acceptodds
Under review as a conference paper at ICLR 2027

Robust Federated Multi-View Clustering

Abstract

Federated Multi-View Clustering (FMVC) has attracted widespread attention in the field of privacy-preserving data analysis. However, existing methods usually suffer from noise interference in practical distributed scenarios: the raw data collected by different clients generally contain noise, outliers, and local distortions, which lead to unstable local representations and clustering structures, resulting in degraded clustering performance. To address the problem, we propose a robust federated multi-view clustering method (RFMVC). For each client, instead of learning clustering structure directly on the original noisy data, our method learns a denoised feature space and learns clustering structure in it to preserve the key structural information of the original data while effectively suppressing the influence of noise and local perturbations on clustering results. For the server, it employs the tensor Schatten p-norm to impose a unified constraint on the cluster indicator matrices of all clients, and combines it with an adaptive client weighting mechanism to fuse multi-client information, thereby enhancing cross-client consistency while achieving denoising at the global structural level. Experimental results show that the proposed method outperforms methods on multiple real-world multi-view datasets, which verifies its robustness.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated Learning, Multi-view Clustering, Robust Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NDAiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RFMVC, a federated multi-view clustering method in which (i) each client learns a "denoised" representation "Z" ^"(m)" regularized by a fidelity term "‖" "X" ^"(m)" "-" "Z" ^"(m)" "‖" _"F" ^"2" and performs k-means-like clustering on the induced distance matrix, and (ii) the server couples the stacked indicator matrices with a tensor Schatten p-norm plus an auto-weighted client weighting "α" _"m" ^"r" . Experiments on six benchmarks report ACC/NMI/PUR, an ablation, parameter sensitivity and convergence curves.

Critical strengths

1.The problem setting (robustness of FMVC against client-side noise) is relevant and under-explored.

  1. The optimization is fully worked out, and the use of cluster statistics ("N" _"k" "," "s" _"k" "," "q" _"k" ) to avoid materializing the N×NN\times N distance matrix is a sensible practical choice (O(NCd_m) per iteration).

  2. Only a scalar T^((m)) and the indicator Y^((m)) are uploaded, which is a lightweight communication design.

Critical weaknesses
  1. The title, abstract and contributions are entirely about noise, outliers and local distortions, yet all experiments use clean benchmarks. There is no noise injection, no outlier ratio sweep, no corrupted client, no comparison with robust (l_2,1/Huber/trimmed) clustering, and no Byzantine/abnormal-client analysis — even though the weighting in Eq. (14) can assign the largest weight to a collapsed (thus corrupted) client.
  2. FedMVL (Huang et al., 2022), FedDMVC (Chen et al., 2023) and TensorFMVC (Feng et al., 2024a) are discussed in the introduction but never compared, while only two FMVC methods are evaluated.
  3. Formulation and notation errors that block reproduction: Eq. (1) minimizes over α_m, which is a global variable under a cross-client constraint; Eq. (2) defines J^((m))circularly J^((m))=f_m (J^((m)),X^((m))) and J^((m)) is never defined; the dual variable W appears in Eqs. (4)–(6) without definition;
Questions for authors

What is the exact objective of ablation Case 1, and how do you exclude degeneration as the cause of its drop?

Reviewer_1Ck1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a robust federated multi-view clustering method (RFMVC) to effectively alleviate the degradation of clustering performance caused by noise, outliers, and local perturbations in distributed environments. Specifically, the proposed method first learns denoised feature representations on each client and further learns clustering structures in the denoised feature space, thereby preserving the key structural information of the original data while suppressing the influence of noise on clustering results. Then, on the server side, the tensor Schatten p-norm is introduced to impose a global consistency constraint on the cluster indicator matrices of different clients. Combined with an adaptive client weighting mechanism, the method integrates multi-client information to enhance the consistency and robustness of clustering structures across clients.

Critical strengths
  1. This paper proposes a robust federated multi-view clustering framework. The motivation is well justified, and the introduction clearly highlights the limitations of existing federated multi-view clustering methods.
  2. This paper introduces the tensor Schatten p-norm to model higher-order correlations among the clustering structures of different clients, enhancing cross-client consistency and information complementarity.
  3. This paper conducts comparative experiments on six real-world multi-view datasets and achieves promising clustering performance. The experimental evaluation is relatively comprehensive, including ablation studies and parameter sensitivity analyses.
Critical weaknesses
  1. The paper provides limited discussion of related work on existing federated multi-view clustering methods.
  2. Although the effectiveness of denoised representation learning has been demonstrated experimentally, the analysis of how it improves sample distributions and clustering structures remains insufficient.
  3. The optimization process for the denoised representation Z(m)Z^{(m)} is described rather briefly, and the motivation and justification for the corresponding update strategy are not sufficiently explained.
Questions for authors
  1. The authors are encouraged to further expand the discussion of related work in federated multi-view clustering.
  2. The authors are encouraged to provide detailed parameter settings for each dataset, including the specific values of key hyperparameters and the rationale behind their selection.
  3. Regarding the issue raised in Weakness 2, the authors are encouraged to provide additional visualization analyses to more intuitively illustrate the effectiveness of the proposed mechanism.
Reviewer_aJgEOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Multi-view clustering (MVC) has been widely studied, and the authors in this paper considers a special application of MVC algorithms, i.e., the distributed scenarios, and thus, the authors develop a federated MVC method. The authors further discuss the problem due of the negative influence due to the noises or outliers, and this issue is generally ignored by current similar works. Besides, the authors further incorporate some specific designs. As a tensorial method, this paper incorporates schatten p-nrom for cross-client multi-view information fusion, which benefits the consistent and complementary information learning. The paper also has some weaknesses in terms of clarification, which will be discussed in the following part.

Critical strengths
  1. The paper integrates local denoised representation learning and clustering structure learning into a unified optimization framework, combined with global consistency modeling based on the tensor Schatten p-norm, thereby improving the robustness and consistency of clustering results.

  2. The paper designs an adaptive client weighting mechanism to dynamically adjust the contributions of different clients, reducing the influence of low-quality clients on global clustering results.

  3. The paper adopts an efficient optimization strategy based on cluster statistics to avoid constructing large-scale distance matrices. Complexity analysis and convergence experiments further demonstrate the efficiency and stability of the proposed algorithm.

Critical weaknesses

1.The paper applies adaptive client weights only to the local clustering loss, while imposing a separate constraint on the data fidelity term, but provides insufficient justification for this design choice.

2.The paper integrates local denoised representation learning with global consistency modeling, but lacks a clear explanation of how global clustering information influences the client-side denoising process.

3.The paper uses local clustering structures to guide denoised representation learning, but does not sufficiently discuss the potential effects of incorrect cluster assignments or the mechanisms used to mitigate them.

4.The Abstract is unclear and does not cover all the design of the proposed method. Please have a careful check and revision on it.

Questions for authors

1.The paper applies adaptive client weights only to the local clustering loss while imposing a separate constraint on the data fidelity term. Could the authors clarify the rationale behind this design choice and explain why adaptive weighting is not applied to the data fidelity term as well?

2.The paper integrates local denoised representation learning with global consistency modeling. Could the authors explain in greater detail how global clustering information influences or guides the client-side denoising process?

3.The paper uses local clustering structures to guide denoised representation learning. Could the authors discuss the potential effects of incorrect cluster assignments on this process and describe the mechanisms used to mitigate such errors?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.