Classifier Clustering and Feature Alignment for Federated Learning under Distributed Concept Drift

Data heterogeneity is one of the key challenges in federated learning, and many efforts have been devoted to tackling this problem. However, distributed concept drift with data heterogeneity, where clients may additionally experience different concept drifts, is a largely unexplored area. In this work, we focus on real drift, where the conditional distribution $P(Y|X)$ changes. We first study how distributed concept drift affects the model training and find that local classifier plays a critical role in drift adaptation. Moreover, to address data heterogeneity, we study the feature alignment under distributed concept drift, and find two factors that are crucial for feature alignment: the conditional distribution $P(Y|X)$ and the degree of data heterogeneity. Motivated by the above findings, we propose FedCCFA, a federated learning framework with classifier clustering and feature alignment. To enhance collaboration under distributed concept drift, FedCCFA clusters local classifiers at class-level and generates clustered feature anchors according to the clustering results. Assisted by these anchors, FedCCFA adaptively aligns clients' feature spaces based on the entropy of label distribution $P(Y)$, alleviating the inconsistency in feature space. Our results demonstrate that FedCCFA significantly outperforms existing methods under various concept drift settings. Code is available at https://github.com/Chen-Junbao/FedCCFA.

Paper

Similar papers

Peer review

Reviewer uegQ5/10 · confidence 5/52024-07-10

Summary

The paper addresses the significant challenge of data heterogeneity and distributed concept drift in federated learning. The authors propose a novel framework, which integrates classifier clustering and feature alignment to improve model performance and collaboration among clients facing different concept drifts. The key contributions of the paper are as follows: 1. FedCCFA Framework: The proposed framework includes a method for clustering local classifiers at the class level and generating clustered feature anchors to enhance feature alignment. This clustering helps clients with similar data distributions share classifiers, thus improving the generalization performance of the global model. 2. Feature Alignment: The framework introduces an adaptive feature alignment technique that aligns clients' feature spaces based on the entropy of the label distribution. This method helps to alleviate the inconsistencies in feature space due to data heterogeneity. 3. Experimental Validation: Extensive experiments demonstrate that FedCCFA significantly outperforms existing federated learning methods under various concept drift settings. The results show that FedCCFA effectively adapts to distributed concept drift and enhances generalization performance.

Strengths

1. Comprehensive Validation: The research is supported by well-designed experiments and thorough ablation studies, demonstrating the effectiveness of the FedCCFA framework. The mathematical formulations and algorithms are clearly presented, and the results are statistically significant. 2. Clear Presentation: The paper is clearly written and well-structured, making complex concepts accessible. Visual aids, such as graphs and tables, effectively support the explanations. Detailed breakdowns of the experimental setup enhance the reproducibility of the results. 3. Significant Impact: This work addresses a critical gap in federated learning, with potential applications in diverse fields like healthcare, finance, and mobile device collaboration. By improving the adaptability and generalization performance of federated models, the proposed FedCCFA framework provides a robust foundation for future studies to build upon, advancing the current state of federated learning research.

Weaknesses

1. Computation Overhead: The proposed FedCCFA framework involves additional computational steps to train balanced classifiers, which increases the overall computational cost. While the authors attempt to mitigate this by setting small iterations and batch sizes, exploring more efficient methods to achieve balanced classifier training would be beneficial. 2. Limited Evaluation Scenarios: The experiments are primarily conducted on standard datasets. Including more diverse datasets, especially those with real-world distributed concept drift scenarios, would strengthen the validation of the framework's general applicability and robustness. 3. Sensitivity to Hyperparameters: The effectiveness of FedCCFA relies on several hyperparameters, such as the maximum distance in DBSCAN and the scaling factor. While the authors provide some tuning, a more thorough analysis of the sensitivity and robustness to these hyperparameters across different datasets and settings would be valuable. 4. Handling Extreme Data Heterogeneity: The paper addresses data heterogeneity, but extreme cases of data heterogeneity can still pose challenges, as noted with gradient explosions in some scenarios. Further discussion and potential solutions to handle such extreme cases more effectively would improve the robustness of the framework.

Questions

1. Computational Overhead Have you explored alternative methods for balanced classifier training to reduce computational overhead? It’s better to consider leveraging advanced optimization techniques or lightweight pre-processing steps to achieve balanced classifier training more efficiently. 2. Evaluation on Diverse Datasets Do you plan to evaluate FedCCFA on more diverse, real-world datasets reflecting practical distributed concept drift scenarios? It’s better to extend evaluation to real-world datasets would provide a comprehensive understanding of the framework's applicability and robustness. 3. Sensitivity Analysis of Hyperparameters Have you conducted a thorough sensitivity analysis of key hyperparameters across different datasets and settings? A detailed sensitivity analysis could help understand the impact of these hyperparameters on performance and provide guidelines for their selection. 4. Handling Extreme Data Heterogeneity What solutions have you considered for effectively handling extreme cases of data heterogeneity and preventing gradient explosions? It’s better to investigate adaptive methods that dynamically adjust based on real-time training stability monitoring to handle extreme data heterogeneity more robustly

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

1. Adaptability to Various Drift Patterns: The paper evaluates performance under specific concept drift scenarios but may not cover all possible drift patterns. Expanding the evaluation to include a wider variety of drift patterns would provide a more comprehensive understanding of the framework’s adaptability. 2. Real-World Implementation: The paper demonstrates effectiveness in controlled experimental settings. Discussing potential challenges and solutions for deploying FedCCFA in real-world environments, including scalability and communication efficiency, would strengthen the paper. 3. Detailed Analysis of Classifier Clustering: The paper proposes class-level classifier clustering but provides limited analysis on the clustering's dynamics and potential pitfalls. A deeper analysis of how clustering decisions impact overall model performance and stability would be beneficial.

Reviewer EQ7w6/10 · confidence 4/52024-07-11

Summary

This paper explores the impact of distributed concept drift on federated learning (FL), and proposes a novel FL framework, FedCCFA, to adapt to distributed concept drift with data heterogeneity (i.e., label distribution shift). Extensive experiments demonstrated the effectiveness and generality of the proposed method.

Strengths

1. Federated learning (FL) under concept drift holds significant academic and practical value. The method proposed in this paper is both reasonable and effective. 2. The experiments are extensive, not only comparing the proposed method with SOTA methods but also simulating various types of drift. This demonstrates the method's generality across different drift scenarios. 3. The paper is well-organized and well-written, enhancing the clarity and impact of its findings.

Weaknesses

1. In the introduction, the authors use the distribution of medical images as an example to describe data heterogeneity. However, the actual task addressed in the paper is label distribution shift, which does not align with the previous example. It is recommended that the authors directly use label distribution shift to describe the problem addressed in the paper 2. The authors calculate the global feature anchor by averaging (i.e., Eq. (7)). However, given the presence of distributed concept drift among different clients, simply averaging may be unreasonable. It raises the question of whether concept drift adaptation should be considered. 3. The adaptive alignment weight is calculated based on the entropy of the label distribution. However, there lacks the relevant motivation, references, and theoretical justification for this strategy. 4. Federated learning (FL) under distributed concept drift primarily addresses concept drift across clients and over time. This is similar to the research problem of multistream classification under concept drift [1-3]. It is recommended that the authors discuss these works to enhance readers' understanding of related research. [1] An adaptive framework for multistream classification, CIKM 2016 [2] Fusion: An online method for multistream classification, CIKM 2017 [3] Online boosting adaptive learning under concept drift for multistream classification, AAAI 2024 5. The paper does not provide open-source code and datasets, which compromises the reproducibility of the experiments.

Questions

discussed in weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

yes

Reviewer erp16/10 · confidence 3/52024-07-12

Summary

The paper proposes a federated learning framework called FedCCFA, which addresses the challenges posed by distributed concept drift and data heterogeneity. The authors introduce innovative solutions such as classifier clustering and adaptive feature alignment to enhance collaboration and improve model performance under various concept drift settings. The authors demonstrate the effectiveness of their approach through experimental results showing significant improvements over existing methods.

Strengths

1. The paper tackles an important and relatively unexplored problem in federated learning, namely concept drift and data heterogeneity. 2. The extensive experimental results demonstrate the effectiveness of FedCCFA, showing significant performance improvements over existing methods in various concept drift scenarios.

Weaknesses

1. I am unable to fully understand the motivation of the paper. Figure 1 does not clearly illustrate how Decoupled Clustering addresses the real drift problem compared to previous decoupled methods. While it shows better performance than FedAvg and pure Decoupled methods, the underlying reasons remain unclear. Why does introducing clustering improve performance? 2. The writing of the paper can be improved. For example, the specific meaning of \(\phi_{i,c}\) is not explained, leading to significant confusion in understanding equation (4). The authors need to explain each symbol to enhance the readability of the paper. Additionally, the term "balanced batch" in line 160 needs to be explained, specifically how it was sampled.

Questions

1. In the experiments, which classifiers were ultimately clustered in the class-level clustering? What does the final distance matrix look like? 2. Equation (5) uses uniform aggregation to avoid privacy concerns. Why does this method avoid privacy concerns? Could taking the average directly lead to performance loss in classifiers?

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors point out that FedCCFA requires training balanced classifiers, which will increase computational complexity to some extent.

Reviewer HDwb7/10 · confidence 4/52024-07-12

Summary

This paper proposes to consider two cross-coupled important issues, i.e., federated learning and concept drift. To address such a challenging problem, the FedCCFA framework has been proposed composed of classifier clustering and feature alignment modules. The former is designed to cope with concept drift and enhance the generalization performance of the model. The latter is utilized to prevent data heterogeneity of clients. The overall design is sound, interesting, and intuitive. The whole paper is well-organized and well-written. The experimental evaluation is comprehensive and convincing.

Strengths

- A relatively new and important problem of federated learning under concept drift is addressed. - Contributions of this work are in the framework perspective and the module/strategy design perspective. - Both classifier clustering and feature alignment are sound and interesting ideas. - The whole paper is well-organized. - Comprehensive evaluation comparing many counterparts (including SOTAs) on benchmark datasets.

Weaknesses

- Lack of a framework pipeline to intuitively explain the overall design. - Time complexity analysis is missing. - Visual classification results on the datasets are preferable. - The source code is not opened.

Questions

See the weaknesses.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer U5Wz6/10 · confidence 4/52024-07-15

Summary

The paper introduces a solutions based on clustering to group classifiers in the federated learning setting. Such an approach is meant to deal with concept drift in data so as to track and adapt the evolutions of the classifiers of the clients in federated learning. On of the core points is to align features and this is carried out by ether adaptive feature alignment procedure suggested in the paper.

Strengths

- The idea of letting the global/local models evolve over time is interesting and relevant - The clustering is based on the classifier space (the feature extraction phase is fixed) - Results in the concept drift are promising

Weaknesses

Some technical points of the paper are not adequately addressed such as: - the partitioning of feature extractor + classifier is relevant and requires to share a common feature extractor. It is not clear to this reviewer whether the feature extractor is fixed or adapted over time - clustering of models requires to assume that the distribution of the model parameters shares some "locality" (e.g., this holds for linear models but not in general for NNs). - performance in stationary-conditions are less relevant. This is a crucial point (e.g., not suffered by other solutions) that requires to be further investigated

Questions

- What's the need to partition the models into feature extractor and classifier? - Which is the assumption guaranteeing the clustering of classifier parameters in the parameter space? - What's the reason of the reduction in accuracy in no-change condition (i.e., Table 1) and how to mitigate this point?

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

see previous box

Reviewer erp12024-08-12

Thanks for the authors' detailed response and additional experiments. My concerns have been adequately addressed. I will keep my positive score.

Reviewer EQ7w2024-08-12

Thank you for the authors' response. All of my concerns have been addressed, and I will increase the score to 6.

© 2026 NYSGPT2525 LLC