- 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
-
The paper studies an important problem, and the presentation is easy to follow.
-
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
-
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.
-
Tuning strategy differs across methods. DFL searches nine values of , 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.
-
The theory, which the paper claims as the main contribution, is weakly linked to the experiments. In Theorem 2, the rates have exponents like . With for BIOS and for MOJI, the bound shrinks extremely slowly as grows. Also, Theorem 4 covers binary and RKHS heads, while the fresh head in Table 5 is a one-hidden-layer network.
-
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
-
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.
-
Could you clarify the tuning budget for each baseline? Could the adversarial baselines get the same nine-point grid search?
-
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
Connection Between Theory and Experiments (W3)
"In Theorem 2, the rates have exponents like . With for BIOS and for MOJI, the bound shrinks extremely slowly."
Ambient versus intrinsic dimension. The exponent reflects the ambient dimension , which enters only through the approximation of Hölder functions on . For inputs supported near a low-dimensional set, approximation results for deep ReLU networks (Nakada & Imaizumi, 2020; Chen et al., 2022) allow to be replaced by the intrinsic dimension. TwoNN estimates of the intrinsic dimension are approximately:
- 32 for MOJI ();
- 32 for BIOS ();
- 45 for CelebA ().
For , 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 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 to 0.57 at . 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 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 . The same argument yields for every in the unit ball of the distance-induced RKHS of (lines 340–344). Binary recovers Theorem 4, and continuous 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 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 ():
| DC (DFL) | HSIC, median | HSIC, ×0.1 | HSIC, ×10 | STD-DenseNet | |
|---|---|---|---|---|---|
| MCDP of jointly trained head | 17.1 | 20.9 | 17.7 | 16.8 | 43.5 |
| New MLP-3 head: accuracy / TPR / MCDP | 73.25 / 21.22 / 21.02 | 70.07 / 42.05 / 42.74 | 70.21 / 42.39 / 42.61 | 71.42 / 37.52 / 37.51 | 70.11 / 43.30 / 43.21 |
| 0.036 | 0.225 | 0.223 | 0.165 | 0.183 |
The HSIC penalty yields jointly trained heads with small gaps. However, its representations retain as much information about 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 . 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 -conditioned decoder, the same 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:
- distance covariance, rather than HSIC, makes the fairness persist in the representation;
- 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.
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)
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):
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)
Yes. All methods now receive the same nine-point grid on every dataset, including CelebA:
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):
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.