Federated Learning with Bilateral Curation for Partially Class-Disjoint Data

Partially class-disjoint data (PCDD), a common yet under-explored data formation where each client contributes a part of classes (instead of all classes) of samples, severely challenges the performance of federated algorithms. Without full classes, the local objective will contradict the global objective, yielding the angle collapse problem for locally missing classes and the space waste problem for locally existing classes. As far as we know, none of the existing methods can intrinsically mitigate PCDD challenges to achieve holistic improvement in the bilateral views (both global view and local view) of federated learning. To address this dilemma, we are inspired by the strong generalization of simplex Equiangular Tight Frame~(ETF) on the imbalanced data, and propose a novel approach called FedGELA where the classifier is globally fixed as a simplex ETF while locally adapted to the personal distributions. Globally, FedGELA provides fair and equal discrimination for all classes and avoids inaccurate updates of the classifier, while locally it utilizes the space of locally missing classes for locally existing classes. We conduct extensive experiments on a range of datasets to demonstrate that our FedGELA achieves promising performance~(averaged improvement of 3.9% to FedAvg and 1.5% to best baselines) and provide both local and global convergence guarantees. Source code is available at:https://github.com/MediaBrain-SJTU/FedGELA.git.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer J5UX6/10 · confidence 2/52023-07-01

Summary

This paper addresses a challenge in Federated Learning referred to as partially class-disjoint data (PCDD), where each client contributes a part of classes (instead of all classes) of samples. Without full classes, the local objective will contradict the global objective, yielding the angle collapse problem for locally missing classes and the space waste problem for locally existing classes. This is a real-world challenge since it is not uncommon for example that some classes will be well sampled in certain regions, but not others. Prior art mainly focus on the general heterogeneity without specially considering partially class disjoint challenges. Without full classes, the local objective will contradict the global objective, yielding the angle collapse for locally missing classes and the waste of space for locally existing classes. The goal is to achieve holistic improvement in the bilateral views (both global view and local view) of federated learning. The authors propose FedGELA where the classifier is globally fixed as a simplex ETF while locally adapted to the personal distributions. Globally, FedGELA provides fair and equal discrimination for all classes and avoids inaccurate updates of the classifier, while locally it utilizes the space of locally missing classes for locally existing classes. The proposed approach builds upon simplex equiangular tight frame (ETF), which provides each class the same classification angle and generalizes well on imbalanced data. Specifically, in their FedGELA approach, the classifier is globally fixed as a simplex ETF while locally adapted based on the local distribution matrix to utilize the wasted space for the existing classes. In the global view, FedGELA merges class features and their corresponding classifier vectors, which converge to ETF. In the local view. it provides existing major classes with larger feature spaces and encourages to utilize the spaces wasted by locally missing classes. Contributions are summarized as : - Study algorithmic implication of a real-world challenge (partially class-disjoint data (PCDD), namely angle collapse and space waste - Propose FedGELA and theoretically show the local and global convergence analysis for PCDD with the experimental verification - Evaluate on multiple benchmark datasets under the PCDD case and a real-world dataset to demonstrate the bilateral advantages of FedGELA over the state of the art methods.

Strengths

1. Related work well covers comparison among a range of FL methods and why PCDD not covered. Examples of why prior art does not address PCDD include: generic federated leaning adopt a uniform treatment of all classes, then attempt mitigate personal differences; personalized federated learning places less emphasis on locally missing classes and selectively shares parameters/prototypes to minimize the impact of personal characteristics. While these methods might directly or indirectly help mitigate the data shifts caused by PCDD, neither achieve holistic improvement for global and local views 2. Performance evaluaton on 3 relevant datasets (SVHN, CIFAR10, CIFAR100), against all the top state of the art algorithms as baselines, and showng it outperforms. FedGELA consistently exceeds all baselines.

Weaknesses

Overall, the paper represents is a solid contribution - well defined problem not addressed by prior art and representative of real-world problem for fedrated learing. Solid treatment of prior art, and differentiation from prior methods. Weaknesses: 1. Only 1 real PCDD federated application Fed-ISIC2019 was evaluated - however I am not aware of other benchmarks I would recommend. 2. Performance improvements against to the best baseline for all tests were all <3% performance improvement.

Questions

1. Given the performance improvements were in the <3% against the best baselines - can you provide more information? For example does it perform better worse under certain conditions? if so, pls distinguish, and if possible explain what this may imply about either the algorithm and limitations, or perhaps a limitation in the benchmark wrt fully characterizing the real world challenge?

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

No negative societal effects. Limitations explained in Weaknesses section.

Reviewer SUwJ6/10 · confidence 3/52023-07-04

Summary

This paper mainly focuses on the partially class-disjoint data (PCDD) problem in federated learning (FL) settings, which is a common yet challenging problem in distributed data sources. Inspired by a classifier structure (simplex equiangular tight frame, ETF), the authors of the paper propose FedGELA to tackle the PCDD problem. FedGELA is a variant of FedAvg with local model adaptation (personalization): They first define the classifier $W$, which is the ETF that the classifier should converge to. Here, they also take the client local data distribution ($\phi$) into account. Afterwards, the feature extractor $H$ will be optimized locally at each client and communicated via server-client communication. Finally, the global feature extractor and the $W$ at central server, as well as the local feature extractors and the adapted $W$ at clients will be returned.

Strengths

The proposed method is motivated very well. The schematic illustration is also clear. The theoretical analysis is sound. Experiments and the results are good.

Weaknesses

From my understanding, FedGELA focuses only on the alignment in the feature embedding space, which has been done by many previous works (FedGen [1], FedProto [2], …). Therefore this paper lacks significance to some extent. Also, the definition of $W$ (ETF, the global classifier), as well as the locally adapted ones looks straightforward.

Questions

1. You claim that FedGELA could mitigate PCDD from both Global and Local view. But in Algorithm 1, the global model $H^T$ is simply an average of the client model and there is no server-side optimization. Could you please explain this? 2. In Table 1, you claim that FedGELA could mitigate the model skew, could you please explain this point in more details? Also, in terms of “save space”, since you are transmitting the whole feature backbone $H$, which is the resnet18 in your experiments, what is the reason of space saving? 3. Why do you use $H$ to represents 2 different terms, features (Line 114) and global backbone (in the Algorithm)? It’s a bit confusing during reading. 4. Could you shortly explain the converged angles in your figures? Is there a specific meaning of the values? E.g. in Figure 2. 5. In Equation 4, you model the client data skew in the label space via $\phi$, which is based on the number of samples from different classes. Have you experimented with other options? This looks a bit too straightforward to me. 6. FedGen[1] is a method which augment the feature embedding space using a shared feature generator, which could possibly mitigate the issue of “waste of space” by generating synthetic embeddings from minority classes. Could you provide a comparison with this work? [1] Zhu, Zhuangdi, Junyuan Hong, and Jiayu Zhou. "Data-free knowledge distillation for heterogeneous federated learning." ICML, 2021. [2] Tan, Yue, et al. "Fedproto: Federated prototype learning across heterogeneous clients." AAAI 2022. Minors: 1. Line 115, $E_W$ and $E_H$ should be introduced at their first appearance. 2. Line 298, the selection of $E_W$ seems to be dataset specific, is there any default values suggestions?

Rating

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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

FedGELA is only tested in the client data with label skew. Is it also applicable to the data with feature skew?

Reviewer SUwJ2023-08-18

concerns are addressed

The rebuttal from the authors addresses the concerns in my review, and the additional experiments also indicate the effectiveness of their proposed method.

Authorsrebuttal2023-08-18

Thank you very much for your confirmation. If the concerns have been addressed, would you like to raise the rating score? We will carefully follow your comments and include all the experiments and discussions in the revision. Best, The authors of Submission7489

Reviewer UNoN5/10 · confidence 3/52023-07-06

Summary

This paper introduces a novel Federated Learning Algorithm to address the Partially class-disjoint data (PCDD) problem. The approach is based on the simplex equiangular tight frame (ETF) phenomenon to solve the angle collapse issue and introduces a second projection to personalize an adapted structure to save space. The main contributions can be summarized in three aspects: identifying the angle collapse and space waste challenges in the PCDD problem, introducing the novel FedGELA algorithm, and conducting a range of experiments to evaluate its performance. The paper also includes a theoretical analysis with convergence analysis.

Strengths

1. The paper is well-written with a proper structure and clear explanations. The presentation of the authors' ideas is easy to follow due to the effective use of figures and notations. 2. The methodology of the FedGELA algorithm is interesting, and the mathematical deductions are sufficient. The algorithm is clear and provides enough information for reproducibility. 3. The algorithm has been wisely experimented, and the plots are suitable and clear.

Weaknesses

1. The authors claim that "none of the existing methods can intrinsically mitigate PCDD challenges to achieve holistic improvement in the bilateral views of federated learning." However, the PCDD problem seems closely related to the general non-iid (non-independent and identically distributed) problem. The main differences between these two problems have not been explained. 2. Based on my understanding, if PCDD is different from the non-iid problem, it should perhaps be related to the multi-label problem. However, the presentation of the paper, the experimental data, and the methods of experimental comparison all tend to be more inclined towards non-iid problems. Non-iid is a common problem setting, which contradicts the first author's claim of contributions. 3. The performance improvement is limited. 4. I disagree with the statement that "restricting local structure will waste feature space and limit the training of the local model on existing classes." I believe the notion of "waste of space" is unfounded as it appears to have no impact on computational efficiency or performance improvement. Conclusion: The methodology and algorithm presented in this paper are interesting, and the paper is written in high quality. However, there seems to be an important flaw in the problem setting. PCDD appears not to be a new issue but rather a non-iid problem under some special conditions.

Questions

see weaknesses

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

see weaknesses

Reviewer nvCV6/10 · confidence 2/52023-07-23

Summary

The authors study the problem of federated learning over partially class-disjoint data and propose using equiangular tight frame (ETF) techniques that allows achieving better performance in both the global and personal learning tasks. They show that the existing federated learning approaches suffer either from angle collapse for locally missing classes or from waste of space for locally existing classes, and propose their approach FedGELA which solves both the issues.

Strengths

+ Extensive experiments comparing the proposed approach, FedGELA, with the existing federated learning approaches. + Detailed theoretical analysis of the proposed approach. + Highlighting the issues of angle collapse and waste-of-space in the federated learning with partially class-disjoint data.

Weaknesses

- Borrows the existing EFT techniques and hence the novelty seems to be limited. - Overall improvement in average accuracy is marginal (~1.5%) over the existing approaches.

Questions

I like the experimental evaluations and thorough comparison with the prior federated learning works, but I'm concerned about the algorithmic novelty of the proposed approach. Most of the techniques seem to be adapted from the prior known literature on EFT. Can the authors highlight the technical difficulties in directly applying the prior techniques in solving the PCDD problem?

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

2 fair

Limitations

I don't think there are any negative societal impacts of this work.

Authorsrebuttal2023-08-15

Invitation to rolling discussion for the possible remaining concerns

Dear Reviewer, We have thoroughly considered your comments and provided detailed response to address your concerns about technical novelty, clarification, more comparison and more verification. We would like to ask you whether you have remaining or more concerns, so that we can try our best to **timely answer you** during this reviewer-author discussion phase, instead of giving some incomplete demonstrations when approaching to the deadline of this phase. Best, The Author of Submission7489

Authorsrebuttal2023-08-18

Dear Reviewer SUwJ: We appreciate your questions and suggestions, which helps us improve the submission. As your rating score is negative, we would like to know that whether our detailed responses have addressed your concerns. If not, we would like to have a further discussion and explanation with your remaining questions. As the deadline is approaching and we have not received your response, we really appreciate that the reviewer can feedback your points on this submission and promote the discussion. Thank you very much. Best, The authors of submission7489

Reviewer J5UX2023-08-15

I have read the authors rebuttal and rebuttals to some of the other reviewers. The authors' rebuttal addresses well the points raised in my review, and are recommended for inclusion in the final submission, if accepted. I do not have further questions for the authors.

Authorsrebuttal2023-08-15

We sincerely appreciate the positive support of the reviewer. We will carefully follow your constructive comments and include the corresponding contents in the revision to improve the submission. Best, The Author of Submission7489

Reviewer nvCV2023-08-16

Thank you for providing clarifications. I have no further questions for the authors.

Authorsrebuttal2023-08-16

Thank you very much for your confirmation. We will carefully include the contents regarding your suggestion in the revision. Best, The Author of Submission7489

Authorsrebuttal2023-08-18

Dear Reviewer UNoN: As you do have a few arguments about some points in our submission, we would like to kindly ask whether our explanations address your concerns. If not, we would like to have a further discussion with you. We appreciate the reviewer's challenges on some points, which make us improve the submission, and welcome your further discussion. Thank you very much. Best, The author of submission7489

Reviewer SUwJ2023-08-18

The rebuttal from the authors addresses the concerns in my review, and the additional experiments also indicate the effectiveness of their proposed method.

Authorsrebuttal2023-08-18

Thank you for your positive support. We will carefully follow your comments to improve the submission in the revision. Best, The authors of Submission7489

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC