DRAUC: An Instance-wise Distributionally Robust AUC Optimization Framework

The Area Under the ROC Curve (AUC) is a widely employed metric in long-tailed classification scenarios. Nevertheless, most existing methods primarily assume that training and testing examples are drawn i.i.d. from the same distribution, which is often unachievable in practice. Distributionally Robust Optimization (DRO) enhances model performance by optimizing it for the local worst-case scenario, but directly integrating AUC optimization with DRO results in an intractable optimization problem. To tackle this challenge, methodically we propose an instance-wise surrogate loss of Distributionally Robust AUC (DRAUC) and build our optimization framework on top of it. Moreover, we highlight that conventional DRAUC may induce label bias, hence introducing distribution-aware DRAUC as a more suitable metric for robust AUC learning. Theoretically, we affirm that the generalization gap between the training loss and testing error diminishes if the training set is sufficiently large. Empirically, experiments on corrupted benchmark datasets demonstrate the effectiveness of our proposed method. Code is available at: https://github.com/EldercatSAM/DRAUC.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer uBXL6/10 · confidence 1/52023-07-07

Summary

This paper considers AUC optimization in the distribution shift setting, which is of great value to the machine learning community. The author formulates the problem as a min-max-min-max problem and relaxes it as an upper bound minimization problem. The experimental results also confirm the effectiveness of this proposal.

Strengths

The theoretical results seem reasonable. However, as I am not an expert on distributionally robust optimization, I cannot justify the theoretical significance of this paper.

Weaknesses

Table 1 shows the results of CIFAR10-LT, but there is no description of the experimental settings in the context. The proposed method performs poorly on CIFAR10-LT, which seems to contradict the claim of the article. How can this phenomenon be explained? Miscellaneous: The colorful curve in Figure 1 may not be friendly for black-and-white print versions. Providing a legend directly may be a better choice. The definition of $b^*$ in Eq (5) seems to be incorrect.

Questions

Is there more discussion on the results of CIFAR10-LT, why is the performance of the proposal similar to AUCMLoss, and sometimes even weaker than AUCMLoss? --- Thanks for the detailed clarifications, which have addressed my concerns. I would like to keep my score.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors do not provide a discussion about the limitations and potential negative societal impact.

Reviewer u8fo5/10 · confidence 1/52023-07-08

Summary

Area under the ROC curve (AUC) has been widely used for imbalanced classification, but optimizing the metric requires the assumption that training and test data come from the same distribution, which is not realistic. The authors tackle to explore distributionally robust optimization under the context of AUC optimization. They first integrate DRO and AUC and reformulate it to a tractable optimization problem using surrogate loss functions. Finally, they proposed distribution-free DRAUC and distribution-aware DRAUC optimization frameworks. The proposed frameworks are validated theoretically and empirically via experiments on multiple datasets.

Strengths

- The authors propose DRAUC-df by combining DRO and AUC optimization in a tractable way and further extend it to DRAUC-da by considering perturbations on the positive and negative distributions separately. - Two proposed frameworks are validated on several standard datasets, which achieved meaningful performance improvement comparing to the baseline (AUCML).

Weaknesses

- Experiments are only conducted on small-scale datasets, like as CIFAR or MNIST with ResNet and CNN architectures. It may need to validate on more realistic large-scale datasets, like as AUCML on large-scale medical image datasets with more recent network architectures (EfficientNet, DenseNet). - This paper addresses the robustness under worst local distribution from DRO perspective, but it restricts to binary classification problem (potentially due to the use of AUC metric) and validated only on binary version of datasets, so I have reservations about the suitability of this problem setting.

Questions

- In L97, why the Wasserstein constraint in the optimization algorithms can be removed with Theorem 2? - Typos in the definition of \ell_{0,1} and Eq. (2)

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations and potential societal impacts are not addressed.

Reviewer 9FMm5/10 · confidence 3/52023-07-09

Summary

The paper proposes an end-to-end framework to optimize the distributional robust AUC. The authors address several key challenges of this task and robust generalization can be achieved with theoretical guarantee. Experiments on several corrupted long-tailed benchmarks demonstrate the superiority of the proposed method.

Strengths

1. The paper addresses several key challenges to integrating AUC optimization with DRO, including computational infeasibility, intractable solution, and label bias. The proposed method is well-motivated and technically sound. And the generalization risk is bounded with theoretical guarantee. 2. Experiments on several corrupted long-tailed benchmarks demonstrate the superiority of the proposed method. 3. The paper is well-written and clearly organized. Sufficient preliminaries on AUC and DRO are provided.

Weaknesses

1. The proposed method is only evaluated on small-scale datasets such as MNIST and CIFAR-10/100. It remains unclear whether the proposed method can be scaled to large datasets. 2. It is well known that DRO suffers from the problem of yielding overly-pessimistic models with low confidence [1]. This weakness might be inherited in the proposed method as Wasserstein DRO is adopted. [1] Hu, Weihua, et al. "Does distributionally robust supervised learning give robust classifiers?." International Conference on Machine Learning. PMLR, 2018.

Questions

Please check the weakness.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitation of the proposed method may lie in scalability and the issue of overly-pessimistic models.

Reviewer 2tJ94/10 · confidence 3/52023-07-25

Summary

The authors propose a Distributionally Robust AUC (DRAUC) optimization method that combines Wasserstein distance based DRO with AUC optimization. The combination is designed to be robust for data perturbations. The authors provide theoretical analysis from perturbation and generalization perspective, and justify the practical performance on the perturbed data version for MNIST and CIFAR10.

Strengths

1) The motivation is clear: combining DRO (perturbation robust) with AUC. 2) Authors provide both theoretical and practical justifications.

Weaknesses

1) The idea is straightforward and the contribution is limited. Perturbation robustness idea has been explored for adversarial defense (e.g. Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472–7482. PMLR, 2019; Yin, Dong, Ramchandran Kannan, and Peter Bartlett. "Rademacher complexity for adversarially robust generalization." International conference on machine learning. PMLR, 2019). From my view point, this paper directly extend the perturbation robustness to AUC learning. 2) There are some issues for presentation: i) the $\phi_\lambda(z,\mathbf\theta)$ in line 94 doesn't make sense because the $z'$ will be simply optimized as the value $z$. The problem exists through the whole main content. ii) The '=' should be removed for E.q. 15 to be consistent with other presentations. iii) The $\mathcal W$ below line 144 should be $\mathcal W_c$. iv) The notations for $\hat Q_+$ and $\hat Q_-$ are abused for E.q. 18 and 19, which are confusing. v) Algorithm 1 and Algorithm 2 are similar, which can be further simplified or merged. 3) It might be better to include (D. Zhu, G. Li, B. Wang, X. Wu, and T. Yang. When auc meets dro: Optimizing partial auc for deep learning with non-convex convergence guarantee. In International Conference on Machine Learning, pages 27548–27573. PMLR, 2022) as another baseline for experiments because it also combines DRO with AUC, although it is not designed for perturbation robustness. 4) The Figure 3 is not very informative. I can only observe CE loss is worse than AUC-type losses.

Questions

Please see the weaknesses.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Please see the weaknesses.

Reviewer uBXL2023-08-16

Thank you for the clarifications and additional supporting results.

Area Chair D8vA2023-08-18

It is not clear why pairwise formulation requires (N^2) complexity. Since the algorithm can be always stochastic based on mini-batch samples, it do not need all pairs in the data. For pairs in the mini-batch, it is fairly light compared with model size. The method base on pairwise loss formulation is a natural baseline.

Authorsrebuttal2023-08-20

>**Q:** It is not clear why pairwise formulation requires (N^2) complexity. Since the algorithm can be always stochastic based on mini-batch samples, it do not need all pairs in the data. For pairs in the mini-batch, it is fairly light compared with model size. The method base on pairwise loss formulation is a natural baseline. **A:** We appreciate your insightful feedback. The primary challenge associated with the pairwise formulation lies in generating the local worst distribution. The pairwise formulation of DRAUC is defined as: ${DRAUC}\_\epsilon(f\_\theta, \widehat P) = 1 -\max\_{\widehat Q:\mathcal W\_c(\widehat Q,\widehat P) \le \epsilon} \mathbb E\_{\widehat Q} [{\ell(f\_\theta(x^+)-f\_\theta(x^-))}]$ Generating the local-worst distribution $\widehat Q $ suffers from two main difficulties. Firstly, as highlighted in our response to reviewer 2tJ9, the strong duality becomes very complicated when dealing with pairwise loss. Consequently, the Wasserstein perturbation becomes nearly infeasible. Secondly, generating $\widehat Q $ is loss-dependent, implying that we need to know all the training details to deliver a malicious attack. In our endeavor to secure a performance guarantee for our model, we cannot limit the scope of information accessible to an attacker. This results in a computational cost remaining at $O(N^+N^-dK)$, even within a stochastic setting. We will provide further explanations regarding the unsuitability of the pairwise formulation in Sec 3.1. While the pairwise formulation might not be ideal for a distributionally robust setting, it can be considered as a baseline of natural training. The outcomes are presented as below: **The results of pairwise AUC(PAUC):** |Datasets|Models|0.01-C|0.05-C|0.10-C|0.20-C|0.01|0.05|0.10|0.20| |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| |CIFAR10|resnet20|56.31|75.75|82.69|84.07|57.43|84.35|90.94|94.30| |CIFAR10|resnet32|61.81|75.24|80.44|84.06|65.01|81.46|89.90|94.21| |CIFAR100|resnet20|52.15|61.19|64.98|69.64|52.58|63.24|67.52|73.40| |CIFAR100|resnet32|53.52|61.72|65.21|69.17|54.24|63.65|67.73|73.30| |MNIST|small_cnn|94.58|97.24|98.39|98.90|99.14|99.86|99.92|99.96| |MNIST|resnet20|90.28|92.68|90.35|98.65|98.65|99.87|99.89|99.95|

Area Chair D8vA2023-08-18

Thank you for adding more results though Melanoma is still a small data. One related question here is why Wasserstein DRO? Other techniques such as adversarial training, f-divergence could also improve the generalization. Do authors make comparison to these methods?

Authorsrebuttal2023-08-20

>**Q:** Thank you for adding more results though Melanoma is still a small data. One related question here is why Wasserstein DRO? Other techniques such as adversarial training, f-divergence could also improve the generalization. Do authors make comparison to these methods? **A:** Thank you for your feedback! Our apologies for not providing a clear introduction in our manuscript. We indeed draw comparisons with methods that employ f-divergence. The competitors we refer to in our paper, specifically **ADVShift** ([a]) and **DROLT** ([b]), utilize KL-divergence, which is a particular instance of f-divergence, as their metric. Based on our experiments, our method outperforms these methods. Regarding adversarial training, we introduce PGD as an additional competitor. Following the setting in [c], we generate adversarial examples within an $l_2$-norm ball with a radius of $128/255$ and a step size of $15/255$. We only tune the learning rate on the validation set, and the testing results are shown in the subsequent table. **The results of PGD-10:** |Datasets|Models|0.01-C|0.05-C|0.10-C|0.20-C|0.01|0.05|0.10|0.20| |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| |CIFAR10|resnet20|60.86|72.02|79.59|84.82|62.31|75.58|83.54|88.86| |CIFAR100|resnet20|52.70|58.72|59.50|60.90|53.00|59.84|60.61|62.14| |CIFAR10|resnet32|60.01|69.13|77.84|83.62|61.14|72.05|81.54|87.31| |CIFAR100|resnet32|53.05|55.25|58.69|57.30|53.33|55.82|59.67|58.20| |MNIST|resnet20|96.71|98.17|98.07|98.78|99.59|99.97|99.98|99.99| |MNIST|small_cnn|94.42|98.21|98.83|98.90|99.53|99.93|99.97|99.98| From the results, it is evident that PGD narrows the generalization gap. However, in a long-tailed setting, PGD exhibits suboptimal performance in clean scenarios, which is also potentially attributable to the robust overfitting phenomenon detailed in [c]. Thus, its robust outcomes are not as high as those of our methods. [a]Zhang, Jingzhao, et al. "Coping with label shift via distributionally robust optimisation." arXiv preprint arXiv:2010.12230 (2020). [b]Samuel, Dvir, and Gal Chechik. "Distributional robustness loss for long-tail learning." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. [c]Rice, Leslie, Eric Wong, and Zico Kolter. "Overfitting in adversarially robust deep learning." International Conference on Machine Learning. PMLR, 2020.

Area Chair D8vA2023-08-18

Thank you for the rebuttal!

Reviewer 2tJ92023-08-18

Thanks for your rebuttal

The reviewer has read the rebuttal and appreciate the efforts made by the authors. However, the reviewer still holds the original evaluation score. Although my previous concerns about the writing issues can be fixed, I don't feel the paper is well prepared by the deadline of NeurIPS. Moreover, the experimental results don't include variance or standard deviation and might not be repeated independently multiple times.

Authorsrebuttal2023-08-20

>**Q:** The reviewer has read the rebuttal and appreciate the efforts made by the authors. However, the reviewer still holds the original evaluation score. Although my previous concerns about the writing issues can be fixed, I don't feel the paper is well prepared by the deadline of NeurIPS. Moreover, the experimental results don't include variance or standard deviation and might not be repeated independently multiple times. **A:** Thank you for your comment! We are glad to hear that our rebuttal has well addressed your previous concerns. Thanks to the valuable suggestions the reviewers proposed, our manuscript has become more complete and convincing; we really appreciate that! Regarding the matter of the error bar, we would like to clarify that due to the considerable computational expense, it is widely accepted within the realms of Wasserstein DRO and adversarial training to refrain from repeated experiments, as evidenced by references [a, b, c, d]. Nevertheless, in order to allay your concerns, we have repeated our method on CIFAR10 three times using varied seeds and present the average results and their standard deviation: **The results were repeated 3 times with different seeds on CIFAR10 using ResNet20:** |Methods|0.01-C|0.05-C|0.10-C|0.20-C|0.01|0.05|0.10|0.20| |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| |DRAUC|65.18(1.07)|79.87(0.64)|85.78(0.48)|89.35(0.13)|68.05(0.69)|86.19(0.16)|90.78(0.37)|94.38(0.20)| |CDRAUC|65.43(0.18)|79.89(0.55)|85.74(0.41)|89.62(0.31)|67.79(0.50)|84.37(0.65)|90.69(0.34)|94.3(0.19)| The results indicate that the variances of our method are small. [a]Sinha, Aman, et al. "Certifying some distributional robustness with principled adversarial training." arXiv preprint arXiv:1710.10571 (2017). [b]Bui, Tuan Anh, et al. "A unified wasserstein distributional robustness framework for adversarial training." arXiv preprint arXiv:2202.13437 (2022). [c]Madry, Aleksander, et al. "Towards deep learning models resistant to adversarial attacks." arXiv preprint arXiv:1706.06083 (2017). [d]Jiang, Ziyu, et al. "Robust pre-training by adversarial contrastive learning." Advances in neural information processing systems 33 (2020): 16199-16210.

Reviewer 9FMm2023-08-18

Thank the authors for the rebuttal. R1 addresses my concern about the scalability of the proposed method since the results on Tiny-ImageNet-200 are provided. However, I can't entirely agree with R2. Wasserstein DRO also suffers from overly pessimistic issues as indicated by [1]. And f-divergence DRO is an important baseline for evaluation, but the author did not seem to provide the results of f-divergence DRO for comparison. [1] Frogner, Charlie, et al. "Incorporating unlabeled data into distributionally robust learning." arXiv preprint arXiv:1912.07729 (2019).

Authorsrebuttal2023-08-20

>**Q:** Thank the authors for the rebuttal. R1 addresses my concern about the scalability of the proposed method since the results on Tiny-ImageNet-200 are provided. However, I can't entirely agree with R2. Wasserstein DRO also suffers from overly pessimistic issues as indicated by [1]. And f-divergence DRO is an important baseline for evaluation, but the author did not seem to provide the results of f-divergence DRO for comparison. **A:** Thank you for your constructive comment! We are glad to hear that your concern regarding scalability has been sufficiently addressed. With respect to f-divergence, we regret the oversight in our manuscript. Notably, our competitors, **DROLT**([a]) and **ADVShift**([b]), employ DRO based on KL-divergence to enhance model robustness, and our methods consistently outperform theirs across all experiments. To our knowledge, [c] reveals a phenomenon quite different to that described in [d]. The former suggests that in Wasserstein DRO, the training process may collapse, leading the model to a trivial solution, even when the perturbation radius $\epsilon$ is relatively small compared to the distance between training distribution and true distribution. To check whether this anomaly manifests in our setting, we conduct a simple two-step experiment on CIFAR datasets. In the first step, we estimate an upper-bound of Wasserstein distance between training and testing distributions. In the second step, we set $\epsilon$ to be larger than the estimated distance, and see if similar issue occurs in our context. **Step 1:** Assume that both training and test sets of the original CIFAR are sampled from the same distribution, and our corrupted CIFAR datasets are sampled from the "real distribution". Directly solving the Wasserstein DRO between these distributions remains intractable. However, given that the corrupted datasets are generated from the original testing sets, a good estimate of the Wasserstein distance can be obtained by calculating the mean example-wise distance between corrupted and clean datasets, resulting in: **The estimated Wasserstein distance between CIFAR and different levels of CIFAR-C:** |Datasets|Level1|Level2|Level3|Level4|Level5| |:-:|:-:|:-:|:-:|:-:|:-:| |CIFAR10|2.80|4.07|4.67|5.58|6.63| |CIFAR100|2.83|4.11|4.67|5.57|6.61| **Step 2:** Train our model with $\epsilon \in \\{1,2,3,4,5,6,7,8\\}$, which is much larger compared to our experimental parameters. In the 16 groups of experiments we run on CIFAR10 and CIFAR100, only $\epsilon=7$ in CIFAR100 fails to converge. However, by choosing smaller learning rate, our model can evade collapsing into a trivial solution. This experiment proves that the training collapse problem is not that fatal in our setting, even when $\epsilon$ is substantially larger than the distance to the true data distribution, and can be avoided by simply tuning the learning rate. However, shifting the discussion about overly-pessimistic to a broader perspective, beyond a specific setting, we must probe if our methods encounter performance degradation under large, unconstrained uncertainty set. This becomes similar to what we discussed in Prop. 2, i.e., if we place no constraint on the distributional attacker, it will tend to attack the tail class, potentially causing label noise. This might typify the over-pessimistic problem inherent in our setting. Nonetheless, we have proposed our solution by adding structural constraint, namely DRAUC-Da, in our paper. [a]Samuel, Dvir, and Gal Chechik. "Distributional robustness loss for long-tail learning." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. [b]Zhang, Jingzhao, et al. "Coping with label shift via distributionally robust optimisation." arXiv preprint arXiv:2010.12230 (2020). [c]Frogner, Charlie, et al. "Incorporating unlabeled data into distributionally robust learning." arXiv preprint arXiv:1912.07729 (2019). [d]Hu, Weihua, et al. "Does distributionally robust supervised learning give robust classifiers?." International Conference on Machine Learning. PMLR, 2018.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC