acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Fair Anomaly Detection via Representation Learning

Abstract

Fair anomaly detection (AD) methods typically trade detection accuracy for group fairness, especially under severe group imbalance. We argue that this trade-off is not inherent to fully unsupervised AD but stems from how fairness constraints distort learned representations. We identify two failure modes: vanishing protected-group influence under imbalance and dimensional collapse caused by overly restrictive fairness regularization. We show theoretically that many existing fairness regularizations dilute the protected group's gradients as imbalance grows and, even with the AD loss included, suppress all group-predictive features as the fairness coefficient increases. We propose FADIG, which combines an adaptively re-balanced autoencoder with a fairness-aware contrastive objective, and prove that it removes the dependence of group contributions on the imbalance ratio and makes collapsed representations suboptimal. On four image and tabular datasets, FADIG achieves higher accuracy with lower group disparity than existing methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Anomaly detection, fairness

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

This paper introduces a method to perform anomaly detection. The author first investigated why the current methods fall short of addressing the issues by introducing the trade-off between performance and fairness. They found that two possible explanations can be vanishing protected influence and dimensional collapse. Motivated by this finding, they propose FADIG, a fully self-supervised, fairness awareness training method. It has two components, (1) FAC encourages representation from different sensitive groups to align (2) Adaptive rebalanced AE reweighs the training dynamics.

Critical strengths
  1. The paper investigates the possible reason causing poor performance in protected group is interesting. The current method using fair regularizer can encounter a trade-off between the loss predictive information and fairness metrics. The paper using singular value spectrum of their missing part is interesting.

  2. The learning objective design is sound, at least it potentially addresses the two issues pointed out by the authors.

  3. The experiment’s performance is strong.

Critical weaknesses
  1. The author explains that the possibility that causing poor performance of protected group by minority groups influences diluted representation dimensional collapse. Per my experience, the alternative way is that minority group is overfit during training.

  2. Proof of Lemma3 has a flaw, where E[∥∇θRp∥]≤a\mathbb{E}[\| \nabla_{\theta} R_p\|] \leq a, E[∥∇θRU∥]≥b\mathbb{E}[\| \nabla_{\theta} R_U\|] \geq b, it can not deduct E[∇θRP∇θRU]≤ab\mathbb{E}[\frac{\nabla_{\theta} R_P}{\nabla_{\theta} R_U}] \leq \frac{a}{b}. Because we can not say E[AB]≤E[A]E[B]\mathbb{E}[\frac{A}{B}] \leq \frac{\mathbb{E}[A]}{\mathbb{E}[B]}.

  3. I would like to see the TSNE on the learned latent and compare with other methods like many fair representation learning work did.

Questions for authors
  1. Eq (3) and Eq (4) are actually reweighing on the two groups, I would like to see why the current method to compute ϵ\epsilon is better than GroupDRO (https://arxiv.org/pdf/1911.08731).

  2. The authors argues that enforces z⊥sz \perp s is not a good way, the existing literature I know share the same idea is FADES (https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=10655684) which I would like to compare which method is more promising. It can think of yy as the anomaly label in the paper.

3 comments
AuthorsOpenReview ↗

Rebuttal by Author (1/2)

We thank the reviewer for recognizing the interesting singular-value spectrum analysis, sound learning objective design, and strong empirical performance of our method.

Weakness 1: The author explains that the possibility that causing poor performance of protected group by minority groups influences diluted representation dimensional collapse. Per my experience, the alternative way is that minority group is overfit during training.

We thank the reviewer for raising this alternative explanation. We agree that overfitting can affect minority-group performance when the available training samples are limited. However, our work focuses on a different mechanism under the capacity-limited reconstruction setting commonly considered in unsupervised anomaly detection. Specifically, in Section 4.2, our analysis assumes that the generative model has limited reconstruction capacity and that normal patterns are more readily fitted than anomalous patterns, which is a common assumption in many papers [1] [2] [3]. Under this setting, the key challenge is not necessarily overfitting the minority group, but rather insufficiently learning its normal patterns due to its limited contribution to the optimization objective. Consequently, normal samples from the protected group may exhibit higher reconstruction errors and be incorrectly identified as anomalies.

[1] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. arXiv preprint arXiv:2306.02602, 2023.

[2] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35: 4571–4584, 2022.

[3] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9737–9746, 2022.

Weakness 2: FairAD performs significantly worse than the results reported in the original FairAD paper.

We want to clarify that FairAD in the two papers refers to different methods. FairAD in this submission is the baseline method in [1] while FairAD in your referred paper is the old name of our method. We changed the name of our proposed method as we found that the name had been used by this baseline method. There is no performance gap issue for FairAD in two papers as they refer to two different methods.

[1] Feng Xiao, Xiaoying Tang, and Jicong Fan. Fairness-aware anomaly detection via fair projection. arXiv preprint arXiv:2505.11132, 2025.

Weakness 3: Proof of Lemma 3.3 has a flaw

We thank the reviewer for carefully examining Lemma 3.3 and identifying these issues. We agree that the original proof does not justify the bound on the expectation of the gradient ratio. However, the intended imbalance-dependent scaling relationship remains valid when formulated as the ratio of expected gradient magnitudes, without requiring stronger assumptions. Specifically, under Assumptions 3.1 and 3.2, we establish the corrected result:

{\mathbb E[\|\nabla_\theta R_U\|]} \leq \frac{k}{c}\sqrt{\frac{n}{m}}.$$ Let $g_i=\nabla_\theta\ell_i(\theta)$. Under Assumption 3.1, the per-sample gradients are independent, zero-mean, and bounded by $\|g_i\|\leq k$. By Cauchy–Schwarz and the vanishing cross terms, $$\begin{aligned} \mathbb E\left[\left\|\nabla_\theta R_P\right\|\right] &=\mathbb E\left[\left\|\sum_{i=1}^{n}g_i\right\|\right]\\ &\leq \sqrt{\mathbb E\left[\left\|\sum_{i=1}^{n}g_i\right\|^2\right]}\\ &=\sqrt{\sum_{i=1}^{n}\mathbb E[\|g_i\|^2]} \leq k\sqrt n. \end{aligned}$$ Combined with Assumption 3.2, $\mathbb E[\|\nabla_\theta R_U\|]\geq c\sqrt m$, which directly establishes the corrected bound. **Importantly, this correction preserves the central theoretical observation.** Under the stated assumptions, the protected group's expected aggregate gradient magnitude becomes relatively smaller as group imbalance increases, following the same $O(\sqrt{n/m})$ scaling. The distinction is that the revised lemma characterizes the ratio of expected gradient magnitudes rather than the expectation of their ratio and this result remains consistent with our motivation for mitigating group-size-driven optimization imbalance. We appreciate the reviewer's feedback and have updated the lemma and proof accordingly. The changes are highlighted in red. **These revisions clarify the statistical interpretation while retaining the imbalance-dependent scaling result motivating FADIG.**
AuthorsOpenReview ↗

Rebuttal by Author (2/2)

Weakness 4: I would like to see the TSNE on the learned latent and compare with other methods like many fair representation learning work did.

We visualize our method and fairness-aware baseline methods in Figure 7 (in Appendix E.2) in the updated PDF file. The visualization shows that our proposed FADIG exhibits anomaly-rich peripheral regions in both groups, although some normal–anomaly overlap remains, suggesting that it preserves group-specific structure alongside information relevant to anomaly discrimination.

Question 1: Eq (3) and Eq (4) are actually reweighing on the two groups, I would like to see why the current method to compute ϵ\epsilon is better than GroupDRO.

Thank you for highlighting this connection. Unlike GroupDRO, which emphasizes groups with larger absolute losses, FADIG accounts for the fact that high reconstruction error in unlabeled data may indicate anomalies rather than under-fitted normal patterns. This distinction matters because our unlabeled training data contain anomalies and high reconstruction error may indicate an anomaly rather than an under-fitted normal pattern, so reducing worst-group reconstruction loss may not necessarily improve detection. Lemma 4.3 (Lemma 4.4 in the updated PDF) motivates our weighting through this fitting trade-off. We have added the discussion of GroupDRO in the related work (Line 484 highlighted in red) in the updated version. For a fair comparison, we adapt GroupDRO to the unsupervised setting by replacing the prediction loss with reconstruction loss and maintain the same network architecture. We compare reconstruction-based GroupDRO with FADIG under identical experimental settings. The detailed experimental setup can be found in Appendix E.1. All values are percentages, averaged across runs.

Dataset (K)MethodAUROC ↑Recall@K ↑Recall Diff ↓AUPRC ↑AUPRC Diff ↓
MNIST-USPS (1200)GroupDRO82.43 ± 0.5554.21 ± 0.8336.32 ± 2.8454.09 ± 0.684.59 ± 2.51
FADIG91.28 ± 0.4667.19 ± 0.333.77 ± 2.1866.47 ± 0.332.52 ± 1.09
MNIST-Invert (500)GroupDRO89.36 ± 0.2754.00 ± 0.3258.66 ± 0.3552.22 ± 1.6016.92 ± 0.46
FADIG97.99 ± 0.0771.82 ± 0.639.78 ± 3.1072.01 ± 0.201.72 ± 0.59
COMPAS (350)GroupDRO56.98 ± 3.0723.72 ± 2.2412.68 ± 8.4823.12 ± 1.089.05 ± 0.23
FADIG61.45 ± 0.4734.38 ± 0.365.97 ± 4.3427.42 ± 1.564.79 ± 3.33
CelebA (5000)GroupDRO50.45 ± 0.757.93 ± 0.913.88 ± 3.689.46 ± 0.345.73 ± 0.38
FADIG59.43 ± 0.4211.96 ± 0.494.72 ± 1.2612.21 ± 0.201.41 ± 0.70

FADIG achieves higher AUROC, recall, and AUPRC on all four datasets, with lower AUPRC disparity throughout and lower recall disparity on three. On CelebA, GroupDRO achieves slightly smaller recall difference at the cost of near-chance AUROC. These results support FADIG’s overall effectiveness.

Question 2: The authors argues that enforces z⊥sz\perp s is not a good way, the existing literature I know share the same idea is FADES which I would like to compare which method is more promising. It can think of yy as the anomaly label in the paper.

Thank you for highlighting FADES. We share its motivation that indiscriminately removing sensitive information can discard useful predictive structure. However, FADES uses target labels and sensitive attributes during training, including in its conditional-mutual-information objective. Our setting provides sensitive attributes but no anomaly labels. Therefore, treating its target variable as the anomaly label would introduce additional supervision, making the original method incompatible with our evaluation protocol. Despite of this, we have discussed its revelance to our paper in the related work (highlighted in red in Section 6 Line 482) in the updated version.

AuthorsOpenReview ↗

Reply to Reviewer

Dear Reviewer VPfn,

Thank you again for your valuable feedback. As the rebuttal period is approaching its end, we would greatly appreciate it if you could take a moment to review our responses. Please let us know if you have any remaining concerns or questions. We would be happy to provide further clarification.

Thank you for your time and consideration!

The Authors

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

The paper studies fairness-aware unsupervised anomaly detection (AD) under group imbalance, where the protected group is much smaller than the unprotected one. It attributes the usual performance-fairness trade-off of existing fair AD methods to two representation-level failure modes: (1) vanishing protected-group influence, where sample-additive fairness regularizers give the protected group a gradient share of order sqrt(n/m) (Lemma 3.3); and (2) dimensional collapse, where fairness regularizers that are minimized by a collapsed encoder suppress group-dependent structure as the fairness coefficient grows (Lemma 3.4, Proposition 3.5). The proposed method, FADIG, combines (i) a fairness-aware contrastive loss L_FAC = L_fair + L_unif, which pulls protected and unprotected representations together and spreads representations within each group, with each term normalised by its own pair count, and (ii) an autoencoder whose reconstruction loss is re-weighted between groups by a data-driven weight epsilon computed from group-wise losses (Lemma 4.3). Experiments on MNIST-USPS, MNIST-Invert, COMPAS and CelebA, plus ACSIncome, VisA and a graph dataset in the appendix, report higher AUCROC / Recall@K / AUPRC than fair and fairness-agnostic baselines with low Recall Diff and AUPRC Diff. Diagnostics of gradient ratios (Fig. 3) and effective rank (Fig. 4) are given to support the two failure modes.

Updated the score based on the rebuttal of the authors.

Critical strengths
  1. Relevant problem. Fair AD under group imbalance matters in practice (fraud, medical, security), and the observation that several fair AD baselines detect at close to chance level (AUCROC around 50%, Tables 1-2) is a useful finding for the community.
  2. Simple method with few hyperparameters. FADIG adds two loss terms to an autoencoder; the contrastive coefficient alpha is fixed to 4 on all datasets, and Fig. 6 suggests results are insensitive to it.
  3. Broad experimental coverage, Results are averaged over four seeds with standard deviations, and code is released.
Critical weaknesses
  1. The paper presents its main contributions as theoretical ("We show theoretically ...", contributions 1-3), but as written these are not supported. There are issues with Lemma 3.3, Theorem 4.2 and Lemma 4.3.
  2. Lemma 3.3: Invalid step from a ratio of expectations. The proof establishes \E∥∇θRP∥≤O(n)\E\|\nabla_\theta R_P\| \le \mathcal{O}(\sqrt{n}) and \E∥∇θRU∥≥cm\E\|\nabla_\theta R_U\| \ge c\sqrt{m}, and then concludes $
\E ⁣[∥∇θRP∥∥∇θRU∥]≤O(n)cm\E\!\left[\frac{\|\nabla_\theta R_P\|}{\|\nabla_\theta R_U\|}\right] \le \frac{\mathcal{O}(\sqrt{n})}{c\sqrt{m}}

$ This does not follow: in general \E[X/Y]≠\E[X]/\E[Y]\E[X/Y] \neq \E[X]/\E[Y], and Jensen's inequality gives \E[1/Y]≥1/\E[Y]\E[1/Y] \ge 1/\E[Y], i.e.\ the opposite direction. Assumption 3.2 controls only \E∥∇θRU∥\E\|\nabla_\theta R_U\|, which places no bound on \E[1/∥∇θRU∥]\E\big[1/\|\nabla_\theta R_U\|\big]; if ∥∇θRU∥\|\nabla_\theta R_U\| is close to zero with non-negligible probability, the expected ratio can be arbitrarily large. The stated result is therefore not proven. Inconsistent independence claim. ∇θRP\nabla_\theta R_P is described as "a sum of nn bounded (possibly dependent) random vectors", and in the next sentence as a sum of independent vectors. The matrix Bernstein inequality is neither stated nor cited.} The proof invokes it once ("By the matrix Bernstein inequality:") and proceeds directly to Eq. (6). The authors should cite the exact source and version of the inequality used (there are several, e.g. Hermitian and rectangular, with different dimension factors), restate it with its assumptions (independence, zero mean, bounded norm) and its variance term, and show how Eq. (6) follows from it. 3. Theorem 4.2: Part (1): undefined quantities. ∇θLfair(P)\nabla_\theta \mathcal{L}^{(P)}_{\mathrm{fair}} and ∇θLfair(U)\nabla_\theta \mathcal{L}^{(U)}_{\mathrm{fair}} ("the gradient contributed by the protected unprotected group") are never defined. The expression given for (P)(P) (line971) is the full gradient of \Lfair\Lfair, and (U)(U) is only asserted "similarly" (line979), with no expression. 4. Lemma 4.3: Assumptions missing from the statement. The five inequalities and the resulting range ΔUAΔUA+ΔPN<ϵ<ΔUNΔUN+ΔPA\frac{\Delta^{UA}}{\Delta^{UA}+\Delta^{PN}} < \epsilon < \frac{\Delta^{UN}}{\Delta^{UN}+\Delta^{PA}} are derived correctly, but they rely on ΔUN>ΔUA\Delta^{UN} > \Delta^{UA} and ΔPN>ΔPA\Delta^{PN} > \Delta^{PA}, which appear only in the proof ("naturally holds") and not in the statement of the lemma. And should explain why they are natural. 5. References: Zhang et al. (2022) and Zhang et al. (2023), "Fairness-aware contrastive learning with partially annotated sensitive attributes", are the same paper listed twice (lines 730-735), both under ICLR. 6. Typos: "celebatab dataset" (line 1324); "our objecive" (line 1294); "Recontrast" vs. "ReContrast" (Tables 11, 16 vs. main text);

Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Rebuttal by Author (1/2)

We thank the reviewer for recognizing the practical relevance of our problem, the simplicity and robustness of FADIG, and the comprehensive experimental evaluation with reproducible results.

Weakness 1: The paper presents its main contributions as theoretical, but as written these are not supported. There are issues with Lemma 3.3, Theorem 4.2 and Lemma 4.3.

We thank the reviewer for carefully examining our theoretical arguments. We agree that some statements require corrections or further details. We have (i) replaced the invalid expectation-of-a-ratio step in Lemma 3.3 with a bound on the ratio of expected gradient magnitudes, (ii) explicitly defined the group-specific contributions in Theorem 4.2, and (iii) added the necessary assumptions (Assumption 4.3) for Lemma 4.3 (Lemma 4.4 in the updated PDF file) and revised the claim about the adaptive weight as appropriate. We have also aligned the abstract, contribution statements, and discussion with the corrected results. The mathematical analysis is complemented by empirical gradient diagnostics (Figure 3), effective-rank analysis (Figure 4), and performance–fairness comparisons (Figure 2).

Weakness 2: Invalid expectation inequality and independence inconsistency in Lemma 3.3

We thank the reviewer for carefully examining Lemma 3.3 and identifying these issues. We agree that the original proof does not justify the bound on the expectation of the gradient ratio. However, the intended imbalance-dependent scaling relationship remains valid when formulated as the ratio of expected gradient magnitudes, without requiring stronger assumptions. Specifically, under Assumptions 3.1 and 3.2, we establish the corrected result:

{\mathbb E[\|\nabla_\theta R_U\|]} \leq \frac{k}{c}\sqrt{\frac{n}{m}}.$$ Let $g_i=\nabla_\theta\ell_i(\theta)$. Under Assumption 3.1, the per-sample gradients are independent, zero-mean, and bounded by $\|g_i\|\leq k$. By Cauchy–Schwarz and the vanishing cross terms, $$\begin{aligned} \mathbb E\left[\left\|\nabla_\theta R_P\right\|\right] &=\mathbb E\left[\left\|\sum_{i=1}^{n}g_i\right\|\right]\\ &\leq \sqrt{\mathbb E\left[\left\|\sum_{i=1}^{n}g_i\right\|^2\right]}\\ &=\sqrt{\sum_{i=1}^{n}\mathbb E[\|g_i\|^2]} \leq k\sqrt n. \end{aligned}$$ Combined with Assumption 3.2, $\mathbb E[\|\nabla_\theta R_U\|]\geq c\sqrt m$, which directly establishes the corrected bound. **Importantly, this correction preserves the central theoretical observation.** Under the stated assumptions, the protected group's expected aggregate gradient magnitude becomes relatively smaller as group imbalance increases, following the same $O(\sqrt{n/m})$ scaling. The distinction is that the revised lemma characterizes the ratio of expected gradient magnitudes rather than the expectation of their ratio and this result remains consistent with our motivation for mitigating group-size-driven optimization imbalance. Regarding independence, we have removed the inconsistent phrase "possibly dependent" in the proof and clarified that the derivation relies on the within-group independence stipulated in Assumption 3.1. We have also replaced the unnecessary matrix Bernstein argument with the direct second-moment proof above. We appreciate the reviewer's feedback and have updated the lemma and proof accordingly. The changes are highlighted in red. **These revisions clarify the statistical interpretation while retaining the imbalance-dependent scaling result motivating FADIG.**
AuthorsOpenReview ↗

Rebuttal by Author (2/2)

Weakness 3: Undefined group-specific gradient contributions in Theorem 4.2

We thank the reviewer for identifying the ambiguity in notation. We provide additional clarification below to address this concern. Our theoretical analysis concerns each group's contribution to the parameter gradient. Let ZP={ziP}Z_P=\{z_i^P\} and ZU={zjU}Z_U=\{z_j^U\}. We clarify the definitions as follows:

\sum_{i=1}^{n} (\frac{\partial z_i^P}{\partial\theta})^\top \nabla_{z_i^P}\mathcal L_{\mathrm{FAC}},$$ $$\nabla_\theta\mathcal L_{\mathrm{FAC}}^{(U)} = \sum_{j=1}^{m} (\frac{\partial z_j^U}{\partial\theta})^\top \nabla_{z_j^U}\mathcal L_{\mathrm{FAC}},$$ such that $$\nabla_\theta\mathcal L_{\mathrm{FAC}} = \nabla_\theta\mathcal L_{\mathrm{FAC}}^{(P)} + \nabla_\theta\mathcal L_{\mathrm{FAC}}^{(U)}.$$ Regarding the use of "similarly," the cross-group alignment term is symmetric with respect to the two groups. The protected- and unprotected-group contributions are obtained by differentiating the similarity function with respect to its first and second arguments, respectively. Under Assumption 4.1 and the symmetry of our similarity function, the same bounding argument applies to both groups. Moreover, since $\operatorname{sim}(z_i,z_j)=\exp(\cos(z_i,z_j))\ge e^{-1}$, the logarithmic normalization is bounded away from zero. Therefore, **the group-size-independent $O(1)$ upper bounds remain valid under the clarified gradient decomposition**. We have revised the proof on Theorem 4.2 to explicitly distinguish the protected-group contribution from the full gradient and clarify the symmetric derivation for the unprotected group with additional steps. **Weakness 4: Missing assumption in Lemma 4.3** > We agree these assumptions must be stated. They express the modeling premise of reconstruction-based anomaly detection that normal patterns are more readily fitted than anomalous patterns within each group, which is **a common assumption** in many papers [1] [2] [3]. Since $\Delta^t=L_0^t-L_1^t$ measures the improvement in reconstruction loss when fitting subgroup $t$, the inequalities mean that normal-example fitting yields larger improvements than anomaly-example fitting. We have added $\Delta^{UN}>\Delta^{UA}$ and $\Delta^{PN}>\Delta^{PA}$ as an assumption (**Assumption 4.3**) in the paper. > >[1] Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. arXiv preprint arXiv:2306.02602, 2023. > >[2] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35: 4571–4584, 2022. > >[3] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9737–9746, 2022. **Weakness 5/6: Duplicate references, Typos and inconsistent names** > We appreciate these corrections. We have removed the duplicate reference and corrected the typos.
AuthorsOpenReview ↗

Reply to Reviewer

Thank you for raising the score! We hope we have addressed all of your concerns. Please let us know if you have any questions. We sincerely appreciate your constructive comments and thoughtful evaluation.

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

The paper argues that in unsupervised anomaly detection, minority protected groups are flagged disproportionately. This flagging is attributed to vanishing protected group gradient influence and, additionally, due to fairness regularization resulting in latent representation that loses group specific information. FADIG is proposed to address these challenges by i) introducing a fairness aware contrastive loss, and ii) weighting minority protected groups in reconstruction scoring. Empirical results show the effectiveness of the proposed method when compared to exciting fairness aware methods.

Critical strengths
  1. The paper clearly states its main goal which is clear right off the bat.
  2. Reweighting for autoencoder is simple and straightforward way to encourage learning from underrepresented groups. More importantly, weighting is analytical rather than tuned.
  3. Empirical results are quite strong, and I think they substantiate the claims well.
Critical weaknesses

No critical weaknesses as such, but I have some questions noted below.

Questions for authors
  1. I’m curious to know why fairness related loss component is needed. Partly because underrepresented protected group (small group), and data imbalance (also about small cohort), aren’t both the same thing? So, why the contrastive loss is important here?
  2. The theoretical results establish the bounds. But in an empirical setting, how useful is bounding the norm? As it may not prevent gradient skew or diminishing signals in stochastic mini batching.
  3. I’m not very clear on cross group pairing (Sec 4.1, Eq 2) as it we are dealing with severe grpup imabalance, there could be a number of batches without any protected group instances. How do you address this in your experiments?
3 comments
AuthorsOpenReview ↗

Rebuttal by Author

We thank the reviewer for recognizing the clear research objective, simplicity and analytical foundation of our reweighting approach, and strong empirical results that substantiate our claims.

Question 1: I’m curious to know why fairness related loss component is needed. Partly because underrepresented protected group (small group), and data imbalance (also about small cohort), aren’t both the same thing? So, why the contrastive loss is important here?

We want to clarify that these two components address different complementary problems. Adaptive reconstruction weighting controls the relative importance of group-wise reconstruction losses and offsets group-size dominance, while it does not preserve the expressive geometry of learned embeddings. Fairness-aware contrastive learning explicitly promotes cross-group alignment while discouraging within-group collapse. The alignment term encourages shared patterns, whereas the within-group term retains diversity that can support anomaly discrimination. Our ablation study (Figure 5, Appendix D.4) directly supports the necessity of contrastive learning. FADIG-R (α=0\alpha=0), which retains only adaptive reconstruction weighting, performs worse than the complete FADIG in both recall and recall disparity. Furthermore, removing either the cross-group alignment term (FADIG-F) or the within-group uniformity term (FADIG-U) degrades performance. These results demonstrate that fairness-aware contrastive learning provides additional benefits beyond reconstruction reweighting.

Question 2: The theoretical results establish the bounds. But in an empirical setting, how useful is bounding the norm? As it may not prevent gradient skew or diminishing signals in stochastic mini batching.

We agree that Theorem 4.2 is intended to establish a group-size normalization property rather than a guarantee of nonzero or equal gradients at every optimization step. In Section 5.4, we provide an empirical test for this situation. Figure 3 tracks protected-to-unprotected gradient ratios across epochs and different group-imbalance levels. FADIG shows more stable ratios in the evaluated settings than FairSVDD and FairOD.

Question 3: I’m not very clear on cross group pairing (Sec 4.1, Eq 2) as it we are dealing with severe grpup imabalance, there could be a number of batches without any protected group instances. How do you address this in your experiments?

To calculate the pairwise terms of Eq. (2), a mini-batch needs at least two protected and two unprotected examples. In our implementation, we use a large batch size to reduce such a probability of no protected samples in the batch and ignore this iteration if it indeed happens. If the batch size is small, a feasible way is to apply group-stratified mini-batching, so that each batch can be constructed with the minimal number of required samples, and the similarity averages can be computed with the actual per-group pair counts. We have added the details in Appendix D.2.

Thank you for taking time to address my questions. Together with ablation, I think it makes sense that the two loss components adress complementary issues. Sec 5.4 and Appendix D.2 resolve most of my concerns. I'll revise my score.

AuthorsOpenReview ↗

Reply to reviewer

Thank you for your positive feedback. We are glad that our clarifications and experimental results have addressed your concerns. We sincerely appreciate your constructive comments and thoughtful evaluation.

Comments on the paper

1 comment

Dear Reviewers,

Please respond to the authors' rebuttal. Thank you.

Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.