acceptodds
Under review as a conference paper at ICLR 2027

Deep Fair Learning: Task-aware Fair Representations via Joint Distance-Covariance Regularization

Abstract

Ensuring fairness is essential as machine learning increasingly informs consequential decisions. However, many fairness-aware methods focus on the outputs of individual predictors, without directly controlling sensitive information retained in the underlying representations. We propose Deep Fair Learning (DFL), which combines distance covariance regularization with predictive loss to jointly learn representations and downstream predictors, promoting fairness at both levels while preserving task-relevant information. Its marginal and class-conditional formulations target independence and separation, respectively. Under suitable regularity conditions, we establish non-asymptotic joint excess-risk rates and convergence of the learned representation up to natural invariances. We further derive fairness-inheritance bounds linking representation-level dependence to downstream disparities over suitable predictor classes, extending fairness guarantees beyond the jointly trained predictor. Experiments on tabular, text, and image benchmarks show that DFL achieves lower fairness gaps than competing methods in many evaluated settings while maintaining competitive predictive accuracy, with fairness gains largely preserved after downstream retraining.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fair representation learning, algorithmic fairness, distance covariance, joint optimization, excess risk.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3bPdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to learn fair representations that stay fair when the downstream head is retrained. To this end, the authors propose Deep Fair Learning (DFL), which trains an encoder and a classifier head together with a distance covariance (DC) penalty between the representation and the sensitive attribute. The theory gives an excess-risk rate for the joint estimator, recovery of the representation up to rigid transformations, and a bound on group score gaps for new heads in a distance-kernel RKHS ball. Experiments on five datasets show that DFL ranks first or second in 30 of 48 comparisons, and that most gains persist after the head is retrained.

Critical strengths
  1. The paper studies an important problem, and the presentation is easy to follow.

  2. The proposed penalty is simple and effective. It handles general cases including independence or separation, and discrete, continuous, or multivariate sensitive attributes without an adversarial optimization or a kernel-bandwidth parameter.

Critical weaknesses
  1. The unconstrained model (STD) for Adult and Bank dataset is logistic regression, not the same DenseNet without the fairness term. DFL even beats STD in terms of accuracy on Adult dataset in Table 2. Since model selection requires 95% of STD accuracy, a weak STD loosens the accuracy floor, making it difficult to interpret if the performance gain comes from the change of the base model.

  2. Tuning strategy differs across methods. DFL searches nine values of α\alpha, while on CelebA the adversarial baselines use one fixed weight (line 2132). So it is difficult to tell whether the gains come from the DC penalty, the encoder, or the tuning budget.

  3. The theory, which the paper claims as the main contribution, is weakly linked to the experiments. In Theorem 2, the rates have exponents like βg/(2βg+p)\beta_g/(2\beta_g + p). With p=768p = 768 for BIOS and p=2,304p=2,304 for MOJI, the bound shrinks extremely slowly as nn grows. Also, Theorem 4 covers binary ZZ and RKHS heads, while the fresh head in Table 5 is a one-hidden-layer network.

  4. The most relevant method is not compared in experiments. FairDisCo and the HSIC-based methods are listed in Table 1 but not in the experiments. Since the main design choice is the DC penalty, a DC vs. HSIC comparison under the same encoder is needed.

Questions for authors
  1. Could you report STD for Adult and Bank using the same DenseNet with no fairness term? That can help clarify where the performance gain comes from.

  2. Could you clarify the tuning budget for each baseline? Could the adversarial baselines get the same nine-point grid search?

  3. Could you compare with FairDisCo or an HSIC-based method, like by swapping the DC penalty for an HSIC penalty in the same architecture?

6 comments

We thank the reviewer for the constructive review, and for recognizing the importance of the problem and the simplicity and generality of the proposed penalty. We have conducted all of the requested experiments. We address each comment below.

Architecture-Matched Unconstrained Model (W1, Q1)

"The unconstrained model (STD) for Adult and Bank dataset is logistic regression, not the same DenseNet … a weak STD loosens the accuracy floor." / Q1: "Could you report STD for Adult and Bank using the same DenseNet with no fairness term?"

STD-DenseNet is the DFL network with the fairness term removed, trained with the same optimizer, early-stopping rule, and seeds. Test results are averaged over five seeds (Appendix J.2, Table J.1, line 2540):

MajoritySTD-LR (paper)STD-DenseNet (accuracy / TPR / MCDP)DFL (accuracy / TPR / MCDP)
Adult74.5777.6778.85 / 3.67 / 31.8778.34 / 3.17 / 4.39
Bank88.6590.6991.09 / 2.29 / 24.5390.22 / 5.19 / 11.39
MOJI50.0071.1970.10 / 43.30 / 43.4773.45 / 17.98 / 17.06
  1. Effect of the accuracy floor. Recomputing the 95% floor from STD-DenseNet instead of STD-LR changes none of the selected weights, for any method, on Adult or Bank.
  2. Source of the improvement. We agree with the reviewer that the accuracy advantage of DFL over STD-LR on Adult is attributable to the network architecture. Against the matched network, the fairness improvement is attributable to the penalty: DFL reduces MCDP by 86% (Adult) and 54% (Bank), at a cost of 0.5 and 0.9 points of accuracy. On MOJI, DFL is more accurate than the matched network (73.45% versus 70.10%), which suggests that the label-informed objective also acts as a beneficial regularizer.
  3. Selection rule. The majority-class predictor attains 74.6% (Adult) and 88.7% (Bank), both within a 95% accuracy floor. We therefore added a balanced-accuracy floor to the selection rule, applied on validation data only. Under this rule, DFL attains the smallest gap among the nine fairness methods on Adult: 3.78, compared with 3.95 (LAFTR-EO-CE), 4.04 (HSIC), and 5.22 (CFair-EO).

The revised Sec. 4.2 reports this comparison alongside the original results (lines 419–421), and the complete matched table is Table J.1 (line 2540).

Equal Tuning Budgets (W2, Q2)

"DFL searches nine values of α\alpha, while on CelebA the adversarial baselines use one fixed weight." / Q2: "Could the adversarial baselines get the same nine-point grid search?"

Yes. All methods now receive the same nine-point grid on every dataset, including CelebA:

  • α∈0.1,…,0.9\alpha\in\\{0.1,\dots,0.9\\} for the DC and HSIC objectives;
  • 10−2,10−1.5,…,102\\{10^{-2},10^{-1.5},\dots,10^{2}\\} for the adversarial weights.

All methods use the same validation rule (Appendix J.1). To verify that no method is limited by the boundary of its grid, we additionally extended every grid by three weights at its fair end (Table J.1, line 2540).

MOJI (accuracy / TPR / MCDP):

  • Nine-point budgets. DFL attains the highest accuracy: 73.45 / 17.98 / 17.06, compared with 68.30 / 16.82 / 17.43 for CFair-EO and 70.44 / 23.12 / 23.90 for LAFTR-EO-CE. CFair-EO reaches a gap comparable to that of DFL only at an accuracy that is 5.2 points lower.
  • Extended grids. DFL attains 67.30 / 2.88 / 5.07. At the same accuracy, CFair-EO attains 67.37 / 14.58 / 15.19, and LAFTR-EO-CE 67.42 / 24.16 / 28.55. At 71.3% accuracy (α=0.05\alpha=0.05), DFL already attains a smaller gap than every adversarial operating point (Figure J.1).

Adult. DFL attains the smallest gap among all methods (3.78, compared with 3.95 for LAFTR-EO-CE and 5.22 for CFair-EO), and its trade-off front lies below both adversarial fronts.

CelebA. We acknowledge that, with the larger tuning budget, CFair-EO attains lower gaps than DFL at the jointly trained head. This fairness appears to reside in the head rather than in the representation: a probe recovers gender from CFair-EO's representation with AUC 82–88 on three of the four tasks, compared with 64–65 for DFL.

After the head is replaced by a new MLP-3, DFL attains the smaller gap on three of the four tasks: Young 12.8 versus 13.9, Attractive 13.1 versus 22.4, and Wavy Hair 17.0 versus 24.8. On Smiling the two are comparable (7.0 versus 6.7). LAFTR-EO-CE returns to the level of the unconstrained model on every task (e.g., Attractive 42.7 versus 40.3 for STD-DenseNet). The complete results are given in Table J.3 (line 2584) and in our response to Reviewer npFd (part 3), and they are summarized in Sec. 4.2 (lines 431–432).

In summary, the tuning budget does not account for the advantage of DFL. With equal budgets, DFL performs best on the tabular and text benchmarks. Where a tuned adversarial method is fairer at its own jointly trained head, DFL is fairer once the learned features are reused, which is the setting the paper addresses.

Connection Between Theory and Experiments (W3)

"In Theorem 2, the rates have exponents like βg/(2βg+p)\beta_g/(2\beta_g+p). With p=768p=768 for BIOS and p=2304p=2304 for MOJI, the bound shrinks extremely slowly."

Ambient versus intrinsic dimension. The exponent reflects the ambient dimension pp, which enters only through the approximation of Hölder functions on [0,1]p[0,1]^p. For inputs supported near a low-dimensional set, approximation results for deep ReLU networks (Nakada & Imaizumi, 2020; Chen et al., 2022) allow pp to be replaced by the intrinsic dimension. TwoNN estimates of the intrinsic dimension are approximately:

  • 32 for MOJI (p=2304p=2304);
  • 32 for BIOS (p=768p=768);
  • 45 for CelebA (p=512p=512).

For β=2\beta=2, the exponent for MOJI increases from 0.0009 to 0.056, a factor of approximately 64. We have added this remark after Theorem 2 (lines 252–255).

Empirical convergence. The alignment distance d∼d_\sim of Theorem 3 between the encoder learned on a subsample of MOJI and the encoder learned on the full sample decreases from 0.83 at n=3,125n=3{,}125 to 0.57 at n=50,000n=50{,}000. The latter value equals the distance between two full-sample encoders trained with different seeds, indicating that the representation has converged to the level of optimization noise.

"Theorem 4 covers binary ZZ and RKHS heads, while the fresh head in Table 5 is a one-hidden-layer network."

  • Tightness. The bound of Theorem 4 is attained over its RKHS ball (remark after Theorem 4, lines 340–344; the proof is given in our response to Reviewer PCnM, part 1). The DFL penalty therefore equals, up to a constant, the worst-case disparity over this head class.

  • Non-binary ZZ. The same argument yields ∣Cov(f(Z),h(R))∣≤M2DC(Z,R)|\mathrm{Cov}(f(Z),h(R))|\le\frac M2\sqrt{\mathrm{DC}(Z,R)} for every ff in the unit ball of the distance-induced RKHS of ZZ (lines 340–344). Binary ZZ recovers Theorem 4, and continuous ZZ is covered as well. On continuous age, DFL attains the lowest violation among four specialized baselines (see our response to Reviewer npFd, parts 6–7). Categorical ZZ is covered by Remarks F.8–F.9.

  • Heads outside the RKHS. Corollary F.7 covers hard decision rules through their RKHS approximation error. Empirically, the persistence predicted by the theory extends beyond the RKHS class:

    • for MLP-3 heads, MCDP is 4.62 (Adult) and 5.11 (Bank);
    • for gradient-boosted trees, it is 7.59 and 6.67;
    • STD-DenseNet attains 32.9 and 24.7.

    This holds across all nine operating points (Figure J.2, line 2606). The adversarial and output-level methods, which lack a corresponding guarantee, lose most of their fairness under the same heads (e.g., LAFTR-EO-CE reaches 31.9 on Adult with GBDT). The theoretical analysis is therefore consistent with the behavior that distinguishes DFL from these alternatives.

Comparison with HSIC and FairDisCo Under the Same Encoder (W4, Q3)

"Since the main design choice is the DC penalty, a DC vs. HSIC comparison under the same encoder is needed." / Q3: "Could you compare with FairDisCo or an HSIC-based method, like by swapping the DC penalty for an HSIC penalty in the same architecture?"

We conducted both comparisons on DFL's encoder, with the same grid and seeds (Appendix J.5, Table J.5, line 2647).

(a) HSIC in place of DC. We replaced the conditional DC penalty with a class-conditional unbiased HSIC penalty with a Gaussian kernel. We used three bandwidths (the median heuristic, and 0.1 and 10 times that value), with the HSIC term rescaled to match the DC term at initialization. Results on MOJI (p=2304p=2304):

DC (DFL)HSIC, medianHSIC, ×0.1HSIC, ×10STD-DenseNet
MCDP of jointly trained head17.120.917.716.843.5
New MLP-3 head: accuracy / TPR / MCDP73.25 / 21.22 / 21.0270.07 / 42.05 / 42.7470.21 / 42.39 / 42.6171.42 / 37.52 / 37.5170.11 / 43.30 / 43.21
dCor(Z,R∣Y)\mathrm{dCor}(Z,R\mid Y)0.0360.2250.2230.1650.183

The HSIC penalty yields jointly trained heads with small gaps. However, its representations retain as much information about ZZ as the raw inputs: dCor is 0.17–0.23, compared with 0.22 for the raw features. A new head consequently returns to the unconstrained level at every bandwidth.

This is consistent with the known sensitivity of Gaussian-kernel HSIC to the bandwidth in high dimension, which allows the encoder to satisfy the penalty while retaining ZZ. Distance covariance has no bandwidth, and through Theorem 4 it directly controls downstream disparity. On the tabular datasets, DFL also retains its advantage after head replacement: new-head MCDP is 4.6 versus 9.2 on Adult, and 5.1 versus 10.5 on Bank.

(b) FairDisCo-style estimator. Since FairDisCo is a variational autoencoder, we implemented a FairDisCo-style model on the same encoder. It uses an autoencoder with a ZZ-conditioned decoder, the same DC(R,Z)\mathrm{DC}(R,Z) penalty, a nine-point grid, and a linear head on the frozen representation.

  • On MOJI, it attains 59.8% accuracy, compared with 70.10% for STD-DenseNet and 73.45% for DFL.
  • On Adult and Bank, no weight satisfies the balanced-accuracy floor.
  • On Bank, its fairness does not persist after head replacement (MCDP 6.4 to 21.4, compared with 5.5 to 5.1 for DFL).

Taken together, (a) and (b) isolate the two design choices of DFL:

  1. distance covariance, rather than HSIC, makes the fairness persist in the representation;
  2. joint, task-aware estimation, rather than a task-agnostic encoder, preserves task-relevant information.

We thank the reviewer again for these precise suggestions, which have substantially strengthened the empirical support for the paper's central claim. Please let us know if there is any additional information or clarification we can provide to further support your understanding and confidence in our work.

Reviewer_npFdOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Deep Fair Learning that aims to create fair representations. The authors argue that many fairness methods constrain only the outputs and that leaves sensitive information in the features for the classifier to pick up. DFL trains a small encoder and a classifier together, with a distance covariance penalty that makes the features less dependent on the sensitive attribute. A second term rewards features that carry label information. The penalty has two versions, one for demographic parity and one for equalized odds. The theory gives convergence rates for the learned model and bounds the fairness gap of a limited class of new classifiers on the frozen features. Experiments on tabular, text, and image data show DFL achieves lower fairness gaps compared to other baselines.

Critical strengths
  1. The paper is well motivated. Learned features are often reused to train new classifiers, and fairness that is enforced only on the outputs might disappear after retraining. DFL targets the features to address this.
  2. Distance covariance is zero only when the features and the sensitive attributes are independent, so it also captures non linear dependence. The two versions of the penalty presented in the paper map into demographic parity and equalised odds.
  3. The experiments cover tabular, text, and image data and compare DFL with nine fairness methods. On Moji dataset, DFL has the lowest TPR.
  4. The paper tests whether its fairness gains survive retraining with a head-replacement experiment. After the classifier is replaced, the accuracy changes by less than one point.
Critical weaknesses
  1. The penalty is incremental since FairDisco uses distance covariance and Guo et al. [1] use distance correlation for fair representation.
  2. Table 5 tests head replacement only for DFL and only with a one-hidden layer classifier. So it does not show that DFL beats other representation methods or holds up under stronger classifiers.
  3. The comparison is uneven. STD is a logistic regression on Adult and Bank (H.5) rather than the DFL network. CFair-EO and LAFTR-EO-CE use one fixed weight on CelebA while DFL searches nine values of α\alpha .
  4. The ablation is limited. Section 4.4 compares the representation-level and output-level penalties on BIOS only and at a single tuning. There the representation-level penalty is more accurate (77.85 vs 64.78), but the output-level penalty has a lower TPR gap (5.12 vs 8.97). Removing the label-reward term changes accuracy and both gaps by less than one standard deviation (Table I.6), so it is unclear whether this term is needed.
  5. The main text is not self-contained. The metric formulas, the linear joint classifier, half of the comparisons only appear in the appendix. I would suggest moving them into the main text.

References

[1] Guo, Dandan, et al. "Learning fair representations via distance correlation minimization." IEEE Transactions on Neural Networks and Learning Systems 35.2 (2022): 2139-2152.

Questions for authors
  1. Why is age on Bank split into two groups at 25 instead of being used as a continuous attribute, given that Section 2 says distance covariance handles continuous sensitive attributes?
  2. Why did you choose a one-hidden-layer classifier for head replacement? Do you expect the fairness gains to hold for more complex classifiers?
10 comments

We thank the reviewer for the detailed review and for recognizing the motivation of the work, the properties of distance covariance, and the breadth of the benchmark. We address each comment below.

Novelty Relative to FairDisCo and Guo et al. (W1)

"The penalty is incremental since FairDisco uses distance covariance and Guo et al. use distance correlation for fair representation."

We agree that distance covariance itself is not new, and we do not claim it as a contribution. The contribution of DFL lies in how the dependence penalty is estimated (jointly with the prediction head) and in what is established about the resulting estimator. We examined this point empirically using the same encoder for all methods.

(1) Joint, task-aware estimation. FairDisCo learns a task-agnostic representation, a variational autoencoder with a distance-based penalty, and fits the predictor afterwards. We implemented a FairDisCo-style model on DFL's encoder. It uses an autoencoder with a decoder conditioned on ZZ, the same DC(R,Z)\mathrm{DC}(R,Z) penalty, and a linear head on the frozen representation (Appendix J.5, lines 2633–2642). With the identical dependence measure:

  • on MOJI, it reaches 59.8% accuracy, compared with 73.45% for DFL and 70.10% for the unconstrained network;
  • on Bank, its fairness does not persist after head replacement (MCDP 6.4 to 21.4), whereas that of DFL does (5.5 to 5.1).

(2) The regularized object. Guo et al. minimize the distance correlation of representations but provide no joint analysis and no guarantee for downstream reuse. Related methods (Huang & Liu; DISCO) penalize predictions. When the same DC penalty is applied to the model outputs, the jointly trained head is fair, but this fairness does not persist after head replacement. New-head MCDP is 18.9 versus 4.6 for DFL on Adult, and 22.4 versus 5.1 on Bank.

(3) Theoretical results specific to the joint estimator. To our knowledge, no prior distance-covariance method provides:

  • a joint excess-risk rate for the encoder–head pair that includes the composition term (Theorem 2);
  • recovery of the representation up to rigid motions, with predictive calibration (Theorem 3);
  • an end-to-end fairness bound for heads fitted on the learned encoder (Eq. 9).

The revised manuscript also shows that the inheritance bound is attained. The training penalty therefore equals, up to a constant, the worst-case disparity over the head class.

Head Replacement for All Methods and More Expressive Heads (W2, Q2)

"Table 5 tests head replacement only for DFL and only with a one-hidden-layer classifier." / Q2: "Why did you choose a one-hidden-layer classifier for head replacement? Do you expect the fairness gains to hold for more complex classifiers?"

The one-hidden-layer head was chosen because it matches the classifier used in the paper and because Theorem 4 concerns a smooth (RKHS) class of heads. We agree that a more informative test includes all methods and more expressive heads, and we have conducted it.

For every method, we froze the encoder and trained new heads with the predictive loss only and without access to ZZ: logistic regression, one- and three-layer MLPs, and gradient-boosted trees (GBDT). We also fitted a nonlinear probe for ZZ (Appendix J.3). Entries are accuracy / TPR / MCDP; the probe column reports AUC, where 50 is chance.

Adult:

EncoderJointly trained headMLP-3GBDTZZ-probe AUC
DFL75.96 / 1.56 / 3.5879.65 / 3.40 / 4.6280.51 / 2.79 / 7.5953.9
CFair-EO76.68 / 2.19 / 4.6081.18 / 3.86 / 8.6282.39 / 4.11 / 10.2258.7
LAFTR-EO-CE77.44 / 3.03 / 3.7482.02 / 8.49 / 31.6183.27 / 9.09 / 31.8978.1
Output-level DC76.88 / 2.48 / 5.3584.59 / 3.66 / 18.8985.55 / 4.18 / 19.5766.6
STD-DenseNet78.85 / 3.67 / 31.8785.48 / 6.84 / 33.0786.14 / 6.52 / 32.9275.4

Bank:

EncoderMLP-3GBDTZZ-probe AUC
DFL88.82 / 1.81 / 5.1189.58 / 0.92 / 6.6752.0
CFair-EO91.34 / 3.39 / 20.6391.13 / 2.69 / 20.8868.1
LAFTR-EO-CE90.59 / 2.98 / 21.7390.43 / 2.83 / 23.0570.2
STD-DenseNet91.45 / 3.36 / 25.0791.10 / 3.95 / 24.7268.0

MOJI and CelebA (MCDP of a new MLP-3 head):

  • MOJI: DFL 21.0, CFair-EO 25.7, output-level DC 26.8, LAFTR-EO-CE 41.6, HSIC 42.7, STD-DenseNet 43.2.
  • CelebA with ZZ = Male, mean of four tasks: DFL 17.6, CFair-EO 23.2, LAFTR-EO-CE 39.5, STD-DenseNet 38.0.

We can see that with GBDT heads, which lie outside any RKHS, DFL attains MCDP 7.59 (Adult) and 6.67 (Bank). The adversarial methods attain 10–23, and STD-DenseNet 25–33. Also, the fairness of the adversarial methods resides largely in their jointly trained heads. LAFTR-EO-CE moves from 3.74 to 31.61 on Adult after head replacement, and a probe recovers ZZ from its representation with AUC 78.

Comparability of Baselines and Tuning Budgets (W3)

"STD is a logistic regression on Adult and Bank (H.5) rather than the DFL network."

We have added STD-DenseNet, the DFL network without the fairness term (Appendix J.2, Table J.1, line 2540; Sec. 4.2, lines 419–421).

  • Adult. DFL reduces MCDP by 86% (31.87 to 4.39) at a cost of 0.5 points of accuracy (78.34% versus 78.85%).
  • Bank. DFL reduces MCDP by 54% (24.53 to 11.39) at a cost of 0.9 points.
  • MOJI. DFL is more accurate than the matched network (73.45% versus 70.10%) and reduces the TPR gap by 58% (43.30 to 17.98).

Recomputing the 95% accuracy floor from STD-DenseNet rather than STD-LR changes none of the selected weights on Adult or Bank. The weaker reference model therefore did not relax the comparison.

"CFair-EO and LAFTR-EO-CE use one fixed weight on CelebA while DFL searches nine values of α\alpha."

All methods now receive the same nine-point grid on every dataset, CelebA included, with the same validation rule.

MOJI. DFL is the most accurate method (73.45%) and reduces the TPR gap by more than half. When every grid is extended by three fair-end weights, DFL attains TPR 2.88 and MCDP 5.07 at 67.3% accuracy. At the same accuracy, CFair-EO attains 14.58 and 15.19.

CelebA. We acknowledge that, with the larger tuning budget, CFair-EO attains lower gaps than DFL at the jointly trained head. This fairness, however, resides largely in the head rather than in the representation: a probe recovers gender from CFair-EO's representation with AUC 82–88 on three of the four tasks, compared with 64–65 for DFL. The table reports the gap (mean of TPR and MCDP) of the jointly trained head and of a new MLP-3 head:

TaskDFLCFair-EO (tuned)LAFTR-EO-CE (tuned)STD-DenseNet
Young14.0 → 12.88.0 → 13.915.0 → 34.832.1 → 30.9
Attractive13.4 → 13.112.9 → 22.428.0 → 42.740.2 → 40.3
Smiling7.0 → 7.06.1 → 6.77.2 → 12.412.1 → 12.0
Wavy Hair17.6 → 17.010.3 → 24.821.2 → 35.836.5 → 36.6

After head replacement, DFL attains the smaller gap on three of the four tasks and a comparable gap on Smiling. The gap of CFair-EO nearly doubles on Attractive and Wavy Hair, and LAFTR-EO-CE returns to the level of the unconstrained model. The gaps of DFL remain unchanged or decrease. Since the paper concerns fair representations that are reused by downstream predictors, we regard this as the relevant comparison. It is summarized in the revised Sec. 4.2.

Ablations Across Datasets and Operating Points (W4)

"Section 4.4 compares the representation-level and output-level penalties on BIOS only and at a single tuning … it is unclear whether this term [the label reward] is needed."

We agree that a single tuning on one dataset is insufficient. We have replaced it with a matched ablation on MOJI, Adult, and Bank over the full nine-point grid of α\alpha, with five seeds and head replacement (Appendix J.4). Each cell reports the jointly trained head (accuracy / TPR / MCDP) and, after the semicolon, the MCDP of a new MLP-3 head:

DFLDFL without label rewardOutput-level DC
MOJI73.45 / 17.98 / 17.06; 21.068.99 / 26.66 / 26.26; 34.372.94 / 23.16 / 22.83; 26.8
Adult78.34 / 3.17 / 4.39; 4.678.66 / 4.44 / 5.68; 22.178.52 / 4.59 / 6.19; 18.9
Bank90.22 / 5.19 / 11.39; 5.189.83 / 5.38 / 15.38; 22.290.73 / 2.46 / 13.62; 22.4

Necessity of the label-reward term.

  • Accuracy. On MOJI, the reward increases accuracy by 4.5 points and reduces the TPR gap by 8.7 points. DFL is more accurate than its no-reward variant at each of the nine values of α\alpha, by 2.4–4.5 points.
  • Fairness across the grid. On Adult, MCDP over the whole grid ranges from 4.4 to 5.6 with the reward and from 5.7 to 17.1 without it, at comparable accuracy.
  • Persistence. Without the reward, new-head MCDP increases from 4.6 to 22.1 (Adult) and from 5.1 to 22.2 (Bank). The ZZ-probe AUC increases from 53.9 to 67.7 (Adult) and from 67.0 to 83.2 (MOJI).

Our interpretation is that the reward preserves label-relevant structure in RR, so that the fairness penalty removes information about ZZ rather than task information.

Representation-level versus output-level penalty.

  • On MOJI, the representation-level penalty is better on all three metrics (+0.5 accuracy, −5.2 TPR, −5.8 MCDP), and it is never less accurate at any of the nine values of α\alpha.
  • On Adult and Bank, the two penalties yield similar jointly trained heads, but only the representation-level penalty retains its fairness after head replacement: 4.6 versus 18.9 on Adult, and 5.1 versus 22.4 on Bank.

On BIOS, the two penalties reach similar accuracy under the shared training procedure. The multi-dataset results above therefore replace the single-tuning comparison of Sec. 4.4 and Table I.6.

Self-Contained Main Text (W5)

"The metric formulas, the linear joint classifier, half of the comparisons only appear in the appendix."

We agree that the principal evidence should appear in the main text. Within the ten-page limit for revised submissions, the main text now includes:

  • the comparison of all methods under head replacement, which was absent from the original main text;
  • the matched evaluation protocol (Sec. 4.1, lines 403–408): STD-DenseNet, the majority predictor, equal nine-point budgets, and the balanced-accuracy floor;
  • the comparisons with STD-DenseNet and under equal budgets on CelebA;
  • a summary of the results for continuous sensitive attributes.

In the final version we will include more details in the main text to help improve the clarity of the paper.

Continuous Sensitive Attributes — Setup and Theory (Q1)

"Why is age on Bank split into two groups at 25 instead of being used as a continuous attribute, given that Section 2 says distance covariance handles continuous sensitive attributes?"

We thank the reviewer for this question. We used the threshold of 25 because it is the standard group definition for the Bank benchmark, and because the paper's metrics (TPR-RMS and MCDP) are defined for caregrorical sensitive attribute. DFL itself requires no modification for a continuous attribute. We have therefore evaluated DFL with ZZ equal to age on Adult and Bank, following the definitions, metrics, and baselines of the literature on continuous sensitive attributes (Appendix J.6, lines 2643–2680).

Definitions. For the score s^=P(Y=1∣X)\hat s=P(Y{=}1\mid X), independence requires s^⊥Z\hat s\perp Z, and separation requires s^⊥Z∣Y\hat s\perp Z\mid Y (Mary et al., ICML 2019; Grari et al., IJCAI 2020; Jiang et al., ICLR 2022).

Metrics (test set, ×100):

  • GDP =EZ,∣E[s^∣Z]−E[s^]∣=\mathbb E_Z\\,|\mathbb E[\hat s\mid Z]-\mathbb E[\hat s]|, the generalized demographic parity of Jiang et al. Their histogram estimator and the FairQuant metric of Grari et al. give the same ranking of methods.
  • GEO =∑yP(Y=y),GDP(s^,Z∣Y=y)=\sum_y P(Y{=}y)\\,\mathrm{GDP}(\hat s,Z\mid Y{=}y), its class-conditional analogue for separation.

Baselines. All baselines use DFL's encoder, the same grid, five seeds, and the same validation rule, with class-conditional versions used for separation:

  • the GDP regularizer (Jiang et al.);
  • the χ2\chi^2 upper bound on HGR maximal correlation (Mary et al.);
  • FREM (Kong, Kim & Kim, 2025), an expected integral probability metric between P(R∣Z)P(R\mid Z) and P(R)P(R);
  • an adversary that regresses age on RR (Louppe et al., 2017; Madras et al., 2018).

DFL. The objective is unchanged: DC(Z,R)\mathrm{DC}(Z,R) for independence and cDC(Z,R∣Y)\mathrm{cDC}(Z,R\mid Y) for separation, with ZZ equal to age in years. Since DC(aZ,R)=∣a∣,DC(Z,R)\mathrm{DC}(aZ,R)=|a|\\,\mathrm{DC}(Z,R), standardizing age only rescales the penalty. Age is removed from the input features. These experiments comprise 1,220 additional runs.

Theory (remark after Theorem 4, Sec. 3.3, lines 340–344). For distance-induced kernels, DC=4,HSIC\mathrm{DC}=4\\,\mathrm{HSIC} (Sejdinovic et al., 2013). Consequently, for every head with ∣h∣HR≤M\\|h\\|_{\mathcal H_R}\le M and every ff with ∣f∣HZ≤1\\|f\\|_{\mathcal H_Z}\le 1, ∣Cov(f(Z),h(R))∣≤M2DC(Z,R).|\mathrm{Cov}(f(Z),h(R))|\le \tfrac M2\sqrt{\mathrm{DC}(Z,R)}. Proof. Cov(f(Z),h(R))=⟨h,CRZf⟩≤∣h∣,∣CRZ∣op∣f∣≤M∣CRZ∣HS=MHSIC=M2DC\mathrm{Cov}(f(Z),h(R))=\langle h,C_{RZ}f\rangle\le\\|h\\|\\,\\|C_{RZ}\\|_{\mathrm{op}}\\|f\\|\le M\\|C_{RZ}\\|_{\mathrm{HS}}=M\sqrt{\mathrm{HSIC}}=\tfrac M2\sqrt{\mathrm{DC}}. □\square

For binary ZZ, this inequality reduces to Theorem 4. For real-valued ZZ, the distance-induced RKHS satisfies ∣f∣HZ2=∫f′(z)2,dz\\|f\\|^2_{\mathcal H_Z}=\int f'(z)^2\\,dz. A single penalty therefore simultaneously controls the covariance between the head and every function of age with bounded energy, including smoothed versions of every threshold indicator 1age≥t\mathbf 1\\{\text{age}\ge t\\}.

Empirical counterpart. On Bank, a single DFL model trained with continuous age reduces MCDP at all six age thresholds, none of which is used during its training (Table J.7, line 2693):

Modeltt=253035405060Mean
STD-DenseNet24.5313.256.914.667.1951.1217.94
DFL, Z=1age≥25Z=\mathbf 1\\{\text{age}\ge25\\}12.986.422.551.174.7525.928.97
DFL, ZZ = age11.925.652.591.764.5921.998.08

The continuous model attains a lower disparity than the binary model even at t=25t=25, the threshold on which the binary model was trained. The largest improvement is for older clients (t=60t=60), a group that the binary encoding does not target.

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

The paper proposes Deep Fair Learning (DFL), a method to learn fair representations. It learns an encoder and classification head and uses distance covariance (DC) penalties to enforce fairness and align representations with the targets. In particular, DC between sensitive attribute and encoded representation is penilised and DC between target and encoded representation is rewarded.

Critical strengths
  • The paper is easy to follow and the method is simple, with no adversarial or complicated training procedures.
  • Provides theoretical guarantees for the method.
  • The evaluation covers varied modalities.
Critical weaknesses
  • Some relevant related work seems to be missing:

    • [1] uses a similar penalty structure, with the same u-statistic estimate and O(n^{(-\frac{\beta}{2beta+p} risk bound.
    • [3] shows that the DP gap over an RKHS ball equals a scaled MMD, which is similar to Theorems 4.
    • [5,6,7] also use DC for fairness or bias mitigation.
  • The guarantees are too narrow and not practically useful compared to existing bounds (like the ones in [8]). For the datasets tested the convergence rate of the bound is too slow to be actionable and the assumptions needed are not verifiable for NN.

  • The evaluation only selects a single operating point rather than reporting accuracy–fairness trade-off curves. It also does not include any competitor from 2024 onwards.

[1] X. Zhou, C. Y. Ip, J. Huang. Fair Sufficient Representation Learning. arXiv:2504.01030, 2025.

[3] Y. Ni, X. Huo. Kernel-based Equalized Odds: A Quantification of Accuracy-Fairness Trade-off in Fair Representation Learning. NeurIPS 2025.

[5] R. Huang, H. Liu. Bridging Fairness Gaps: A (Conditional) Distance Covariance Perspective in Fairness Learning. arXiv:2412.00720, 2024.

[6] D. Guo, C. Wang, B. Wang, H. Zha. Learning Fair Representations via Distance Correlation Minimization. IEEE TNNLS, 2022.

[7] E. Kavak, T. N. Wolf, C. Wachinger. DISCO: Mitigating Bias in Deep Learning with Conditional Distance Correlation. ICML 2026.

[8] U. Gupta, A. Ferber, B. Dilkina, G. Ver Steeg. Controllable Guarantees for Fair Outcomes via Contrastive Information Estimation. AAAI 2021.

Questions for authors
  • Can you discuss how [1] and [3] relate to this work?
  • Why are the proposed DC guarantee preferable over certificates or bounds applicable to more generic heads [8, 9, 10]?
  • Can you include results for an "unfair" (architecture matched) baseline and for the optimal input agnostic model (i.e. majority class in Adul, Bank, etc.).
  • What is the rationale used to select the competitors to compare against?
12 comments

We thank the reviewer for the positive assessment of the clarity and simplicity of the method and the breadth of the evaluation, and for the helpful references. We hope our response addresses your concerns.

Related Work and Novelty (W1, Q1)

We thank the reviewer for the positive assessment of the clarity and simplicity of the method and the breadth of the evaluation, and for the helpful references. All of the suggested works are now discussed in the revised manuscript (Sec. 1, lines 56–58; App. A, lines 868–887). The distinction is the same in each case: DFL estimates the encoder and the prediction head jointly, and its guarantees concern the fairness of heads fitted after training, including the effect of estimation error.

"[1] uses a similar penalty structure, with the same u-statistic estimate and O(n−β/(2β+p))O(n^{-\beta/(2\beta+p)}) risk bound."

FSRL (Zhou, Ip & Huang) is a two-stage method. It first learns a representation using a label-dependence reward, a sensitive-dependence penalty, and an energy-distance term toward a Gaussian reference, and then fits a predictor on the frozen representation. DFL instead estimates the pair (g,f)(g,f) jointly, which leads to results without a counterpart in FSRL:

  • the excess-risk bound of Theorem 2 contains a composition term, through which encoder error propagates into the head with factor sf=min⁡βf,1s_f=\min\\{\beta_f,1\\};
  • Theorem 3 establishes recovery of the representation up to rigid motions, together with predictive calibration;
  • Eq. (9) provides an end-to-end fairness bound for heads fitted on the learned encoder.

The degree-four U-statistic and the general form of the nonparametric rate are standard tools for distance-covariance estimators. We do not claim them as contributions, and the revised manuscript states this explicitly.

"[3] shows that the DP gap over an RKHS ball equals a scaled MMD, which is similar to Theorem 4."

We agree that Theorem 4 and the result of Ni & Huo rest on the same RKHS argument, and we now cite their work accordingly. For the distance-induced kernel, the bound of Theorem 4 is in fact attained (remark after Theorem 4, lines 340–344): sup⁡∣h∣≤M∣E[h(R)∣Z=1]−E[h(R)∣Z=0]∣=M∣μ1−μ0∣=M2p0p1DC(Z,R),\sup_{\\|h\\|\le M}\big|\mathbb E[h(R)\mid Z=1]-\mathbb E[h(R)\mid Z=0]\big| = M\\|\mu_1-\mu_0\\| = \frac{M}{2p_0p_1}\sqrt{\mathrm{DC}(Z,R)}, with maximizer h⋆∝μ1−μ0h^\star\propto\mu_1-\mu_0, where μz\mu_z are the group-conditional mean embeddings.

Proof. By the reproducing property, E[h(R)∣Z=z]=⟨h,μz⟩\mathbb E[h(R)\mid Z=z]=\langle h,\mu_z\rangle, so the gap equals ⟨h,μ1−μ0⟩\langle h,\mu_1-\mu_0\rangle, and the Cauchy–Schwarz bound is attained at h⋆h^\star. For binary ZZ, the distance-induced kernel is kZ(z,z′)=zz′k_Z(z,z')=zz'. The cross-covariance operator CRZC_{RZ} therefore has rank one, and ∣CRZ∣HS=p0p1∣μ1−μ0∣\\|C_{RZ}\\|_{\mathrm{HS}}=p_0p_1\\|\mu_1-\mu_0\\|. Since DC=4,HSIC=4∣CRZ∣HS2\mathrm{DC}=4\\,\mathrm{HSIC}=4\\|C_{RZ}\\|^2_{\mathrm{HS}} for distance-induced kernels (Sejdinovic et al., 2013), the second equality follows. □\square

This has two implications.

  1. The training penalty of DFL equals, up to a constant, the worst-case disparity over the head class; it is not merely an upper-bounding surrogate.
  2. The contribution of DFL relative to [3] lies on the estimation side. Eq. (9) extends this population identity to an encoder learned from data by accounting for the joint excess risk Δn\Delta_n and the representation error d∼d_\sim.

The same argument extends Theorem 4 to continuous ZZ (lines 340–344; see our response to Reviewer npFd, part 6).

"[5,6,7] also use DC for fairness or bias mitigation."

Huang & Liu [5] and DISCO [7] apply (conditional) distance-covariance penalties to predictions. Guo et al. [6] minimize the distance correlation of representations but do not provide guarantees for downstream reuse. Our new experiments indicate that the regularized object is consequential. An output-level DC penalty, in the style of [5] and [7] and trained on DFL's encoder, loses its fairness after head replacement: Adult MCDP increases from 5.35 to 18.89, and Bank MCDP from 12.53 to 22.44. The representation-level penalty of DFL retains it: 3.58 to 4.62 on Adult and 5.50 to 5.11 on Bank (Table J.4, line 2619). DISCO (ICML 2026) is recent concurrent work, which we now discuss in the related-work section.

Practical Value of the Guarantees (W2, Q2)

"For the datasets tested the convergence rate of the bound is too slow to be actionable and the assumptions needed are not verifiable for NN."

Dependence on dimension. The exponent in Theorem 2 reflects a worst case in the ambient dimension pp, which enters only through the approximation of Hölder functions on [0,1]p[0,1]^p. When the inputs concentrate near a low-dimensional set, approximation results for deep ReLU networks (Nakada & Imaizumi, JMLR 2020; Chen et al., 2022) allow pp to be replaced by an intrinsic dimension dd. We estimated dd with the TwoNN estimator on three subsamples of 10,000 points:

MOJIBIOSCelebAAdultBank
ambient dimension pp230476851210162
intrinsic dimension dd32.032.144.78.09.1

For β=2\beta=2 on MOJI, the exponent increases from 0.0009 to 0.056, a factor of approximately 64. We have added this remark after Theorem 2 (lines 252–255).

Empirical convergence. We trained DFL on nested subsamples of MOJI (three seeds each) and measured the alignment distance d∼d_\sim of Theorem 3 to the encoder trained on the full sample. The distance decreases from 0.83 at n=3,125n=3{,}125 to 0.61 at n=12,500n=12{,}500 and 0.57 at n=50,000n=50{,}000. The last value coincides with the distance between two full-sample encoders trained with different seeds (0.57). The learned representation thus reaches the level of optimization noise well before the full sample size.

Assumptions of Theorem 4. Theorem 4 requires only a bounded representation, which the clipping step enforces, and non-degenerate group probabilities. Identifiability and smoothness are used only for the separate estimation and recovery results. Empirically, the persistence predicted by Theorem 4 holds for MLP heads and also for gradient-boosted trees: Adult MCDP 7.59 and Bank 6.67, compared with 32.9 and 24.7 for STD-DenseNet.

"Why are the proposed DC guarantees preferable over certificates or bounds applicable to more generic heads [8, 9, 10]?"

We can only find [8] in reviewer's comments, and we will mainly discuss this paper. We view the two approaches as complementary, and the revised Appendix A (lines 880–884) discusses it. We note three properties of the DC guarantee.

  1. Estimability. For [0,1][0,1]-valued heads, the worst-case disparity over all measurable heads equals the total variation distance between the group-conditional distributions of the representation. Total variation between high-dimensional continuous distributions cannot be estimated consistently without density assumptions. Certificates that cover all heads must therefore bound an auxiliary quantity, such as the mutual-information upper bound in [8], or restrict the class of encoders. Example F.5 shows that some restriction of the head class is necessary for any DC-based bound, whereas under exact independence Proposition F.4 covers every measurable head.
  2. Exactness and computability. The DC bound is attained over its head class (lines 340–344; proof in part 1 of our response). The certified quantity is the training statistic itself, an unbiased U-statistic that requires no density estimation, mutual-information bound, adversary, or bandwidth.
  3. Applicability to learned encoders. Through Eq. (9), the guarantee applies to an unrestricted deep encoder estimated from data, including estimation error. Empirically, the restriction to an RKHS head class does not appear to be binding: the fairness of DFL persists for gradient-boosted trees (General Response, part 1).

Trade-off Curves and Choice of Baselines (W3, Q4)

"The evaluation only selects a single operating point rather than reporting accuracy–fairness trade-off curves."

We now report the full nine-point trade-off curve of every method, under equal tuning budgets (Appendix J.2–J.3).

Fronts of the jointly trained heads (Figure J.1, line 2569).

  • On Adult, DFL's front lies below those of CFair-EO and LAFTR-EO-CE over its entire accuracy range.
  • On MOJI, where every method was additionally given three weights at the fair end of its grid, DFL's front lies below those of the adversarial and output-level methods. At comparable accuracy:
MOJI, at comparable accuracyAccuracyTPRMCDP
DFL (α=0.02\alpha=0.02)67.302.885.07
CFair-EO (γ=316\gamma=316)67.3714.5815.19
LAFTR-EO-CE (γ=316\gamma=316)67.4224.1628.55

At 71.3% accuracy (α=0.05\alpha=0.05), DFL already attains a smaller gap than every adversarial operating point at any accuracy. Increasing the adversarial weights further does not extend the adversarial trade-offs: the gaps of LAFTR-EO-CE increase to 26–32 at large weights, and the fairest point of CFair-EO is both less accurate and less fair than DFL at α=0.05\alpha=0.05.

Fairness after head replacement across the grid (Figure J.2, line 2606). We also report fairness after head replacement over the entire grid, which reflects the setting targeted by the paper. On Adult, the mean MCDP of a new three-layer MLP head on DFL's encoders is 8.2, compared with 22.9 for CFair-EO, 28.6 for LAFTR-EO-CE, and 26.5 for the output-level penalty. On Bank, it is 10.8 compared with 20.7–22.6, and each of the nine DFL encoders is fairer than every adversarial and output-level encoder.

"It also does not include any competitor from 2024 onwards." / Q4: "What is the rationale used to select the competitors?"

Since the paper concerns where sensitive information is controlled, the baselines were chosen to represent the main alternatives:

  • prediction-level fairness: AdvDebias, FairMixup, DRAlign, and DiffMCDP (ICML 2024);
  • post-hoc representation editing: INLP, RLACE, and SUP (NAACL 2024);
  • joint adversarial representation learning with the same equalized-odds target: CFair-EO and LAFTR-EO-CE.

The revision adds the following comparisons, all on DFL's encoder and with equal tuning budgets:

  • an HSIC penalty at three bandwidths;
  • a FairDisCo-style task-agnostic estimator with the same DC penalty;
  • an output-level DC penalty, as in Huang & Liu and DISCO (ICML 2026);
  • four methods for continuous sensitive attributes, including FREM (2025) and the GDP regularizer (ICLR 2022), described in our response to Reviewer npFd.

After head replacement, DFL attains lower disparities than each of these alternatives (Appendix J.3–J.6). The rationale is now stated in Sec. 4.1 (lines 403–405).

Architecture-Matched and Input-Agnostic Baselines (Q3)

"Can you include results for an 'unfair' (architecture matched) baseline and for the optimal input agnostic model (i.e. majority class in Adult, Bank, etc.)?"

We have added both baselines.

  • STD-DenseNet is DFL's network with the fairness term removed, trained with the same procedure and seeds.
  • Majority is the input-agnostic majority-class predictor.

Test results are averaged over five seeds (Appendix J.2, Table J.1, line 2540; Sec. 4.2, lines 419–421):

Majority (accuracy)STD-DenseNet (accuracy / TPR / MCDP)DFL (accuracy / TPR / MCDP)
Adult74.5778.85 / 3.67 / 31.8778.34 / 3.17 / 4.39
Bank88.6591.09 / 2.29 / 24.5390.22 / 5.19 / 11.39
MOJI50.0070.10 / 43.30 / 43.4773.45 / 17.98 / 17.06

Relative to its own architecture without the fairness term, DFL reduces MCDP by 86% on Adult and 54% on Bank, at a cost of 0.5 and 0.9 points of accuracy. On MOJI, DFL is more accurate than the unconstrained network and reduces both gaps by approximately 60%.

The majority-class baseline also led us to strengthen the selection protocol. An accuracy-only floor can therefore admit near-constant predictors for any method. All matched results therefore use an additional balanced-accuracy floor: 95% of STD-DenseNet's validation balanced accuracy, applied on validation data only.

Under this stricter rule, DFL attains the smallest gap among the nine fairness methods on Adult (3.78, compared with 3.95 for LAFTR-EO-CE, 4.04 for HSIC, and 5.22 for CFair-EO). On Bank, the best methods are within one standard deviation of each other. Replacing the STD-LR floor by the STD-DenseNet floor changes none of the selected weights on Adult or Bank.

We thank the reviewer again for these suggestions. We believe the additional experiments provide the empirical counterpart of the theoretical analysis: fairness enforced on the representation persists when the learned features are reused, whereas output-level or adversarial control largely does not.

Comments on the paper

1 comment

Dear AC and reviewers

We sincerely thank all three reviewers for their careful reading and constructive comments. We have run all of the requested experiments and uploaded a revised manuscript in which every change is shown in blue. The main additions are:

  • a new Section 4.5 with Table 6;
  • brief additions in Sections 1–4;
  • a new Appendix J, with six subsections, two figures, and seven tables.

The reviewers' requests fall into five groups, each addressed in the individual responses:

  1. an architecture-matched unconstrained model and the majority-class predictor (PCnM Q3, npFd W3, 3bPd W1/Q1);
  2. equal tuning budgets and accuracy–fairness trade-off curves (PCnM W3, npFd W3, 3bPd W2/Q2);
  3. head replacement for all methods and with more expressive heads (npFd W2/Q2, 3bPd W3);
  4. comparisons with HSIC and FairDisCo (npFd W1, 3bPd W4/Q3);
  5. continuous sensitive attributes (npFd Q1).

Principal finding: persistence of fairness under head replacement. We froze the encoder of every method and trained a new three-layer MLP head with the predictive loss only. This head has no fairness term and no access to ZZ. The table reports the test MCDP of the new head at each method's validation-selected weight. The MCDP of the jointly trained head is in parentheses (five seeds).

DFLCFair-EOLAFTR-EO-CEOutput-level DCHSICSTD-DenseNet
Adult4.6 (3.6)8.6 (4.6)31.6 (3.7)18.9 (5.4)9.2 (3.0)33.1
Bank5.1 (5.5)20.6 (11.1)21.7 (9.1)22.4 (12.5)10.5 (8.9)25.1
MOJI21.0 (17.1)25.7 (17.4)41.6 (23.9)26.8 (22.8)42.7 (20.9)43.2
CelebA (mean of 4 tasks)17.6 (18.2)23.2 (15.2)39.5 (24.4)––38.0

Several competing methods attain small gaps at their own jointly trained head but lose most of this fairness once the head is replaced, and this pattern does not depend on a particular operating point or head class:

  • Across the full weight grid, the mean new-head MCDP of DFL on both Adult and Bank are the lowest compared with CFair-EO, LAFTR-EO-CE, and the output-level penalty.
  • With gradient-boosted trees, a head class outside any RKHS, DFL attains MCDP 7.59 (Adult) and 6.67 (Bank). CFair-EO attains 10.22 and 20.88, and LAFTR-EO-CE 31.89 and 23.05.
  • A nonlinear probe recovers ZZ least accurately from DFL's representation:
    • Adult, AUC 53.9;
    • Bank, 52.0;
    • MOJI, 67.0, versus 72.7 for CFair-EO and 92.5 for LAFTR-EO-CE;
    • all four CelebA tasks, 55–65, versus 60–88 for CFair-EO and about 97 for LAFTR-EO-CE.

These results support the central claim of the paper: fairness enforced at the level of the representation persists when the learned features are reused by downstream predictors.

Ablations .

The label-reward term. On MOJI it increases accuracy by 4.5 points and reduces the TPR gap by 8.7 points. DFL is more accurate than its no-reward variant at every one of the nine values of α\alpha. Without the reward, new-head MCDP increases from 4.6 to 22.1 (Adult) and from 5.1 to 22.2 (Bank).

Representation-level versus output-level penalty. The same DC penalty applied to the model outputs yields a fair jointly trained head, but this fairness does not persist after head replacement. New-head MCDP is 18.9 versus 4.6 on Adult, and 22.4 versus 5.1 on Bank.

Comparison with HSIC and FairDisCo

HSIC. On MOJI, replacing DC with HSIC yields seemingly fair jointly trained heads at all three bandwidths. However, a new head returns to the unconstrained level (MCDP 37.5–42.7 versus 43.2). In addition, dCor(Z,R∣Y)\mathrm{dCor}(Z,R\mid Y) remains close to that of the raw inputs (0.17–0.23 versus 0.22). DFL retains its improvement (new-head MCDP 21.0; dCor 0.036), and distance covariance has no bandwidth to tune.

FairDisCo-style estimator. A task-agnostic estimator with the same DC penalty loses task information: 59.8% accuracy on MOJI, versus 73.45% for DFL. On Bank, its fairness does not persist after head replacement (MCDP 6.4 to 21.4).

Continuous sensitive attributes. With ZZ equal to age on Adult and Bank, we compared DFL with four methods from the literature on continuous sensitive attributes. DFL is used with its objective unchanged and attains the lowest violation in all four settings (independence and separation on each dataset). Details are given in our response to Reviewer npFd (parts 6–7).

Theory. Two remarks were added to the revised manuscript.

  • After Theorem 4: the inheritance bound of Theorem 4 holds with equality over its head class, and the same argument extends to continuous ZZ.
  • After Theorem 2: the convergence rate depends on the intrinsic rather than the ambient dimension of the data. The estimated intrinsic dimension of MOJI and BIOS is approximately 32.

Fianlly, we thank all the reviewers for thier constructive feedback. We are happy to provide additional information and clarification to help you understand our work.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.