceptance Rates

Abstract

Fair learning methods typically optimize the accuracy–fairness trade-off of a binary classifier at a particular decision threshold. In many applications, however, the deployment threshold is determined by a desired acceptance rate, which may be unknown during training and vary with downstream requirements. This motivates learning a score function that remains both fair and predictive across a range of acceptance rates. Existing score-level fairness measures move beyond a single threshold, but do not explicitly account for the range of acceptance rates over which group disparity persists. We introduce *global fairness*, a score-function fairness measure that aggregates threshold-wise group disparity according to the population score density. Our key observation is that the location of disparity relative to this density determines its acceptance-rate coverage: the same disparity in a high-density score region spans a wider range of acceptance rates than in a low-density region. We show that this density-weighted formulation is equivalent to averaging group disparity uniformly over population acceptance rates. Consequently, global fairness can distinguish score functions whose threshold-wise disparities have the same magnitude and shape but substantially different acceptance-rate coverage, including cases not distinguished by Kolmogorov–Smirnov or Wasserstein-1 measures. To jointly account for fairness and predictive utility, we introduce *global accuracy*, which aggregates predictive performance across acceptance rates, and formulate learning as maximizing global accuracy subject to a global fairness constraint. We establish consistency of the empirical global fairness estimator and characterize global accuracy over the full acceptance-rate range through its relationship with AUC. Experiments on three fairness benchmarks demonstrate favorable global accuracy–fairness trade-offs and more stable fairness across acceptance rates, while remaining competitive under conventional single-threshold evaluation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Algorithmic Fairness, Fair Score Functions, Acceptance Rates, Global Fairness

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aMy8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces global fairness, a score-level metric that integrates demographic-parity disparities across population acceptance rates. It proposes a smoothed estimator with a consistency guarantee and proves that global accuracy is an affine, strictly increasing function of AUC under certain continuity assumptions. The method is evaluated against six baselines on ACSIncome, Adult, and COMPAS. The authors report improved accuracy-fairness trade-offs, although the extent to which the experiments substantiate this claimed advantage remains an important question.

Critical strengths
  1. The metric addresses limitations of evaluating fairness at a single threshold.
  2. The consistency result and relationship between global accuracy and AUC provide useful mathematical interpretation.
Critical weaknesses
  1. It is unclear whether evaluating Fairlearn's mixture and individual classifiers separately allows for a fair comparison with the other methods.
  2. The paper does not provide enough evidence to determine whether the results are consistent across random seeds and data splits or whether the baselines are tuned comparably. Also, improvements in overall fairness may hide larger disparities at certain acceptance rates.
Questions for authors
  1. How are Fairlearn's mixtures and individual classifiers converted into scores for comparison, and how is the accuracy-fairness trade-off evaluated?
  2. Are the results consistent across different random seeds and data splits, and are the baselines tuned using comparable procedures?
  3. Do the disparity-versus-acceptance-rate curves show improvements across the full range, or do the overall results hide larger disparities at certain acceptance rates?
5 comments

W1 & Q1. Fair comparison with Fairlearn's mixture and individual classifiers.

A1. We thank the reviewer for raising this important concern. Fairlearn's exponentiated-gradient algorithm (Agarwal et al., 2018) iteratively generates base classifiers and learns a randomized mixture of binary classifiers, rather than a mixture of continuous scoring functions. Therefore, the original framework does not uniquely specify how to construct a deterministic score function from the learned mixture.

Score construction. In our implementation, each base classifier is a neural network that naturally produces a continuous logit score st,ε(x)s_{t,\varepsilon}(x). We construct a deterministic mixture score by taking the weighted average of the base classifiers' predicted positive-class probabilities using Fairlearn's learned mixture weights:

smix,ε(x)=∑twt,ε σ(st,ε(x)),s_{\mathrm{mix},\varepsilon}(x)=\sum_t w_{t,\varepsilon}\,\sigma\big(s_{t,\varepsilon}(x)\big),

where σ(⋅)\sigma(\cdot) denotes the sigmoid function. Importantly, Fairlearn's mixture weights are learned to combine binary classifiers under its original randomized-classification objective, rather than to optimize the accuracy–fairness trade-off of a deterministic continuous score function. Consequently, directly applying these weights to continuous scores does not necessarily yield a favorable score-level representation.

To avoid restricting Fairlearn's performance to this particular mixture-score construction, we additionally include the continuous scores produced by all individual base classifiers generated during Fairlearn's optimization in its candidate score function set. This allows Fairlearn to be evaluated using both the weighted mixture score and the individual base-classifier scores, without requiring additional training or learnable post-processing parameters.

Evaluation and fairness of comparison. For each candidate score function, we evaluate accuracy and DP/EO gaps at acceptance rates α∈{0.1,…,0.9}\alpha\in\{0.1,\ldots,0.9\}, and average them to obtain global accuracy and global fairness. These metrics are computed using the same protocol as for all competing methods. Importantly, including the individual base-classifier scores expands Fairlearn's candidate set and cannot deteriorate its attainable global accuracy–fairness Pareto frontier relative to evaluating only the mixture scores. Thus, our evaluation provides Fairlearn with additional opportunities to achieve favorable trade-offs rather than artificially restricting its score-level performance.

Empirical evidence. Despite this expanded candidate set, our method achieves more favorable global accuracy–fairness trade-offs, as shown in Figures 2 and 7. Figures 3 and 4 further demonstrate the advantages across individual acceptance rates. In particular, Figure 3 shows that, among models with global accuracy of at least 64.5% on ACS, our method maintains a DP gap below 0.02 across all evaluated acceptance rates, whereas Fairlearn and the other competing methods satisfy this level only over subsets of the range.

We emphasize that our deterministic mixture score is a score-level adaptation of Fairlearn's learned models rather than its original randomized binary classifier, and we do not assume that it inherits the latter's fairness guarantees.

W2 & Q2: Robustness across random seeds, data splits, and baseline tuning.

We thank the reviewer for raising these important concerns regarding experimental robustness and the comparability of baseline tuning.

Robustness across random seeds and data splits. We clarify that Figures 3 and 4 report results obtained from five random model seeds under a fixed training/validation/test split. To further investigate whether the observed improvements remain consistent across different data splits, we conduct additional experiments on ACS using five independently generated data splits, each with five random model seeds. All methods are evaluated on the same data splits to ensure direct comparability. The results are reported as mean ± standard deviation (SD) over the five model seeds for each data split.

First, the following table reports global accuracy GAcc(s)\mathrm{GAcc}(s) under the global fairness constraint ΔG(s)≤0.025\Delta_G(s)\leq 0.025.

Global Accuracy (%) under ΔG(s)≤0.025\Delta_G(s)\leq 0.025 on ACS (DP):

MethodSplit 0Split 1Split 2Split 3Split 4
Fairret64.07±1.1764.46±0.5065.02±0.2364.17±0.1264.34±0.58
Fairlearn65.17±0.4865.17±0.4262.50±2.4863.54±0.4364.67±0.52
Wasserstein-165.67±0.1364.13±1.8361.69±0.8162.49±1.5264.08±1.93
Ours65.80±0.1865.56±0.0765.22±0.1764.58±0.2065.19±0.20

As in Figure 2, ERM, FairScore, and Minimax do not achieve the required global fairness level and are therefore excluded from this constrained-accuracy comparison. Our method achieves the highest mean global accuracy across all five data splits while satisfying the specified global fairness constraint.

Second, we evaluate global fairness ΔG(s)\Delta_G(s) under the global accuracy constraint GAcc(s)≥64.5%\mathrm{GAcc}(s)\geq64.5\%.

Global DP Gap (%) under GAcc(s)≥64.5%\mathrm{GAcc}(s)\geq64.5\% on ACS:

MethodSplit 0Split 1Split 2Split 3Split 4
ERM11.45±0.5013.25±0.2012.23±0.339.43±0.7412.14±0.27
Fairret2.24±0.452.98±0.403.52±0.503.45±0.412.82±0.45
Fairlearn6.60±2.907.27±3.963.25±2.453.32±2.157.26±3.68
Minimax9.78±0.8411.49±0.799.17±0.477.05±0.599.79±0.69
FairScore6.17±0.738.09±0.847.56±1.076.32±0.497.49±0.74
W11.75±0.152.84±1.422.95±0.651.81±0.312.51±1.50
Ours1.20±0.351.28±0.791.02±0.401.74±0.471.57±0.58

Our method consistently achieves the lowest mean global DP gap across all five data splits while maintaining the required global accuracy. Taken together, these results provide additional evidence that the improvements are not specific to a particular random initialization or data partition, but remain consistent across the evaluated model seeds and data splits.

Comparability of baseline tuning and evaluation. As detailed in Appendix C, we employ a common two-layer ReLU scoring architecture with hidden dimension 64 and a 60%/20%/20% training/validation/test split. The baselines are implemented using their publicly available codebases. For each fairness-aware method, including ours, we distinguish between two types of hyperparameters. First, we systematically vary the fairness budget or an analogous fairness regularization-strength parameter to construct a family of candidate score functions, rather than selecting a single value through validation. Second, for all other hyperparameters, we follow the procedures specified in the respective original papers, either tuning them on the validation set or adopting the recommended values. Appendix C.2 provides the corresponding hyperparameter settings and search ranges.

Importantly, all candidate score functions are evaluated using the same global accuracy and global fairness metrics across acceptance rates α∈0.1,0.2,…,0.9\alpha\in\\{0.1,0.2,\ldots,0.9\\}. Each candidate score function therefore corresponds to a paired point (ΔG(s),GAcc(s))(\Delta_G(s),\mathrm{GAcc}(s)), from which we construct the accuracy–fairness Pareto frontiers in Figures 2 and 7.

For the acceptance-rate-wise comparisons in Figures 3 and 4, we apply the same constrained selection criteria to all methods. Specifically, under a prescribed global accuracy threshold, we select the candidate score function with the smallest global fairness gap among those satisfying the accuracy constraint. Conversely, under a prescribed global fairness threshold, we select the candidate with the highest global accuracy among those satisfying the fairness constraint. We then report the corresponding fairness gaps or accuracies at each acceptance rate.

Thus, while the fairness-controlling parameters are systematically varied to characterize each method's accuracy–fairness trade-off, the remaining hyperparameters follow the original methods' recommended tuning procedures or settings. All methods are compared using the same evaluation metrics and constrained selection criteria, ensuring a consistent evaluation protocol across different methods.

W2 & Q3: Fairness disparities across individual acceptance rates

A3. We thank the reviewer for highlighting the possibility that improvements in aggregate fairness could conceal larger disparities at individual acceptance rates.

While Figures 2 and 7 present the overall global accuracy–fairness trade-offs, where global accuracy and fairness are computed by averaging the corresponding metrics across acceptance rates α∈0.1,0.2,…,0.9\alpha\in\\{0.1,0.2,\ldots,0.9\\}, we specifically include Figures 3 and 4 to examine whether improvements in aggregate metrics are accompanied by favorable performance at individual acceptance rates, rather than potentially masking large disparities at certain rates.

Specifically, Figures 3 and 4 report accuracy and fairness gaps at each evaluated acceptance rate under comparable global performance constraints.

In Figure 3, among score functions with global accuracy of at least 64.5% on ACS, our method maintains a DP gap below 0.02 across all nine evaluated acceptance rates, whereas competing methods satisfy this fairness level only over subsets of the acceptance-rate range.

Similarly, Figure 4 shows that, among score functions with global accuracy of at least 56.0% on COMPAS, our method maintains an EO gap below 0.05 across all nine evaluated acceptance rates, while competing methods do not consistently satisfy this fairness level across the entire range.

These results demonstrate that the improvements of our method are not merely reflected in the averaged global fairness metric, but also translate into consistently low disparities across the evaluated acceptance rates.

Beyond these empirical findings, we further establish a quantitative connection between global fairness and maximum disparity across acceptance rates. At the population level, let ds(α)d_s(\alpha) denote the threshold-wise fairness disparity at acceptance rate α\alpha, and suppose that dsd_s is LL-Lipschitz continuous on [0,1][0,1]. Since ΔG(s)=∫01ds(α) dα\Delta_G(s)=\int_0^1 d_s(\alpha)\,d\alpha, for any r∈(0,1]r\in(0,1], we have

\leq \frac{\Delta_G(s)}{r}+Lr.$$ In particular, when $L>0$ and $\Delta_G(s)\leq L$, choosing $r=\sqrt{\Delta_G(s)/L}$ yields $$\sup_{\alpha\in[0,1]}d_s(\alpha) \leq 2\sqrt{L\Delta_G(s)}.$$ **This result shows that controlling global fairness also limits how sharply large disparities can be concentrated at individual acceptance rates.** In particular, a large disparity at one acceptance rate must contribute to the overall global fairness measure through its neighboring rates, with the resulting maximum-disparity bound depending on the Lipschitz constant $L$. Taken together, the acceptance-rate-wise experiments demonstrate consistently low disparities at the evaluated rates, while the theoretical analysis further explains how global fairness can control maximum disparity across the continuous acceptance-rate range when the disparity function is Lipschitz continuous.
Reviewer_BuE4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of learning fair scoring functions when the deployment acceptance rate is not necessarily known at training time. Standard fairness-aware classifiers are typically optimized for a particular operating threshold, and changing this threshold after training may lead to increased fairness disparities or degraded predictive utility.

The authors introduce global fairness, a measure that aggregates threshold-wise demographic-parity disparities weighted by the population score density. This formulation accounts for both the magnitude of disparity and the range of acceptance rates over which it persists. Global accuracy is also introduced as the average classification accuracy across acceptance rates, and the learning problem consists of maximizing global accuracy subject to a global fairness constraint.

The main theoretical contributions establish the equivalence between global fairness and acceptance-rate-averaged disparity, the consistency of the empirical global fairness estimator, and the relationship between global accuracy and AUC. The framework is implemented using a pairwise ranking surrogate and smoothed empirical CDFs, with primal-dual optimization. Experiments on ACS, Adult, and COMPAS demonstrate favorable global accuracy–fairness trade-offs compared with threshold-based and score-level fairness approaches.

Critical strengths

I previously reviewed an earlier version of this work, and I appreciate the substantial improvements introduced in the current submission.

First, the motivation and significance of the proposed problem are now considerably better established. The new motivating example in Section 3 and Figure 1 illustrates how two scoring functions can have identical threshold-wise demographic-parity disparity patterns, while these disparities affect substantially different acceptance-rate ranges. The example provides a concrete explanation of why the proposed metric captures a distinction that is overlooked by KS and Wasserstein-1 distances. The corresponding theoretical characterization and quantitative example considerably strengthen the motivation and originality of the work.

Second, the theoretical guarantees have been strengthened. In particular, Theorem 4.1 establishes consistency of the smoothed empirical global fairness estimator, with an estimation error of order (Op(n−1/2+h)O_p(n^{-1/2}+h)). This is an important addition, as the original submission did not sufficiently justify the connection between the population-level fairness quantity and its empirical approximation.

Third, the experimental evaluation has been improved. The authors now include two particularly relevant score-level fairness baselines, namely the Wasserstein-based approach and FairScore, in addition to the existing threshold-level methods. This provides a more meaningful comparison with approaches addressing fairness beyond a single operating threshold. The parameter analysis in Figure 6 is also a welcome addition, as it provides evidence about the sensitivity of the learning procedure to the smoothing bandwidth (h) and the ranking surrogate parameter (γ\gamma).

Overall, the revised submission provides a clearer motivation, stronger theoretical foundations, and a more convincing empirical evaluation than the previous version.

Critical weaknesses

Although several of my previous concerns have been addressed, a few limitations remain.

First, the theoretical consistency guarantee is established for a fixed scoring function, whereas the actual learning procedure jointly optimizes the scoring function and the empirical fairness constraint:

\quad \text{s.t.} \quad \widehat{\Delta}G(s\phi)\leq\varepsilon.$$ Theorem 4.1 provides an important statistical justification for the fairness estimator, but it does not directly establish that the learned score function satisfies the corresponding population-level constraint ($\Delta_G(s_\phi)\leq\varepsilon$). A uniform generalization guarantee over the hypothesis class would strengthen the theoretical contribution. Nevertheless, the new result represents a substantial improvement over the original submission. Second, the parameter analysis remains somewhat limited. Figure 6 investigates the influence of (h) and ($\gamma$), but only on COMPAS under a particular fairness budget ($\varepsilon=0.3$). Moreover, the results indicate that the choice of ($\gamma$) can noticeably affect performance. Although this analysis partially addresses my previous concern about the reliability of the empirical approximations, it would be useful to establish whether the observed behavior generalizes to other datasets and, particularly, more stringent fairness constraints. Finally, global fairness remains an average measure: $$\Delta_G(s)=\int_0^1\Delta_{DP}(h_{t(\alpha)}),d\alpha.$$ Consequently, a small global fairness gap does not necessarily guarantee small disparities at every deployment acceptance rate. This limitation is inherent to the proposed formulation and does not invalidate its contribution, but the distinction between average fairness and worst-case fairness should be discussed more explicitly, particularly when comparing the approach with KS-based methods.
Questions for authors

Figure 6 provides useful evidence regarding the impact of (h) and (γ\gamma). Could the authors clarify whether similar stability is observed under stricter fairness budgets, where accurate estimation of small group disparities is particularly important? A small additional experiment would help establish the robustness of the empirical approximation in the high-fairness regime.

Given that (ΔG(s)\Delta_G(s)) measures average disparity across acceptance rates, could the authors clarify whether there are conditions under which a small global fairness gap also implies meaningful control of the maximum disparity over a deployment interval? Otherwise, it would be useful to discuss situations where a KS-based objective might be preferable despite the empirical advantages of global fairness.

4 comments

W1: Generalization guarantee for the learned scoring function

A1. We thank the reviewer for this insightful observation. We agree that Theorem 4.1 establishes the consistency of the empirical global fairness estimator for a fixed scoring function, but does not directly guarantee population-level fairness for a data-dependent scoring function obtained by solving Eq. (14).

Nevertheless, the proof technique of Theorem 4.1 can be extended to derive a uniform estimation-error decomposition over a hypothesis class, providing a route toward generalization guarantees for learned scoring functions.

Let S\mathcal S denote the class of candidate scoring functions. For each s∈Ss\in\mathcal S, let Fz,sF_{z,s} denote the conditional CDF of s(X)s(X) given Z=zZ=z, where z∈−1,1z\in\\{-1,1\\}. Define

\quad g_s(x)=|D_s(s(x))|.$$ Let $P$ and $P_n$ denote the population and empirical distributions of $X$, respectively. The population global fairness is then $\Delta_G(s)=Pg_s$. Similarly, letting $\widehat F_{z,s,h}$ denote the smoothed empirical conditional CDF with bandwidth $h$, define $$\widehat D_s(t)= \widehat F_{1,s,h}(t) - \widehat F_{-1,s,h}(t), \quad \widehat g_s(x) = |\widehat D_s(s(x))|.$$ The empirical global fairness can be expressed as $\widehat\Delta_G(s)=P_n\widehat g_s$. **Step 1: Decomposition of the global fairness estimation error.** For any $s\in\mathcal S$, adding and subtracting $P_ng_s$ yields $$|\widehat\Delta_G(s)-\Delta_G(s)| = |P_n\widehat g_s-Pg_s| \leq |P_n(\widehat g_s-g_s)| + |(P_n-P)g_s| \leq \sup_x|\widehat g_s(x)-g_s(x)| + |(P_n-P)g_s| \leq \sum_{z\in\{-1,1\}} \sup_{t\in\mathbb R} |\widehat F_{z,s,h}(t)-F_{z,s}(t)| + |(P_n-P)g_s|,$$ where the last inequality follows from the reverse triangle inequality. **Step 2: Decomposition of the smoothed CDF estimation error.** Let $\widetilde F_{z,s}$ denote the unsmoothed empirical conditional CDF. For the Gaussian CDF kernel, with $U\sim\mathcal N(0,1)$, we have $$\widehat F_{z,s,h}(t) = \mathbb E_U[ \widetilde F_{z,s}(t-hU)].$$ Assume that the group-conditional score densities are uniformly bounded by $M$ over $\mathcal S$. Then, for any $s\in\mathcal S$, $$|\widehat F_{z,s,h}(t)-F_{z,s}(t)|= \Big| \mathbb E_U[ \widetilde F_{z,s}(t-hU)] -F_{z,s}(t) \Big| \leq \left| \mathbb E_U[ \widetilde F_{z,s}(t-hU) -F_{z,s}(t-hU)] \right| + \left| \mathbb E_U[ F_{z,s}(t-hU)] -F_{z,s}(t) \right| \leq \sup_{v\in\mathbb R} |\widetilde F_{z,s}(v)-F_{z,s}(v)| + Mh\mathbb E|U| = \sup_{v\in\mathbb R} |\widetilde F_{z,s}(v)-F_{z,s}(v)| + Mh\sqrt{\frac{2}{\pi}}.$$ The first term captures the empirical CDF estimation error, while the second term bounds the smoothing bias. **Step 3: Uniform estimation-error decomposition.** Define $$\mathcal E_z(\mathcal S) = \sup_{s\in\mathcal S} \sup_{t\in\mathbb R} |\widetilde F_{z,s}(t)-F_{z,s}(t)|$$ and $$\mathcal E_{\mathrm{out}}(\mathcal S)= \sup_{s\in\mathcal S} |(P_n-P)g_s|.$$ Combining the preceding inequalities and taking the supremum over $s\in\mathcal S$, we obtain $$\sup_{s\in\mathcal S} |\widehat\Delta_G(s)-\Delta_G(s)|\leq \mathcal E_1(\mathcal S) + \mathcal E_{-1}(\mathcal S) + \mathcal E_{\mathrm{out}}(\mathcal S) + 2Mh\sqrt{\frac{2}{\pi}}.$$

Step 4: Complexity requirements for uniform convergence.

To obtain a high-probability generalization bound from this decomposition, the relevant function classes must be controlled.

For the conditional CDF estimation errors Ez(S)\mathcal E_z(\mathcal S), the relevant class is the threshold indicator class

\mathcal H=\left\\{x\mapsto\mathbf 1\\{s(x)\leq t\\}: s\in\mathcal S,\ t\in\mathbb R\right\\}.

Uniform convergence of the empirical conditional CDFs can be established under suitable complexity conditions on H\mathcal H, such as finite VC dimension, provided that the group sample sizes are sufficiently large.

For the outer empirical-process error Eout(S)\mathcal E_{\mathrm{out}}(\mathcal S), the relevant class is instead the induced function class \mathcal G=\left\\{x\mapsto |F_{1,s}(s(x))-F_{-1,s}(s(x))|: s\in\mathcal S\right\\}. This distinction is important because the functions in G\mathcal G depend on both the scoring function and its induced group-conditional score distributions. Consequently, controlling the complexity of the threshold indicator class alone does not automatically establish uniform convergence of the outer empirical-process term. Additional complexity control on G\mathcal G, for example through appropriate covering-number or Rademacher-complexity conditions, is needed.

Under suitable uniform convergence conditions for both H\mathcal H and G\mathcal G, together with a bandwidth hh that decreases appropriately with the sample size, the decomposition yields a high-probability uniform estimation-error bound of the form

|\widehat\Delta_G(s)-\Delta_G(s)| \leq \beta_n,$$ where $\beta_n$ accounts for the two conditional CDF estimation errors, the outer empirical-process error, and the smoothing bias. **Step 5: Implication for the learned scoring function.** Let $s_{\widehat\phi}\in\mathcal S$ denote the scoring function learned by solving Eq. (14). On the event where the above uniform bound holds, any empirically feasible solution satisfying $$\widehat\Delta_G(s_{\widehat\phi}) \leq\varepsilon$$ also satisfies $$\Delta_G(s_{\widehat\phi}) \leq \widehat\Delta_G(s_{\widehat\phi}) + \left|\Delta_G(s_{\widehat\phi}) - \widehat\Delta_G(s_{\widehat\phi})\right| \leq \widehat\Delta_G(s_{\widehat\phi}) + \sup_{s\in\mathcal S} |\Delta_G(s)-\widehat\Delta_G(s)| \leq \varepsilon+\beta_n.$$ Thus, **uniform estimation-error control directly translates empirical fairness feasibility into approximate population-level fairness, even when the scoring function is learned from the same data used to estimate global fairness.** We emphasize that Theorem 4.1 itself establishes fixed-function consistency rather than this uniform guarantee. Deriving an explicit convergence rate for the neural-network hypothesis class used in our implementation requires additional complexity analysis of both $\mathcal H$ and $\mathcal G$. Nevertheless, the decomposition above identifies the relevant error sources and provides a principled theoretical route toward population-level fairness guarantees for data-dependent scoring functions. **W2/Q1. Parameter sensitivity under stringent fairness constraints** **A2.** We thank the reviewer for highlighting the importance of evaluating parameter sensitivity under stringent fairness constraints and across different datasets. **First, we apologize for a typographical error in the original Figure 6: the fairness budget used for COMPAS was actually $\epsilon=0.03$, rather than $\epsilon=0.3$.** Therefore, the original experiment already examined parameter sensitivity under a relatively stringent fairness constraint. We have corrected this error in the revised manuscript. **Second, to further assess the robustness of our method, we have extended the parameter sensitivity analysis to two additional datasets under even stricter fairness budgets:** ACS with $\epsilon=0.01$ and Adult with $\epsilon=0.004$. The updated results are presented in [Figure](https://anonymous.4open.science/r/Parameter-analysis/) and Figure 6 of the revised manuscript, which now includes all three datasets. Across these datasets and fairness budgets, we observe consistent trends: **global accuracy remains relatively stable across the considered bandwidths $h$, whereas the choice of $\gamma$ has a more noticeable influence on performance.** In particular, a small value such as $\gamma=0.5$ tends to yield lower accuracy, while moderate values generally provide more stable performance. Based on these observations, we fix $\gamma=2$ across all datasets rather than tuning it separately for each dataset. These additional experiments demonstrate that the observed parameter sensitivity trends extend beyond COMPAS to ACS and Adult, including settings with more stringent fairness constraints. They also provide further empirical support for the robustness of our method to the bandwidth choice, while clarifying the importance of avoiding excessively small values of $\gamma$.

W3/Q2. Average fairness versus worst-case disparity

We agree with the reviewer that global fairness measures average disparity across population acceptance rates and does not, in general, guarantee small disparity at every individual acceptance rate. This distinction is fundamental to our formulation and reflects a different objective from KS-based worst-case fairness.

Importantly, global fairness provides a quantitative guarantee on the acceptance-rate coverage of unfairness, even without additional regularity assumptions.

For any disparity tolerance η>0\eta>0, define

\left\\{ \alpha\in[0,1]: \Delta_{\mathrm{DP}}(h_{t(\alpha)})>\eta \right\\}.$$ By Theorem 3.3, $$\begin{aligned} \eta\operatorname{Leb}(B_\eta(s)) &\leq \int_{B_\eta(s)} \Delta_{\mathrm{DP}}(h_{t(\alpha)})\,d\alpha\leq \Delta_G(s). \end{aligned}$$ Consequently, $$\operatorname{Leb}(B_\eta(s)) \leq \frac{\Delta_G(s)}{\eta}.$$ Therefore, although global fairness does not ensure pointwise fairness at every acceptance rate, a small global fairness gap limits the measure of acceptance rates at which disparity exceeds a prescribed tolerance. This directly formalizes our central motivation of reducing how broadly substantial unfairness persists across possible deployment acceptance rates. Furthermore, although global fairness measures average rather than worst-case disparity, **quantitative control of maximum disparity can be established under additional regularity assumptions**. Let $$d_s(\alpha)=\Delta_{\mathrm{DP}}(h_{t(\alpha)}),$$ and suppose that $d_s$ is $L$-Lipschitz continuous with respect to the acceptance rate $\alpha$. For any deployment interval $I=[a,b]\subseteq[0,1]$ of length $w=b-a>0$, define $M_I=\sup_{\alpha\in I}d_s(\alpha)$. Since Lipschitz continuity implies continuity, there exists $\alpha^{\star}\in I$ such that $d_s(\alpha^{\star})=M_I$. By the Lipschitz condition, for any $\alpha\in I$, $$d_s(\alpha)\geq d_s(\alpha^{\star})-L|\alpha-\alpha^{\star}| =M_I-L|\alpha-\alpha^{\star}|.$$ Now, for any $0<r\leq w$, we can construct a subinterval $J\subseteq I$ of length $r$ containing $\alpha^{\star}$. Since $|\alpha-\alpha^{\star}|\leq r$ for all $\alpha\in J$, it follows that $d_s(\alpha)\geq M_I-Lr$ throughout $J$. Using the nonnegativity of $d_s$ and the definition of global fairness, we obtain $$\Delta_G(s) =\int_0^1 d_s(\alpha)\,d\alpha \geq\int_J d_s(\alpha)\,d\alpha \geq r(M_I-Lr).$$ Rearranging gives $$\sup_{\alpha\in I}d_s(\alpha) \leq \frac{\Delta_G(s)}{r}+Lr, \qquad 0<r\leq w.$$ Furthermore, when $L>0$ and $\Delta_G(s)\leq Lw^2$, choosing $r=\sqrt{\Delta_G(s)/L}$ yields $$\sup_{\alpha\in I}d_s(\alpha) \leq 2\sqrt{L\Delta_G(s)}.$$ When $\Delta_G(s)>Lw^2$, choosing $r=w$ instead gives $M_I\leq\Delta_G(s)/w+Lw$. Thus, **under a bounded Lipschitz constant, reducing global fairness also provides quantitative control of the maximum disparity over any deployment interval of positive length**. We emphasize that this is a conditional guarantee under the stated regularity assumption, rather than an unconditional worst-case guarantee of global fairness. This result shows that, when disparity varies sufficiently smoothly across acceptance rates, controlling global fairness also provides quantitative control of the maximum disparity over a deployment interval. Without such regularity, however, a small average disparity may coexist with a narrow region of relatively large disparity. We therefore agree that KS-based fairness may be preferable in applications requiring strict worst-case disparity control across all possible deployment thresholds. Our formulation instead targets applications where the deployment acceptance rate is uncertain and the objective is to control both the magnitude and the acceptance-rate coverage of disparity while preserving predictive utility.

The additional parameter sensitivity experiments and the clarification regarding average versus worst-case fairness meaningfully address some of my concerns.

The theoretical discussion of generalization is also useful, although it remains a possible direction toward a formal guarantee rather than an established result for the learned scoring function.

Overall, I appreciate the authors' efforts and find the responses satisfactory in clarifying the scope and limitations of their contributions. However, since these clarifications do not fundamentally change my original assessment of the paper, I will maintain my score of 3 (Weak Acceptance).

Reviewer_EesROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies fair classification when the acceptance rate is unknown during training or may change after deployment. They propose a new fairness measure called global fairness, which measures group disparity across the entire range of possible acceptance rates.

Critical strengths
  1. The proposed global fairness measure has a clear mathematical interpretation.
Critical weaknesses
  1. To solve the constrained problem (14), this paper uses the Lagrangian formulation (15) and applies a primal–dual gradient method. However, the optimization problem is nonconvex in the neural network setting, and the paper does not provide convergence guarantees for the proposed primal–dual procedure.

  2. The contribution of this paper is limited. First, the global fairness measure has some limitations. The proposed fairness measure minimizes average disparity, but a small average disparity does not necessarily imply that the disparity is small at every acceptance rate. Hence, global fairness does not guarantee fairness at every acceptance rate. Second, while the proposed global fairness metric offers an interpretable way to aggregate disparity over acceptance rates, its relationship to existing distributional fairness and optimal transport approaches deserves further investigation. In particular, Chzhen et al. (2020) derive an optimal regression predictor satisfying strong demographic parity using Wasserstein barycenters and establish finite-sample fairness guarantees. Although the learning tasks and utility objectives differ, this prior work demonstrates that fairness across all thresholds can already be addressed through score-distribution alignment. The authors should more clearly articulate the advantages of their relaxed fairness formulation and consider additional comparisons with distributional post-processing methods.

Reference: Chzhen, E., Denis, C., Hebiri, M., Oneto, L., and Pontil, M. (2020). Fair Regression with Wasserstein Barycenters. NeurIPS 2020.

Questions for authors

See weaknesses.

2 comments

W1. To solve the constrained problem (14), this paper uses the Lagrangian formulation (15) and applies a primal–dual gradient method. However, the optimization problem is nonconvex in the neural network setting, and the paper does not provide convergence guarantees for the proposed primal–dual procedure.

A1. We thank the reviewer for highlighting the optimization challenge. We agree that the neural-network parameterization makes the constrained learning problem in Eq. (14) nonconvex, and that the primal–dual procedure in Eq. (15) does not, by itself, guarantee convergence to a globally optimal solution.

We would like to distinguish the statistical guarantees established in our paper from optimization guarantees. Theorem 4.1 establishes the consistency of the empirical global fairness estimator, independently of the optimization algorithm. Theorem 4.2 characterizes the population global accuracy objective through its relationship with AUC. These results justify the proposed learning formulation but do not imply convergence of the nonconvex primal–dual procedure.

We employ primal–dual gradient optimization as a practical approach to handling the differentiable empirical fairness constraint. Although this approach is widely used in constrained learning, the nonconvexity introduced by neural-network parameterization prevents a general guarantee of convergence to a globally optimal solution. Establishing weaker guarantees, such as convergence to a stationary point satisfying appropriate first-order optimality conditions, would also require additional assumptions on the objective, constraints, and optimization dynamics, together with a dedicated convergence analysis. Therefore, we do not claim such guarantees for the proposed optimization procedure, and leave their theoretical investigation for future work.

Importantly, the main contributions of this work—the proposed global fairness measure, its interpretation in terms of population acceptance rates, and the consistency analysis of its empirical estimator—are independent of the particular optimization algorithm. The primal–dual procedure serves as a practical implementation of the proposed constrained learning framework, while establishing optimization convergence guarantees for the nonconvex neural-network setting remains an interesting direction for future work.

W2. Limitations of Global Fairness and Connections to Distributional Fairness

A2. We thank the reviewer for these important comments. We agree that global fairness does not guarantee small disparity at every individual acceptance rate, and that its relationship to existing distributional fairness approaches, particularly those based on Wasserstein distances and optimal transport, deserves clarification. We emphasize that our contribution is not to replace worst-case fairness or strong demographic parity, but to introduce a complementary fairness formulation that explicitly accounts for the magnitude and acceptance-rate coverage of group disparity when the deployment acceptance rate is uncertain.

First, global fairness provides a quantitative guarantee on the acceptance-rate coverage of unfairness. Although a small average disparity does not imply pointwise fairness, it limits the range of acceptance rates at which disparity exceeds a specified tolerance. Specifically, for any η>0\eta>0, define

\{\alpha\in[0,1]: \Delta_{\mathrm{DP}}(h_{t(\alpha)})>\eta\}.$$ By the acceptance-rate representation of global fairness (Theorem 3.3), we have $$\eta\,\operatorname{Leb}(B_\eta(s)) \leq \int_{B_\eta(s)} \Delta_{\mathrm{DP}}(h_{t(\alpha)})\,d\alpha \leq \Delta_G(s),$$ which immediately yields $$\operatorname{Leb}(B_\eta(s)) \leq \frac{\Delta_G(s)}{\eta}.$$ Therefore, while global fairness does not guarantee fairness at every acceptance rate, it directly controls the measure of acceptance rates at which group disparity exceeds a prescribed tolerance. This property formalizes our central motivation: **reducing global fairness limits how broadly substantial unfairness can persist across possible deployment acceptance rates**, thereby providing greater flexibility in selecting an acceptance rate with small group disparity. **Second, our formulation differs fundamentally from strong demographic parity and Wasserstein-based distributional alignment.** We appreciate the reviewer's reference to Chzhen et al. (NeurIPS 2020), which establishes an important connection between Wasserstein barycenters and fair regression under strong demographic parity. Strong demographic parity requires the group-conditional score distributions to coincide, thereby ensuring zero demographic-parity disparity at every threshold. In contrast, our framework considers $$\max_s\ \mathrm{GAcc}(s) \qquad \mathrm{s.t.}\quad \Delta_G(s)\leq\varepsilon,$$ which allows a controllable fairness–utility trade-off without requiring exact distributional alignment. This relaxation is particularly relevant when enforcing zero disparity at every threshold would impose an unnecessarily restrictive fairness requirement or compromise predictive utility. Importantly, the novelty of global fairness lies not merely in relaxing strong demographic parity, but in **how disparity is aggregated across thresholds**. As demonstrated by our motivating example in Section 3, two score functions can have identical threshold-wise disparity patterns up to translation, and consequently identical KS and Wasserstein-1 distances, while their disparities persist over substantially different ranges of population acceptance rates. Global fairness distinguishes these score functions by weighting threshold-wise disparity according to the population score density, which is mathematically equivalent to averaging disparity uniformly over acceptance rates. Neither KS nor Wasserstein-1 captures this distinction. Furthermore, our framework directly learns a scoring function through an in-processing approach that optimizes global accuracy, characterized through its relationship with AUC, subject to a global fairness constraint. This differs from the Wasserstein-barycenter approach of Chzhen et al., which addresses fair regression under strong demographic parity. Thus, the methods differ in their fairness requirements, predictive utility objectives, and learning formulations. Our experiments additionally include Wasserstein-based and FairScore baselines to evaluate the proposed approach against existing score-level fairness methods. In summary, we acknowledge that global fairness does not provide a worst-case fairness guarantee and should not be viewed as a replacement for strong demographic parity. Rather, it offers a complementary and deployment-relevant perspective: **when the acceptance rate is unknown or may vary, fairness should account for not only the severity of group disparity but also how broadly it persists across possible acceptance rates**.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.