Federated Compositional Deep AUC Maximization

Federated learning has attracted increasing attention due to the promise of balancing privacy and large-scale learning; numerous approaches have been proposed. However, most existing approaches focus on problems with balanced data, and prediction performance is far from satisfactory for many real-world applications where the number of samples in different classes is highly imbalanced. To address this challenging problem, we developed a novel federated learning method for imbalanced data by directly optimizing the area under curve (AUC) score. In particular, we formulate the AUC maximization problem as a federated compositional minimax optimization problem, develop a local stochastic compositional gradient descent ascent with momentum algorithm, and provide bounds on the computational and communication complexities of our algorithm. To the best of our knowledge, this is the first work to achieve such favorable theoretical results. Finally, extensive experimental results confirm the efficacy of our method.

Paper

References (42)

Scroll for more · 30 remaining

Similar papers

Peer review

Reviewer 5nau3/10 · confidence 3/52023-07-04

Summary

This work aims to address the challenges of imbalanced data in FL. To this end, the authors propose to optimize AUC score. Some experiments are conducted to verify the effectiveness of the proposed method.

Strengths

The paper is easy to follow. The notations are well-defined. The studied problem is promising.

Weaknesses

This work confuses me a lot. The authors believe that data imbalance is a crucial issue under the FL scenario, which is consistent with my understanding. The problem has motivated the data-split fashion, i.e., Latent Dirichlet Sampling [1]. Moreover, many excellent works have demonstrated the effectiveness of their efforts in mitigating data heterogeneity [2, 3]. However, I cannot find these works in this work. This indicates that the authors may overlook some advanced methods in this field. [1] Measuring the effects of non-identical data distribution for federated visual classification [2] Federated optimization in heterogeneous networks [3] SCAFFOLD: Stochastic controlled averaging for federated learning

Questions

cf. Weaknesses

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

cf. Weaknesses

Reviewer BTwE7/10 · confidence 5/52023-07-15

Summary

This paper firstly studies the federated compositional AUC maximization problem, which includes both the local and global imbalanced distributions, and proposes the momentum-based algorithm LocalSCGDAM to solve this problem. The SOTA convergence rates are established and various experiments are used to evaluate the proposed algorithms.

Strengths

Novelty: This paper is the first work to consider the federated compositional AUC maximization problem. Data heterogeneity is a key problem in federated learning and many works focus on it. The Federated compositional AUC maximization problem is more challenging because it considers both the local and global imbalanced distributions. Quality: It proposes a new algorithm to solve the problem. The structure of the algorithm is clear. Both theoretical analysis and experimental verification are provided. The theoretical analysis is very complete. Clarity: The paper is organized well and it is easy to follow.

Weaknesses

1. More motivation should be introduced. It includes why 1) AUC is significant and 2) why FL AUC should be considered. 2. In the experiments, all samples are set to 0.1. Discussion about different imbalanced data settings is welcomed.

Questions

1. In the math, $\mathcal{D}^k_{g}$ and $\mathcal{D}^k_{n}$ are clear and they are the datasets in the inner layer and out layer. But in the AUC maximization, how to define these datasets? 2. This work solves the binary imbalanced data distribution problem well. I am curious whether it is possible to extend it to the multiple-class imbalanced data distribution problem. 3. In the nonconvex optimization, we consider the convergence rate, including sample complexity and communication complexity, to reach an $\epsilon$-stationary point. In this case, what is the convergence rate of your algorithm?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No

Reviewer jnCq5/10 · confidence 3/52023-07-24

Summary

The paper proposes a new federated learning algorithm to address the class imbalance problem. Instead of using cross-entropy loss functions, the proposed algorithm directly optimizes the AUC score by solving a federated stochastic compositional minimux optimization problem. Specifically, the paper proposes to employ the local stochastic gradient with momentum to update the local model parameters. Experiments show that the proposed algorithm achieves better model performance than the other baselines.

Strengths

1. The paper considers a setting where the global distribution is also class-imbalanced, which is interesting. 2. The paper provides a theoretical analysis of the convergence of the proposed algorithm.

Weaknesses

1. There are many FL studies that try to address non-IID challenge in FL. However, the baselines seem to focus on different optimization methods and lack SOTA FL studies that address the non-IID data (e.g, [1][2]). [1] Addressing class imbalance in federated learning [2] No fear of heterogeneity: Classifier calibration for federated learning with non-iid data 2. The theoretical analysis requires assumptions on the outer-level and inner-level functions. I’m not sure how realistic these assumptions are. 3. The algorithm estimates the inner-level function in local training. It may not be applicable in the client sampling setting, where a client may be selected after many rounds.

Questions

1. Is existing FL approaches applicable with LocalSCGDAM? Existing studies usually optimize the standard cross-entropy loss. I’m curious whether they are applicable to the studied loss function and how they compare with LocalSCGDAM. 2. Is the experimental setting satisfies the assumptions in Section 3.2? 3. Is the algorithm applicable in the client sampling setting? Also, how many clients are used in the experiments? I suggest the authors to add experiments to study the scalability. 4. How does the convergence rate of LocalSCGDAM compare with other related studies? 5. Is directly optimizing AUC loss a mainstream method in class-imbalanced learning?

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

One potential limitation is that the algorithm may not work well in the client sampling setting.

Reviewer 5nau2023-08-15

Further comments

Hi, Thanks for the careful responses! Some of my concerns are addressed. However, it is still unclear why FedProx and Scaffold are not considered baseline methods. I suggest the authors add the corresponding results since existing methods can solve the studied imbalance problem partially. A following up confusion, based on the response, is that the mentioned theoretical results show the convergence rate, why it can confirm the effectiveness of the proposed method?

Authorsrebuttal2023-08-20

Thanks for the reviewer's comments. **Q1**. …I suggest the authors add the corresponding results since existing methods can solve the studied imbalance problem partially. **A1**: (1) Please note that our work focuses on optimizing AUC for federated learning. Therefore, we compared with strong and direct baselines of federated learning for optimizing AUC, including CoDA [1] and LocalSGDAM. As prior works [2, 3] have demonstrated that optimizing traditional cross-entropy (CE) loss will yield worse results than AUC maximization for imbalanced data, we tried not to include too many baselines that optimize CE loss in FL. (2) We would like to draw reviewer's attention that CoDA leverages the epoch-wise proximal term as in FedProx and the variance-reduction as in Scaffold for solving the min-max formulation of AUC maximization. Therefore, CoDA is much a stronger baseline than FedProx and Scaffold. (3) Nevertheless, we have conducted an experiment to compare with Scaffold and FedProx for optimizing CE loss on STL10 dataset with $p=4$, and the results are shown below. It can be observed that our algorithm can outperform those two baselines with a large margin. | | Scaffold | FedProx | CoDA | LocalSCGDAM(Ours) | |-----|----------|---------|-------|-------| | AUC | 0.788 | 0.778 | 0.801 | 0.820 | **Q2**. …why it can confirm the effectiveness of the proposed method?... **A2**. The effectiveness of the proposed method can be explained below. (1) We optimize the compositional AUC formulation, which has been shown to learn much better feature representation than conventional AUC maximization [4]. (2) Our algorithm design and theoretical analysis ensure that our algorithm has a small sample complexity. This is also important as with the same number of epochs, an algorithm with a higher sample complexity may find a worse solution. Therefore, our theoretical analysis ensures that our algorithm can quickly find a good solution to compositional AUC maximization. Together, it can explain the effectiveness of our algorithm. (3) We would like to emphasize that it is not trivial to attain the $O(1/(K\epsilon^4))$ complexity for solving the compositional AUC loss as in our paper. For instance, [5, 6] show that the stochastic gradient descent (SGD) or stochastic compositional gradient descent (SCGD) algorithm under federated learning setting has the sample complexity of $O(1/\epsilon^{8})$ or $O(1/\epsilon^{5})$, which cannot match the sample complexity $O(1/(K\epsilon^{4}))$ of SGD for non-compositional optimization problems. In our theoretical analysis, we show that our algorithm LocalSCGDAM can achieve the sample complexity of $O(1/(K\epsilon^{4}))$, which can match that of traditional federated learning algorithms for non-compositional optimization problems. [1] Guo et al., Communication-efficient distributed stochastic auc maximization with deep neural networks. ICML 2020. [2] Yuan et al., Robust Deep AUC Maximization: A New Surrogate Loss and Empirical Studies on Medical Image Classification. ICCV 2021. [3] Liu et al., Stochastic AUC Maximization with Deep Neural Networks. ICLR 2020. [4] Yuan et al., Compositional Training for End-to-End Deep AUC Maximization. ICLR 2022. [5] Huang et al., "Compositional federated learning: Applications in distributionally robust averaging and meta learning." arXiv preprint arXiv:2106.11264 (2021). [6] Wang et al., "Memory-Based Optimization Methods for Model-Agnostic Meta-Learning and Personalized Federated Learning." JMLR 24 (2023): 1-46.

Reviewer jnCq2023-08-16

Thanks for your response. Most concerns are addressed but my concern about client sampling remains. The estimation of inner-level function may be affected by client sampling. I suggest that the authors add experiments to it.

Authorsrebuttal2023-08-20

Thanks for the reviewer's comments. **Q1**. …I suggest that the authors add experiments to it. **A1**. We would like to emphasize that our experiments focus on the **cross-silo federated learning**, which has been widely used in many applications, such as the biomedical image classification used in our experiment. **For cross-silo federated learning, it is NOT necessary to do client sampling**. In particular, under the cross-silo federated learning setting, there are not a huge number of participants. It is preferable to include all participants to participate in the computation since there does not exist a lot of training data, e.g., biomedical data. Thus, it is not necessary to do client sampling. Nevertheless, we have conducted an experiment about client sampling. In this additional experiment, there are four workers. We use STL10 dataset and the communication period is set to 4. To simulate the client sampling scenario, we randomly select three workers to participate in the computation. The following table shows the test AUC score of our algorithm with/without client sampling. It can be observed that LocalSCGDAM-with-sampling has a similar performance to LocalSCGDAM-without-sampling. | | LocalSCGDAM-with-sampling | LocalSCGDAM-without-sampling | |-----|---------------------------|------------------------------| | AUC | 0.813 | 0.820 |

Reviewer jnCq2023-08-21

Thanks for your further response! I suggest that 1) you highlight you are focusing on cross-silo FL, or 2) conduct experiments in the client-sampling setting in the revision.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC