acceptodds
Under review as a conference paper at ICLR 2027

Worst-Case Control of Fairness for Continuous Sensitive Attributes

Abstract

Fair representation learning (FRL) seeks a representation that is informative for downstream prediction yet independent of a sensitive attribute. For a continuous sensitive attribute, existing methods measure the discrepancy between conditional and marginal distributions at each sensitive value and *average* it over the distribution of the sensitive attribute. Such an average can be small even when a fairness violation is confined to a low-probability range of sensitive values, and thus provides no uniform guarantee over the sensitive domain. We therefore propose the *supremum integral probability metric* (**supIPM**) to measure the worst-case discrepancy between the conditional and marginal distributions of the representation over sensitive values. supIPM controls the worst-case disparity of predictions, which we call -generalized demographic parity. We then propose **supFR**, an algorithm for learning representations that control worst-case fairness violations in downstream predictions. We establish a uniform convergence rate of the empirical supIPM and show that it yields a bound on downstream disparity for prediction heads of bounded Hölder norm. Experiments on tabular and graph benchmarks show that supFR attains prediction performance comparable to existing methods while substantially reducing the worst-case fairness violation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Algorithmic Fairness, Fair Representation Learning, Integral Probability Metrics, Uniform Convergence, Continuous Sensitive Attributes

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3xp3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies fair representation learning when the sensitive attribute SS is continuous. It argues that averaging the fairness violation over the sensitive attribute can hide large violations in low-density ranges. To address this issue, the authors propose supIPM, the largest ReLU-IPM between the representation distribution given S=sS = s and its marginal, over all ss. They give a kernel estimator and a training method named supFR, together with a uniform convergence rate and a bound on the worst-case disparity of H"older smooth prediction heads. Experiments on tabular and graph datasets show much lower representation-level disparity than seven baselines with similar accuracy.

Critical strengths
  1. The paper is well motivated. Figure 1 and the simulation in Figure 2 show how averaging can hide a violation in a rare range of SS.

  2. The proposed method is supported with theoretical guarantees. The bound in Theorem 1 is uniform over encoders, so it applies to the learned encoder, and Corollary 2 extends it to every head in a H"older ball.

  3. The method is validated in various datasets, and the results look promising.

Critical weaknesses
  1. The theory does not cover the cases that motivate the paper. Theorem 1 and Corollary 2 hold only on the interior region [ϵn,1−ϵn][\epsilon_n, 1-\epsilon_n] with ϵn≍1/log⁡n\epsilon_n \asymp 1/\log n, and the constants depend on the lower bound cSc_S of the density of SS. But the violations in the motivation examples are in low-density regions.

  2. The main evaluation metric is close to what supFR optimizes. Both take a worst case over ss of a function of a linear score of the representation. On measures that supFR does not train against (like kernel GDP, binned GDP in Appendix D.5), it's only comparable to the baselines. So part of the large gain may come from matching the training and test criteria.

Questions for authors
  1. Could you clarify the values of ϵn\epsilon_n and cSc_S used in the experiments? How much of the observed SS lies outside the region covered by Theorem 1, and how does the constant change as cSc_S becomes small?

  2. Could you repeat the simulation in Sec 5.1 with a more extreme value s0s_0 than the current setting s0s_0 = 0.8? That can help validate if the method works in the motivating case even where the theory does not apply.

  3. Could you clarify if the performance gain is not from the matching of training and test criteria? The authors may consider reporting a representation-level disparity that supFR does not train against.

Overall, the paper studies an important problem, and proposed method is well motivated. If the authors could clarify the gap between the theory and the motivating cases used in the paper, and whether the performance gain is from training w.r.t. the testing criteria, I'd be happy to further increase my score.

4 comments

We thank the reviewer for the careful review and for pointing out the gap between the theory on the interior region and the low-density and boundary cases that motivate the paper. This comment led to a new guarantee over the entire sensitive domain. Revisions made in response to the reviewer's comments are marked in blue in the manuscript.

Weakness 1

  • We thank the reviewer for this comment, which led to the new Remark 2 in Section 4 and Proposition 7 in Appendix A.6. With the supremum taken over the entire domain [0,1][0,1], boundary included, the same estimator still converges uniformly, at the slower rate in the table below, because the kernel bias is larger at the boundary. Corollary 2 extends in the same way, so a worst-case sensitive value near the boundary is covered. We would be grateful if the reviewer could take a look at the proof.

Uniform convergence of the empirical supIPM with bandwidth γn\gamma_n.

range of sssmoothing biasrn,δr_{n,\delta}
Theorem 1interior Sn=[ϵn,1−ϵn]\mathcal{S}_n=[\epsilon_n,1-\epsilon_n]γn2\gamma_n^2O((log⁡nn)2/5)O\left(\left(\frac{\log n}{n}\right)^{2/5}\right)
Proposition 7entire domain [0,1][0,1]γn\gamma_nO((log⁡nn)1/3)O\left(\left(\frac{\log n}{n}\right)^{1/3}\right)
  • Even under Theorem 1 alone, [ϵn,1−ϵn][\epsilon_n,1-\epsilon_n] grows to [0,1][0,1] as ϵn→0\epsilon_n\to0, so every fixed interior value, such as s0=0.8s_0=0.8 in the simulation, is eventually covered. In our response to Q2, we also report the simulation with s0=0.9s_0=0.9.
  • We would also like to clarify the setting that motivates the paper. We aim to control a violation at a sensitive value s0s_0 around which SS has a small probability, so that the violation contributes little to an average over SS. This setting does not conflict with Assumption 1, which only requires the density of SS to be bounded below by some cS>0c_S > 0. In the simulation of Figure 2, for example, the density at s0s_0 is a tenth of its maximum and the interval [s0−τ,s0+τ][s_0-\tau,s_0+\tau] has probability about 0.022, while cS>0c_S > 0.

Weakness 2

  • We agree that this possibility should be examined. Our aim is to propose a fair representation learning method that prevents worst-case fairness violations, so supFR is designed to control the supremum over ss. Under measures that supFR does not train against, supFR is comparable to the baselines, and we focus on reducing the worst-case violation.
  • For a fair comparison, we add Reg-GDP$_\infty$ as a baseline. It penalizes the worst-case disparity of its own prediction head and is thus trained on the evaluation criterion. At matched prediction performance, supFR is lower even on this prediction-level measure and markedly lower at the representation level, as the table below shows. We also add Reg-GDP$_\infty$ to all figures and tables of the real data analysis in Section 5.2 and Appendix D.

Disparities at matched prediction performance.

supFRReg-GDP$_\infty$
Adult, accuracy 0.842
prediction-level0.03500.0761
representation-level0.00160.0404
Crime, 1−MAE1-\mathrm{MAE} 0.855
prediction-level0.03710.0537
representation-level0.01000.0541
ACSIncome, accuracy 0.784
prediction-level0.03450.0493
representation-level0.00180.0210

Question 1

  • For ϵn\epsilon_n, we min-max scale SS to [0,1][0,1] and take the supremum over [0.005,0.995][0.005, 0.995], that is, ϵn=0.005\epsilon_n=0.005. The table below gives the share of observations in the training and test splits outside this range.

Share of observations outside [0.005,0.995][0.005, 0.995].

AdultCrimeACSIncome
excluded (%)1.29.50.6
  • The share is larger on Crime because every excluded observation there takes the smallest or the largest observed value of SS, which min-max scaling maps to zero and one.
  • The constant cSc_S appears only in the theory, so neither the algorithm nor the evaluation requires its value. The constant CC of Theorem 1 grows like 1/cS1/c_S by Eq. (51) in Appendix A, so the bound Crn,δCr_{n,\delta} becomes looser as cSc_S becomes small, and a larger sample is needed to attain the same error bound.

Question 2

  • We thank the reviewer for this suggestion, which led to the new Appendix C.4.1 with s0=0.9s_0=0.9. We also ran the simulation with s0=0.95s_0=0.95, even closer to the boundary. In both cases the conclusion of Figure 2 holds, and supFR attains a smaller representation-level disparity than FREM, as the table below shows.

Representation-level disparity at matched prediction performance.

supFRFREM
test 1−MAE1-\mathrm{MAE} at least 0.945
s0=0.8s_0=0.80.00180.0068
s0=0.9s_0=0.90.00300.0075
s0=0.95s_0=0.950.00260.0084

Question 3

  • We thank the reviewer for this suggestion. The new Figure 14 of Appendix D.5 reports the kernel GDP of the worst head, that is, the downstream head attaining the representation-level disparity, which supFR does not train against. supFR attains the smallest value among all methods on the three datasets, as the table below shows.

Kernel GDP of the worst head at matched prediction performance.

supFRFREM
Adult, accuracy 0.8420.00090.0096
Crime, 1−MAE1-\mathrm{MAE} 0.8550.00260.0104
ACSIncome, accuracy 0.7840.00090.0077
  • We think that this gain comes from the ReLU discriminator class, which controls the disparity of every Hölder smooth head by Corollary 2 rather than that of a single trained head.

  • We thank the reviewer again for the constructive comments. We hope that our responses have addressed the reviewer's concerns, and we would be happy to answer any further questions.

Thank you for the detailed rebuttal and the revised paper. After reading the authors' rebuttal, I still have two concerns remaining:

  1. For Weakness 1, Proposition 7 is a meaningful addition that addresses my concern of the boundary. However, the guarantee is weaker in the low-density setting which motivates the paper. The bound depends on a single c_S for the whole range, so one sparse region weakens the guarantee everywhere, not just near that region. So I agree the guarantee holds theoretically, but it is not quite clear how informative the guarantee is in practice.

  2. For Weakness 2, the addition of Reg-GDP$\infty$ didn't address my question of training w.r.t. the evaluation metric, as Reg-GDP$\infty$ also optimizes the worst case. Beating it shows supFR is a better worst-case method, but it doesn’t show that supFR’s advantage would survive a test it wasn’t trained for. In fact, the new results in Appendix D seem to suggest the opposite. When unfairness is measured with the adversarial linear head (i.e., the kind supFR trains against), supFR is significantly better, but when the same unfairness measure is applied to the ordinary trained prediction head, supFR is only comparable to the baselines. Since only the head changed, the big advantage seems to be from matching the evaluator to the training criterion. Also, the added results in Table 5 are a bit surprising to me. Reg-GDP$\infty$ directly considers the worst-case, but is worse than Reg-GDP on Adult and ACSIncome dataset. This might be because the Reg-GDP$\infty$ method is under-tuned.

We thank the reviewer for the follow-up comments. Our responses to the two remaining concerns are given below.

Question 1

  • Theorem 1 and Proposition 7 show that the empirical supIPM converges uniformly to the supIPM and give the rate of convergence. By Corollary 2, a small empirical supIPM then guarantees a small disparity at every sensitive value, including an s0s_0 of low density.
  • The lower bound cSc_S appears only in the constant CC and not in the rate. Thus the bound Crn,δCr_{n,\delta} still goes to zero as nn grows when cSc_S is small, and only a larger sample is needed to attain the same error bound.
  • In the simulation of Appendix C.4.1, the density of SS is pS(s)∝1−aexp⁡−(s−s0)2/(2⋅0.12)p_S(s)\propto 1-a\exp\\{-(s-s_0)^2/(2\cdot0.1^2)\\} with a=0.9a=0.9, so that cS/CS≈0.1c_S/C_S\approx0.1. We also ran the simulation with a=0.99a=0.99, for which cS/CS≈0.01c_S/C_S\approx0.01. For both values of aa, supFR attains a smaller representation-level disparity than FREM, as the table below shows. The unfair model is trained without the fairness penalty, that is, with λ=0\lambda=0.

Representation-level disparity at matched prediction performance.

cS/CSc_S/C_SUnfairsupFRFREM
test 1−MAE1-\mathrm{MAE} at least 0.945
a=0.9a=0.9, s0=0.9s_0=0.9 (Figure 6)0.10.10250.00300.0075
a=0.99a=0.99, s0=0.9s_0=0.90.010.05080.00180.0056
  • Relative to the unfair model, the disparity of supFR is 2.92.9\\% for a=0.9a=0.9 and 3.53.5\\% for a=0.99a=0.99. Thus supFR is only slightly less effective for the smaller cSc_S.

Question 2

  • To examine whether the gain comes from matching the evaluator to the training criterion, we change the fairness measure of the evaluation. The most unfair linear head is now chosen to maximize the kernel GDP, a measure that supFR does not train against. supFR again attains the smallest value among all methods on the three datasets, as the table below shows.

Kernel GDP of the most unfair linear head at matched prediction performance.

supFRFREM
Adult, accuracy 0.8420.00090.0101
Crime, 1−MAE1-\mathrm{MAE} 0.8550.00330.0131
ACSIncome, accuracy 0.7840.00100.0086
  • The linear heads used in the evaluation are Hölder smooth, so by Corollary 2 supFR controls their disparity at every ss. The kernel GDP averages this disparity over ss, so it is small as well.
  • Regarding Table 5, we thank the reviewer for pointing out that Reg-GDP$_\infty$ may be under-tuned. Following this comment, we retrained Reg-GDP$_\infty$ on Adult and ACSIncome with a finer grid of λ\lambda and updated Table 5. Its prediction-level disparity is now smaller than that of Reg-GDP on Adult and ACSIncome.

Disparities at matched prediction performance.

supFRReg-GDPReg-GDP$_\infty$
Adult, accuracy 0.842
prediction-level0.03500.06390.0618
representation-level0.00160.03430.0348
ACSIncome, accuracy 0.784
prediction-level0.03450.04870.0378
representation-level0.00180.02060.0159

Final remark

Our aim is to learn a fair representation that controls the unfairness arising at a sensitive value s0s_0 of low density. supFR is therefore designed to reduce the worst-case disparity of downstream heads. We hope that our responses resolve the reviewer's concerns. We would be grateful if the reviewer could let us know of any further questions.

Reviewer_UBwMOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method for fair representation learning, focused on ensuring fairness in the worst case. It uses a supremum norm based generalisation of demographic parity as the fairness metric. It provides a theoretical analysis of the approach and practical algorithm. The method is compared empirically to various competitors on the typical set of benchmarks for the field.

Critical strengths
  • The worst case GDP formulation (although not a novel contribution) is arguably of more practical interest than the average formulation.
  • Empirically, the method performs much better than other competitors.
Critical weaknesses
  • The methods are evaluated on the supremum norm GDP, however competitors are trained on a different objective. The evaluation on other fairness metrics is relegated to the appendix. I believe the discussion of the tradeoff between worst case fairness and average case fairness should be expanded on.
  • The theoretical results are not particularly useful in practice: Plugging in realistic values for the constants at play make the bounds almost meaningless. The theory assumes the true supremum is known but in practice this is obtained with an approximate search.
Questions for authors

Given two approaches, one focused on the average case and one focused on the worst case, I would expect each to outperform the other (all things being equal) when evaluated on its training objective. This does not seem to be the case and the proposed method seems to outperform also in the average case.

A similar effect can be seen in the synthetic example in Fig.2 where the authors attribute the improvements to the reduction in the peak, while FREM is significantly worse on average as well. A 3 way comparison including supFR_MMD would be more meaningful here. It would also be interesting to investigate a scenario where there's a clear tension between the average and worst case objectives.

Similarly to supFR_MMD, some of the other baselines could be adapted to the supremum objective. Comparing against methods that optimise for a different objective is not evaluating the merits of the approach in isolation.

These concerns make me somewhat skeptical of the evaluation.

6 comments

We thank the reviewer for the careful review and for raising the trade-off between worst-case and average-case fairness, which led to a new baseline and a three-way comparison in the revised manuscript. Revisions made in response to the reviewer's comments are marked in blue in the manuscript.

Weakness 1

  • We thank the reviewer for this suggestion. To compare with a competitor trained on the same kind of objective, we add Reg-GDP$_\infty$. It modifies Reg-GDP by penalizing the maximum over ss of Δ^fψ(s)\widehat{\Delta}_{f_{\psi}}(s), the estimated disparity of its prediction head at ss, in place of its average over SS. At matched prediction performance, supFR attains smaller prediction-level and representation-level disparities on all three datasets, as detailed in our response to Q3.
  • The aim of this work is to measure and reduce the worst-case fairness violation, so the main text focuses on the worst-case measures, and Appendix D.5 reports average-case measures such as the kernel GDP. Under these measures supFR is comparable to FREM, for the reason explained in our response to Q1.

Weakness 2

  • We agree your concerns of the theoritical results. Its constant Cd,βC_{d,\beta} and exponent αd,β\alpha_{d,\beta} come from the property of the ReLU-IPM [1] in Eq. (8), which bounds the Hölder IPM by a power of the ReLU-IPM. When the smoothness β\beta of the head is small relative to the dimension dd of the representation, αd,β=2β/(d+3)\alpha_{d,\beta}=2\beta/(d+3) is small and the bound is weak.
  • For smooth heads, however, the guarantee is linear. Once β\beta exceeds (d+3)/2(d+3)/2, αd,β=1\alpha_{d,\beta}=1, so the worst-case disparity of every such head is bounded by a constant multiple of the empirical supIPM plus the estimation error Crn,δCr_{n,\delta} of Theorem 1, which goes to zero as nn grows.
  • The theory indeed assumes the exact supremum over ss, whereas in practice it is found by an approximate search. This gap is small in practice. When we redo the maximization over ss on a fine grid at several epochs of training, the compass search attains the grid maximum up to a few percent, as the table below shows. The full results are given in Table 7 of the new Appendix G.

Compass search against a grid.

ratio to grid maximum, meanratio to grid maximum ≥0.95\ge 0.95 (%)
Simulation, univariate SS1.000100.0
Simulation, bivariate SS0.99696.1
Adult1.002100.0
Crime1.000100.0
ACSIncome1.001100.0

[1] Park et al. "ReLU integral probability metric and its applications." Neural Networks, 2026.

Question 1

  • We agree that, all things being equal, each approach would be expected to perform better on its own objective. This is the case when the average case is measured by the objective of FREM itself, the EIPM computed with the MMD, which we denote by EIPM⁡MMD\operatorname{EIPM}_{\mathrm{MMD}}. At the same prediction performance, FREM attains a smaller EIPM⁡MMD\operatorname{EIPM}_{\mathrm{MMD}} than supFR in the simulation and on Crime and ACSIncome, and a comparable one on Adult, as the table below shows.

EIPM⁡MMD\operatorname{EIPM}_{\mathrm{MMD}} of the representation at the same prediction performance.

supFRFREM
Simulation, models circled in Figure 20.00720.0034
Adult, accuracy 0.8370.03810.0399
Crime, 1−MAE1-\mathrm{MAE} 0.8560.12790.1199
ACSIncome, accuracy 0.7800.05430.0485
  • Thus, in most cases, FREM performs better on its own measure, as the reviewer expected. supFR can nevertheless attain a small value because the two objectives are not symmetric. Since an average over SS cannot exceed the supremum over ss, EIPM⁡V(Z;S)≤supIPM⁡V(Z;S)\operatorname{EIPM}_{\mathcal{V}}(\mathbf{Z};S)\le\operatorname{supIPM}_{\mathcal{V}}(\mathbf{Z};S) for the same discriminator class V\mathcal{V}, so reducing the supIPM also reduces the EIPM, whereas reducing the EIPM leaves the supremum uncontrolled. This asymmetry may explain why supFR did not lose under the average-case measures of Appendix D.5.

Question 2

  • We thank the reviewer for this suggestion. Figure 2 now includes supFRMMD\mathrm{supFR}_{\mathrm{MMD}}, which lies between supFR and FREM. Its disparity departs from that of FREM only near s0s_0, so the gap between supFR and FREM away from s0s_0 comes from the discriminator class.
  • For a scenario with a clear tension, the new Appendix C.4.2 adds to the simulation a dependence on SS over the whole domain. The disparity then grows toward both ends of [0,1][0,1] in addition to the peak at s0s_0.
  • In this setting, the trade-off anticipated by the reviewer appears clearly. FREM attains the smallest kernel GDP and supFR the largest, whereas supFR and supFRMMD\mathrm{supFR}_{\mathrm{MMD}} attain smaller prediction-level and representation-level disparities than FREM, as the table below shows.

Disparities in the simulation with dependence over the whole domain at matched prediction performance.

supFRsupFRMMD\mathrm{supFR}_{\mathrm{MMD}}FREM
test 1−MAE1-\mathrm{MAE} at least 0.945
kernel GDP0.01070.00900.0089
prediction-level0.07690.06710.0948
representation-level0.00400.00700.0118

Question 3

  • We agree that a comparison with methods adapted to the supremum better isolates the merits of the approach, and we add such a method, Reg-GDP$_\infty$, which replaces the average over SS in Reg-GDP by the supremum. At matched prediction performance, supFR attains smaller prediction-level and representation-level disparities than Reg-GDP$_\infty$ on all three datasets, as the table below shows. We also add Reg-GDP$_\infty$ to all figures and tables of the real data analysis in Section 5.2 and Appendix D.

Disparities at matched prediction performance.

supFRReg-GDP$_\infty$
Adult, accuracy 0.842
prediction-level0.03500.0761
representation-level0.00160.0404
Crime, 1−MAE1-\mathrm{MAE} 0.855
prediction-level0.03710.0537
representation-level0.01000.0541
ACSIncome, accuracy 0.784
prediction-level0.03450.0493
representation-level0.00180.0210

Final remark

We thank the reviewer for the sincere concerns. Our aim is to learn a fair representation that controls the unfairness arising at a sensitive value s0s_0 of low density. Our experiments therefore focus on showing that the representation learned by supFR keeps this worst case small even for the most unfair downstream head, which is why we evaluate the worst-case disparity GDP⁡∞\operatorname{GDP}_{\infty} of downstream heads. We are grateful for the reviewer's thoughtful advice.

I appreciate the detailed response.

I am still unconvinced about the accuracy of the supremum search. Even if the approximate search works on the benchmark datasets, it is going to break on high dimensional data and/or on more challenging datasets. I don't see this as a critical flaw, but it is nevertheless a limitation of the approach.

I am also still confused at Fig.2 (and by extent of the results in the appendix). It seems to shows that the method is superior on all metrics: accuracy, worst case fairness and average case fairness alike. This should should not be possible unless worst case and average case fairness are not in tension in practice.

Reviewer_NB1aOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper study fair representation learning for continuous sensitive attributes. The authors introduce a worst-case fairness metric called supIPM, and proposed supFR, which is an algorithm for training fair representations. They proved uniform convergence for the empirical IPM estimator and derive disparity bounds for downstream prediction models under Holder-smoothness assumptions.

Critical strengths
  1. While most of the existing fair representation methods optimize average fairness metrics, causing certain minority groups suffer from discrimination. This paper focus on worst-case fairness over the full sensitive domain, addressing a real-world problem.

  2. This paper provides an uniform convergence theory for empirical supIPM, linking the estimator to population-level worst-case disparity for Holder smooth prediction function.

Critical weaknesses
  1. Theory builds on very strong assumptions that are practically impossible to satisfy in real-world datasets. For example, Assumption 1 require sensitive attribute density strictly bounded away from zero and infinity, and Assumption 2 require twice continuous differentiable joint density of (X,S)(X, S). These assumptions are commonly violated in practice.

  2. The nested min-max optimization has critical practical weaknesses and is not discussed anywhere in the paper. While supFR relies on compass search over sensitive values, this may overlook the true worst case sensitive values, and thus produce misleading results. Moreover, training creates unquantifed gradient bias and higher dimensional sensitive inputs will only make it worse.

Questions for authors
  1. Theoretical foundations are mostly build on Assumptions 1 and 2, which are frequently violated in real world. Can you supplement additional experiments to show how the emipirical supIPM estimator behave when these assumptions break? Can you provide practical heurisics or modified bounds for boundary points of the sensitive supports?

  2. Algorithm updates inner variables across mini-batch training and gradient bias is introduced. Can you quantify these gradient bias? Also, can you provide diagnostics for detecting when compass-search fails to locate the true worst-case sensitive values?

4 comments

We thank the reviewer for the careful review and for the questions. They led to a guarantee over the entire sensitive domain and to new experiments on the search over ss. Revisions made in response to the reviewer's comments are marked in blue in the manuscript.

Weakness 1

  • We understand the reviewer's concern about these assumptions. We would like to note that both are standard in kernel estimation with a continuous sensitive attribute. FREM [1] imposes the same two conditions in its Assumptions 2 and 3, and describes the bound on the density of SS as very mild.
  • In addition, the assumptions are needed only for the theory, and the algorithm itself does not rely on them. Assumption 1 ensures that every sensitive value has observations nearby, so that the kernel estimate at that value is stable, and Assumption 2 controls the bias of kernel smoothing.
  • We would also like to clarify that Assumption 2 requires smoothness only in ss and none in x\mathbf{x}, as noted after Assumption 2 in Appendix A.1, so X\mathbf{X} may be discrete. In our response to Q1, we add an experiment with a small lower bound cSc_S of the density of SS.

Weakness 2

  • We thank the reviewer for these questions, which led to the new Appendix G. There we redo the maximization over ss on a fine grid at several epochs of training and compare it with the compass search. The compass search attains at least 95% of the grid maximum in nearly every comparison, as the table below shows. The full results are given in Table 7 of Appendix G.
  • For the gradient bias, Proposition 8 in Appendix B.2 shows that mini-batch gradients are unbiased when the inner variables are fixed before the mini-batch is drawn, as in Algorithm 2. The remaining bias comes from the approximate search over ss. The gradient columns of the table show that the encoder gradient under the compass search differs from that under the grid maximum by less than 1% in median.

Compass search against a grid.

ratio to grid maximum, meanratio to grid maximum ≥0.95\ge 0.95 (%)encoder gradient, median relative differenceencoder gradient, cosine ≥0.99\ge 0.99 (%)
Simulation, univariate SS1.000100.00.004099.9
Simulation, bivariate SS0.99696.10.007196.1
Adult1.002100.00.005199.7
Crime1.000100.00.002599.9
ACSIncome1.001100.00.0003100.0
  • With a bivariate SS in the new Appendix C.5, the search stays accurate, as the bivariate row above shows. In that setting, supFR also attains a smaller representation-level disparity than FREM at comparable prediction performance, 0.0005 against 0.0030. Details are in Appendices C.5 and G.

[1] Kong et al. "Fair representation learning for continuous sensitive attributes using expectation of integral probability metrics." IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

Question 1 We thank the reviewer for this question, which led to the new Remark 2 in Section 4 and Proposition 7 in Appendix A.6. With the supremum taken over the entire domain [0,1][0,1], boundary included, the same estimator still converges uniformly, at the slower rate in the table below, because the kernel bias is larger at the boundary. Corollary 2 extends in the same way, and no change to the algorithm is needed. We would be grateful if the reviewer could take a look at the proof.

Uniform convergence of the empirical supIPM with bandwidth γn\gamma_n.

range of sssmoothing biasrn,δr_{n,\delta}
Theorem 1interior Sn=[ϵn,1−ϵn]\mathcal{S}_n=[\epsilon_n,1-\epsilon_n]γn2\gamma_n^2O((log⁡nn)2/5)O\left(\left(\frac{\log n}{n}\right)^{2/5}\right)
Proposition 7entire domain [0,1][0,1]γn\gamma_nO((log⁡nn)1/3)O\left(\left(\frac{\log n}{n}\right)^{1/3}\right)
  • In line with the reviewer's concern, these assumptions do not hold exactly in the real datasets used in our experiments. In Adult, age takes only 74 integer values, so SS has no density. Even so, Figure 5 shows that the disparity of supFR remains small over the entire sensitive domain.
  • We also add a simulation in the new Appendix C.4.1. There s0s_0 moves from 0.8 to 0.9, close to the boundary, and the density of SS at s0s_0 is only a tenth of its maximum, so cSc_S is small. supFR again attains a smaller representation-level disparity than FREM, as the table below shows.

Representation-level disparity at matched prediction performance.

supFRFREM
test 1−MAE1-\mathrm{MAE} at least 0.945
s0=0.8s_0=0.8, Figure 20.00180.0068
s0=0.9s_0=0.9, Figure 60.00300.0075

Question 2

  • As in our response to W2, mini-batch gradients are unbiased for fixed inner variables by Proposition 8, and the approximate search over ss changes the encoder gradient by less than 1% in median.
  • We suggest a two-step diagnostic for a failure of the compass search. The compass search starts from TT points and returns the best end point s⋆s^\star. At each step, we first count the number mm of starting points whose end point is s⋆s^\star, which needs no extra computation. If m=1m=1, we also maximize the objective QQ over a fine grid of ss, with maximizer sgrid⋆s^\star_{\mathrm{grid}}, and regard the search as failed if ρ:=Q(s⋆)/Q(sgrid⋆)<0.95\rho:=Q(s^\star)/Q(s^\star_{\mathrm{grid}}) < 0.95.
  • With T=17T=17, no step has m=1m=1 and no step has ρ<0.95\rho < 0.95. When we reduce TT to 3, most steps have m=1m=1, and every step with ρ<0.95\rho < 0.95 has m=1m=1, as the table below shows.

Results of the two-step diagnostic.

SimulationAdultCrimeACSIncome
T=3T=3
steps with m=1m=1 (%)100.091.195.991.8
ρ<0.95\rho < 0.95 among steps with m=1m=1 (%)18.50.15.70.0
ρ<0.95\rho < 0.95 among steps with m≥2m\ge2 (%)–0.00.00.0
T=17T=17
steps with m=1m=1 (%)0.00.00.00.0
ρ<0.95\rho < 0.95 among steps with m≥2m\ge2 (%)0.00.00.00.0

Thank you for your detailed explanation and additional numerical experiments. However, I am still concerned about Assumptions 1 and 2. As you mentioned, the real dataset application presented in this paper indeed violates these assumptions, while these assumptions are essential to the proposed method. Therefore, I maintain the current score.

We thank the reviewer for the reply.

  • Defining the conditional distribution given S=sS=s at every ss requires Assumption 1. Our estimator of this distribution is a kernel regression estimator in ss, and Assumption 1 is standard for the uniform convergence of such estimators [1, 2].
  • The real datasets support Assumption 1. Let p^S\hat p_S be the kernel density estimate of SS, with CS=max⁡sp^S(s)C_S=\max_s\hat p_S(s) and cS=CS/100c_S=C_S/100. As the table below shows, almost all observations SiS_i satisfy p^S(Si)∈[cS,CS]\hat p_S(S_i)\in[c_S, C_S].
cSc_SCSC_Sobservations with p^S(Si)∈[cS,CS]\hat p_S(S_i)\in[c_S, C_S] (%)
Adult0.0202.01599.86
Crime0.0383.785100.00
ACSIncome0.0171.69699.95
  • Assumption 2 enters Theorem 1 only through the constant CC, via ∥M1∥L1(ν)\lVert M_1\rVert_{L_1(\nu)} and ∥M2∥L1(ν)\lVert M_2\rVert_{L_1(\nu)}, which can be estimated from a kernel estimate of p(x,s)p(\mathbf{x},s). It affects the size of the constant but not convergence rate.

We sincerely appreciate the time and effort the reviewer has devoted to the discussion.

[1] Einmahl, U., & Mason, D. M. (2005). Uniform in bandwidth consistency of kernel-type function estimators.

[2] Hansen, B. E. (2008). Uniform convergence rates for kernel estimation with dependent data. Econometric Theory, 24(3), 726-748.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.