- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper focus on fair dataset distillation. Existing PA-aware methods differ in their class-wise alignment targets, yet all represent each class as a single point in feature space. This can miss differences in how widely or unevenly each group’s data is distributed. Authors propose Fair Geometric Distillation (FGD), which uses small, group-specific sets of real examples as targets for the synthetic data. FGD achieves sustainable improvement across six benchmarks. Main innovation of this paper is reframing fair dataset distillation as a distribution-geometry problem: matching group means alone can miss within-group variation, which may contribute to fairness failures on complex data.
Critical strengths
1 This paper shows a clear limitation of exisiting fairness-aware dataset distillation: aligning each group to a single feature-space target may overlook within-group variation. A complete theoretical analysis further supports this argument.
2 The evaluation experiment examines both group-level geometric metrics and downstream fairness across selected six benchmarks. FGD improves the reported geometric metrics across the datasets and achieves the lowest maximum equalized-odds gap in 7 of 9 face-dataset and IPC settings. This result provides useful evidence that the proposed approach FGD can improve fairness-related outcomes in practice, while making its accuracy–fairness trade-offs visible.
Critical weaknesses
Major:
1 Paper’s geometric claim is not fully supported by the proposed objective. In equation 8, synthetic examples are matched one-sidedly to their nearest target medoids, with a group-mean regularizer. But this process does not ensure that all medoids are represented or that within-group covariance is matched.
2 The experimental evidence does not make the robustness of the fairness improvements clear. The main tables report means over five seeds without variability estimates, while the appendix notes that some control differences are within seed noise.
3 Some equations need further explanations.
Minor:
1 An additional visualization could make the proposed method easier to follow.
2 Additional experiments to substantiate the paper’s central claims could better shape this paper. Experiments that can directly assess within-group distribution coverage can be beneficial.
Questions for authors
1 Several equations would benefit from additional clarification. First, for Eq. 8, please specify the direction of the nearest-medoid matching. Do multiple synthetic examples match the same medoid? How the group-mean regularizer is computed and weighted?
2 How does FGD perform as protected-group sizes become more imbalanced or sparse?
2 comments
Major 2: Variability
We add seed spreads to the main table. Bold: lowest DEO in each column.
| Method | IPC | CelebA DEO$_M$ ↓ | CelebA DEO$_A$ ↓ | CelebA Acc ↑ | bFFHQ DEO$_M$ ↓ | bFFHQ DEO$_A$ ↓ | bFFHQ Acc ↑ | UTKFace DEO$_M$ ↓ | UTKFace DEO$_A$ ↓ | UTKFace Acc ↑ |
|---|---|---|---|---|---|---|---|---|---|---|
| DM | 10 | 30.01 ± 2.47 | 28.85 ± 2.26 | 60.97 ± 1.13 | 44.32 ± 1.33 | 34.76 ± 0.17 | 64.90 ± 0.68 | 22.29 ± 2.25 | 18.64 ± 2.06 | 61.24 ± 0.29 |
| DM | 30 | 34.45 ± 2.89 | 32.67 ± 1.68 | 62.45 ± 0.71 | 53.68 ± 1.54 | 42.92 ± 0.97 | 65.02 ± 0.09 | 22.87 ± 1.44 | 20.46 ± 1.24 | 63.48 ± 1.29 |
| DM | 50 | 40.26 ± 2.57 | 38.81 ± 2.14 | 63.57 ± 0.78 | 60.88 ± 1.93 | 50.56 ± 0.90 | 65.28 ± 0.19 | 22.50 ± 1.71 | 20.31 ± 1.86 | 64.97 ± 1.15 |
| FairDD | 10 | 9.37 ± 1.62 | 5.37 ± 0.73 | 64.37 ± 0.20 | 15.84 ± 1.81 | 10.68 ± 1.23 | 65.46 ± 1.38 | 18.37 ± 1.40 | 14.44 ± 1.10 | 61.98 ± 0.49 |
| FairDD | 30 | 11.81 ± 0.93 | 8.02 ± 1.08 | 66.10 ± 0.84 | 17.96 ± 1.89 | 11.44 ± 1.20 | 68.46 ± 0.54 | 19.06 ± 1.60 | 16.56 ± 0.91 | 63.98 ± 1.61 |
| FairDD | 50 | 14.08 ± 1.19 | 9.87 ± 1.67 | 68.50 ± 0.84 | 19.76 ± 1.45 | 14.96 ± 1.74 | 69.00 ± 1.14 | 18.99 ± 0.72 | 16.43 ± 1.20 | 65.19 ± 0.41 |
| COBRA | 10 | 10.00 ± 0.81 | 6.36 ± 1.07 | 64.94 ± 0.19 | 20.24 ± 0.81 | 11.40 ± 0.74 | 68.30 ± 0.34 | 21.00 ± 1.64 | 17.43 ± 0.75 | 49.95 ± 0.45 |
| COBRA | 30 | 10.65 ± 1.04 | 5.98 ± 0.56 | 67.04 ± 0.85 | 16.56 ± 1.99 | 10.56 ± 2.73 | 70.56 ± 0.77 | 21.80 ± 1.40 | 16.93 ± 1.04 | 50.97 ± 0.55 |
| COBRA | 50 | 15.32 ± 1.77 | 10.65 ± 1.41 | 68.48 ± 0.62 | 14.96 ± 1.89 | 10.00 ± 1.53 | 71.40 ± 0.46 | 22.20 ± 1.50 | 17.24 ± 0.72 | 53.80 ± 0.51 |
| FGD | 10 | 5.01 ± 1.21 | 3.18 ± 0.69 | 59.49 ± 0.46 | 13.18 ± 1.04 | 8.60 ± 1.76 | 65.82 ± 1.14 | 21.82 ± 1.42 | 18.18 ± 1.17 | 60.07 ± 0.54 |
| FGD | 30 | 3.65 ± 0.95 | 2.64 ± 0.60 | 59.01 ± 1.74 | 16.01 ± 1.18 | 10.48 ± 0.98 | 68.28 ± 0.40 | 18.99 ± 0.40 | 15.58 ± 1.40 | 59.99 ± 0.77 |
| FGD | 50 | 5.31 ± 1.35 | 3.01 ± 0.63 | 59.99 ± 1.06 | 17.44 ± 1.88 | 11.99 ± 0.80 | 67.01 ± 0.49 | 14.98 ± 1.08 | 11.74 ± 1.86 | 58.96 ± 0.89 |
Minor 2: Directly assessing within-group coverage
We add per-cell coverage diagnostics, all evaluated in fresh feature networks not used during distillation (100 networks, all seeds at IPC 10):
- fraction of medoids matched;
- maximum multiplicity (points per medoid);
- to the medoids;
- Bures discrepancy to the real cell.
Each cell: FGD / FGD-assign / GroupMoment (mean + full covariance on FGD's allocation, best ); per-cell averages over 100 fresh networks and 3 seeds ( and Bures normalised per cell).
| Dataset, IPC | Coverage ↑ | Max. multiplicity ↓ | to medoids ↓ | Bures to real cell ↓ |
|---|---|---|---|---|
| CelebA 10 | 0.41 / 0.63 / 0.70 | 3.3 / 2.6 / 2.3 | 0.89 / 0.77 / 1.14 | 1.02 / 0.92 / 1.13 |
| bFFHQ 10 | 0.43 / 0.74 / 0.75 | 4.4 / 2.3 / 2.2 | 1.06 / 0.83 / 1.24 | 1.03 / 0.94 / 1.17 |
| UTKFace 10 | 0.61 / 0.89 / 0.85 | 3.0 / 2.0 / 2.0 | 1.03 / 0.90 / 1.17 | 1.05 / 1.03 / 1.22 |
- FGD-assign has the lowest to the medoids in all three settings.
- Its coverage is within 0.07 of mean + covariance matching, and higher on UTKFace IPC 10.
- It also has the lowest Bures discrepancy to the real cell in all three settings, below GroupMoment, which matches covariance explicitly. This is Prop. A in practice: the one-to-one coreset loss controls the covariance as a by-product.
We thank the reviewer for the positive assessment of the presentation and the geometric perspective and for concrete requests that strengthened the paper.
Major 1: Eq. (8) does not ensure all medoids are represented or that covariance is matched
A fair observation about the nearest-medoid step, and we measured it: under 100 fresh feature networks not used in distillation, it covers 61% of the re-solved medoids at IPC 10 and 33% at IPC 50.
FGD's design allows a principled fix. Every (class, group) cell has exactly as many medoids as synthetic slots, so the matching can be made one-to-one with a Hungarian assignment σ at negligible cost. We call this FGD-assign:
This delivers what the reviewer asked for, and it improves the method:
The revision presents both variants.
Major 3 / Q1: Clarifying Eq. (8)
Within each (class y, group a) cell, at each training step:
We will rewrite §4 with these steps and move the training-loop diagram into the main text.
Q2: Performance as protected groups become more imbalanced
We subsampled CelebA's minority group to bias ratios 0.7–0.99 and compared FGD and FGD-assign with FairDD on identical data. Seed-paired difference to FairDD (ΔAcc / ΔDEO$_M$):
This is what FGD is designed for so each minority cell keeps its own slots and real targets however rare the group becomes, whereas a class-shared target is increasingly dominated by the majority.