acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Semi-Supervised Federated Multi-View Clustering

Abstract

Federated multi-view clustering enables collaborative clustering of multi-view data distributed across isolated clients while preserving data privacy. However, existing methods largely rely on unsupervised learning, which may overlook available prior information and consequently limit the discriminability of local representations and the reliability of global consensus. To address this issue, we introduce supervisory information into clustering process and propose Adaptive Semi-Supervised Federated Multi-View Clustering (ASFMVC). Specifically, to reduce the dependence on intrinsic data structure alone, we introduce a label-guided representation generation mechanism that incorporates the observed labels into representation learning. Since different views may provide clustering structures of unequal quality, we further design a consensus-aware adaptive weighting strategy to adjust their contributions according to their discrepancies from the current consensus. Moreover, the same supervision is transformed into signed pairwise relations to guide consensus refinement through must-link and cannot-link constraints. These components are integrated into a unified alternating optimization framework with compact communication and of linear complexity. Experiments on six benchmark datasets demonstrate the effectiveness of the proposed method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
federated multi-view clustering, semi-supervised learning, multi-view clustering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HeDyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ASFMVC, a semi-supervised method for vertical federated multi-view clustering, in which each of V clients holds one view of the same N samples, with labels available for a subset of samples. The authors argue that existing federated multi-view clustering methods rely largely on unsupervised or self-supervised signals and thus fail to exploit available prior knowledge.

Building on the communication-efficient federated multi-view clustering framework CeFMC (Liu et al., 2026a), the proposed method introduces three components: (1) label-guided representation learning that incorporates observed labels into each client's local representation, (2) adaptive client weighting based on agreement with the global consensus, and (3) signed pairwise constraints derived from the same labels to refine consensus clustering through must-link and cannot-link relations. The algorithm alternates between client- and server-side updates while retaining compact communication and claimed linear computational complexity in N. Experiments on six benchmark datasets using a 20% label ratio show improvements over unsupervised federated baselines but mixed results against centralized semi-supervised methods. Additional evaluations include ablation studies, varying label ratios, convergence curves, and parameter sensitivity analyses.

Critical strengths
  1. Practically relevant problem setting.The paper addresses an important gap in existing federated multi-view clustering methods, which largely rely on unsupervised or self-supervised learning despite the potential availability of partial labels. This is particularly relevant in applications such as healthcare, where data may be distributed across institutions and only a subset of cases have expert annotations. A well-designed method for incorporating such supervision into federated clustering would be valuable to the community.

  2. Efficient integration of supervision. The proposed method incorporates supervision through local representation learning and server-side consensus refinement while largely preserving the compact communication structure of CeFMC. The alternating updates remain computationally efficient, with no substantial additional per-round communication overhead. This is a useful practical feature given the importance of communication efficiency in federated learning, although the underlying efficiency is primarily inherited from the published CeFMC framework.

Critical weaknesses
  1. Limited novelty relative to CeFMC and insufficient evidence for the proposed components. The method is largely built on the published CeFMC framework (Liu et al., 2026a) with three additive terms, and setting the new terms to zero recovers CeFMC exactly. The underlying objective structure, initialization, alternating update scheme, and communication and complexity analyses are substantially inherited from CeFMC, but the manuscript does not sufficiently acknowledge or distinguish these inherited elements. In particular, the third listed contributio( efficient alternating optimization, compact communication, and linear computational complexity) largely belongs to CeFMC. Moreover, the ablation studies suggest that much of the improvement comes from the must-link/cannot-link constraints, a standard semi-supervised clustering technique, while the other two components provide relatively small gains on several datasets. Consequently, it remains unclear whether the proposed combination constitutes a sufficiently substantial methodological advance beyond incorporating established semi-supervised techniques into CeFMC.

  2. Insufficient theoretical justification and experimental validation. The adaptive weights are not derived from the stated objective, and weighting views by agreement with the current consensus does not necessarily reflect their reliability; it may reinforce early errors by upweighting views that agree with an incorrect consensus. The convergence analysis assumes fixed weights and does not establish convergence of the full algorithm, in which the weights are updated every outer iteration. Empirically, ASFMVC uses 20% labeled samples while all federated baselines are unsupervised, so the reported improvements cannot separate the benefit of the proposed mechanisms from the benefit of the additional labels.

Questions for authors
  1. Are the labeled samples included when computing ACC, NMI, and Purity? As this is important for determining whether the reported improvements reflect better clustering of previously unseen labels rather than agreement with labels already used during training.
  2. The hyperparameter values in Table 7 vary substantially across datasets. How were these values selected? Were ground-truth labels from the evaluation samples used for tuning?
  3. Could the authors compare ASFMVC against CeFMC augmented only with the signed pairwise-constraint term, without label-guided representation learning or adaptive weighting? The existing ablations suggest that signed constraints account for much of the improvement on several datasets, but they do not evaluate this simpler model directly. This comparison is essential for establishing the incremental contributions of the other two components.
  4. The convergence analysis establishes monotonicity only when the adaptive weights are fixed. Does the objective remain non-decreasing when these weights are refreshed after each outer iteration? Can the adaptive weighting rule be derived from a single well-defined optimization objective, or can the authors otherwise justify convergence of the complete algorithm?
  5. The weight on the signed-constraint term is small across all datasets (2^-8 to 2^-6), yet removing the signed-constraint term produces some of the largest performance drops in the ablation study. How does the magnitude of this term compare with the other objective components during optimization?
  6. Where are the observed labels stored, and which parties have access to them? The formulation appears to require supervision at both the clients and the server, and initialization involves transmitting each client’s full initial representation to the server. How do these requirements affect the claimed privacy preservation?
  7. Given the claimed linear computational complexity and inherited communication-efficient structure, the authors should also consider those larger datasets used in CeFMC. Could the authors provide results including runtime and communication cost, to support the scalability claims and enable a more direct comparison with CeFMC?
4 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q2

We thank Reviewer HeDy for the constructive review. We provide point-by-point responses to the concerns raised as follows:

Q1: Novelty beyond the existing framework

A1: We acknowledge that ASFMVC inherits the CeFMC framework, and we do not claim its existing architecture and optimization mechanisms as our contributions. To the best of our knowledge, existing federated multi-view clustering methods predominantly focus on unsupervised learning, and we have not identified an existing semi-supervised federated MVC method directly applicable to our vertical federated setting. Meanwhile, centralized semi-supervised MVC methods assume joint access to all views, making their direct application to vertically partitioned data challenging. Therefore, the key challenge is to use supervision across separate views at both local and global levels, without repeatedly transmitting full representations. Rather than simply combining existing techniques, we adapt three mechanisms to the federated framework.

First, LGR integrates the observed-label matrix into the orthogonality-constrained local representation objective. Accordingly, we modify the client-side SVD update by introducing the supervision-dependent term ηYo(Yo⊤Hv)\eta Y_o(Y_o^\top H_v), enabling local representations to exploit labels without transmitting them to the server.

Second, AW is incorporated into the original consensus alignment through discrepancy-dependent coefficients αv\sqrt{\alpha_v}. This modifies both local pseudo-label optimization and server-side consensus construction, allowing client contributions to be adjusted using existing clustering information without additional high-dimensional communication.

Third, signed constraints are adapted to the server-side discrete consensus update through a balanced and normalized signed graph. The resulting label-update rule jointly considers weighted client agreement and must-link/cannot-link relationships.

Together, these modifications establish a optimization process in which supervision influences both local representations and global consensus, while the updated consensus guides subsequent client optimization, forming an iterative supervision-feedback loop rather than three isolated additions.

In the revised manuscript, we will explicitly acknowledge and cite the components inherited from CeFMC, and more clearly distinguish them from our methodological extensions. Furthermore, by introducing limited supervision into federated multi-view clustering, ASFMVC provides a foundation for future research on supervision-enhanced federated clustering, including more flexible supervision mechanisms and federated optimization strategies.

Q2: Limited incremental contributions of LGR and AW

A2: We acknowledge that the improvements from LGR and AW are relatively modest under some experimental settings. However, their contributions vary with the amount of available supervision. To further investigate this behavior, we conducted additional ablation experiments on Reuters and leave experiments on other datasets for future work due to the tight time limit during the rebuttal period.

Table 1. Ablation study on Reuters under different label ratios (ACC).

Ratiow/o Sw/o LGRw/o AWASFMVC
1%47.84±0.1347.94±0.4347.39±0.1748.04±0.30
5%47.94±0.1048.03±0.5547.72±0.3448.70±0.70
10%48.04±0.1950.00±0.3648.50±0.5150.05±0.80
20%48.18±0.2752.97±2.2449.43±0.4153.00±2.27
30%48.22±0.1053.03±3.0850.17±0.5255.47±3.20
40%48.39±0.1753.78±3.2950.83±3.4055.77±3.54

At 20%, 30%, and 40%, removing AW lowers ACC by 3.57, 5.30, and 4.94 points; removing LGR lowers it by 2.44 and 1.99 points at 30% and 40%. The three components serve different roles, so their effects on final ACC don't need to be equally large. These additional experiments indicate that LGR and AW can provide meaningful improvements under certain supervision conditions.

AuthorsOpenReview ↗

Rebuttal Part 2: Responses to Q3–Q6

Q3: Missing semi-supervised federated baseline

A3: As discussed in A1, existing federated MVC methods predominantly focus on unsupervised learning, and we have not identified a directly applicable semi-supervised federated MVC baseline. Therefore, to enable a more meaningful comparison under matched supervision and federated settings, we construct CeFMC+S by incorporating must-link and cannot-link constraints into the original CeFMC framework, using the same 20% labeled-data budget.

Table 2. ACC comparison under a 20% label ratio.

DatasetCeFMCCeFMC+SASFMVC
BBC77.24±0.0581.43±1.2084.26±0.50
Reuters47.27±0.3049.23±0.4053.00±2.27
HW94.17±0.2195.44±0.2395.59±0.26
Scene-1530.31±4.8038.01±4.9438.20±4.84
Flower1733.93±2.0540.08±1.9640.90±2.08
ORL71.88±2.6172.92±2.4973.72±2.59

CeFMC+S consistently improves upon CeFMC, demonstrating the benefit of incorporating supervision into federated clustering. Besides, ASFMVC achieves higher ACC than CeFMC+S, indicating that the proposed framework offers benefits beyond supervision alone. We will include this additional baseline in the revised manuscript.

Q4: Whether the weighting rule may reinforce early errors

A4: To investigate this particular aspect, we conducted additional noise experiments on Reuters. Specifically, we injected Gaussian noise into the fifth view (View 5) at different noise levels and compared ASFMVC with its variant without adaptive weighting (w/o AW).

Table 3. Noise experiments on Reuters.

Noise level σ\sigmaView discrepancyFusion weight (%)ASFMVC ACC (%)w/o AW ACC (%)
016.3422.8153.00±2.2749.43±0.41
0.519.4520.6952.44±1.2248.88±0.57
1.020.2920.5252.04±2.1848.55±2.61

As the noise level increases, the discrepancy between the corrupted view and the global consensus increases from 16.34 to 20.29, while its normalized fusion weight declined from 22.81% to 20.52%. Moreover, ASFMVC consistently achieves higher mean ACC than w/o AW under the evaluated noise levels. These results indicate that the adaptive weighting mechanism can respond to the degradation of a view by reducing its relative contribution.

Q5: Whether labeled samples are included in ACC/NMI/Purity

A5: Yes. We compute ACC, NMI, and Purity over all samples. The same evaluation protocol is consistently applied to ASFMVC and all compared methods, ensuring a fair comparison under a consistent evaluation setting.

Q6: Hyperparameter selection

A6: We use the ground-truth labels to guide hyperparameter selection based on clustering performance. Following the parameter search and sensitivity analysis practices adopted in existing semi-supervised multi-view clustering studies, such as FSSMSC (Ling et al., 2025) and Hypergraph Learning-Based Semi-Supervised Multi-View Spectral Clustering (2023), we examined candidate hyperparameter configurations based on their clustering performance. To further examine the reasonableness of the dataset-specific settings, we conducted additional parameter sensitivity experiments on all six datasets. The results demonstrate the robustness of ASFMVC to hyperparameter variations and support the reasonableness of our parameter selection strategy. We will include these additional results in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal Part 3: Responses to Q7–Q9

Q7: Magnitude of the signed-constraint term

A7: We evaluated the five components of Eq. (14) on four datasets. Let T1T_1–T5T_5 denote the representation, label-guidance, local-clustering, weighted-consensus, and signed-constraint terms, respectively, before applying their respective external coefficients. Table 4 reports these unscaled terms, while Table 5 reports their weighted contributions to the objective, where J1=T1J_1=T_1, J2=ηT2J_2=\eta T_2, J3=λT3J_3=\lambda T_3, J4=βT4J_4=\beta T_4, and J5=γT5J_5=\gamma T_5.

Table 4. Magnitudes of the five objective terms before applying their respective coefficients.

DatasetT1T_1T2T_2T3T_3T4T_4T5T_5
BBC246.49199.87415.0370.33566.131
Reuters530.115110.68846.6140.491103.121
Scene-1513396.918136.14618.6230.305420.525
Flower178132.797109.87246.8740.967131.100

Table 5. Weighted contributions of the five objective terms after applying their respective coefficients.

DatasetJ1J_1J2J_2J3J_3J4J_4J5J_5
BBC246.4916.24230.0740.6700.258
Reuters530.1150.432186.4571.9630.403
Scene-1513396.9180.13374.4931.2206.571
Flower178132.7971.717187.4953.8662.048

As shown in Table 5, the signed-constraint term makes a relatively small contribution to the overall objective magnitude, partly due to the dominant representation term. However, its contribution is not negligible relative to the components directly involved in global consensus optimization. In particular, ∣J5∣/∣J4∣|J_5|/|J_4| ranges from 0.205 to 5.387, indicating that the signed-constraint term can be comparable to or even larger than the weighted consensus term in magnitude.

More importantly, the influence of signed constraints depends on their effect on discrete label assignments rather than their overall objective magnitude. As shown in Eq. (21), signed constraints introduce an additional correction to the consensus scores. When the scores of competing clusters are close, even a relatively small correction may change the final assignment. Our additional experiments show that the magnitude of the signed-constraint correction can be substantial relative to the score differences between competing clusters.

Q8: Location of labels, supervision and privacy

A8: The observed labels are stored locally on the clients, while the must-link/cannot-link relationships are maintained on the server. Both forms of supervision are treated as prior knowledge and directly incorporated into the corresponding client-side and server-side optimization objectives. We do not transmit observed labels or must-link/cannot-link relationships between clients and the server during communication.

Q9: Large-scale dataset results and efficiency

A9: We conducted additional experiments on YouTubeFace10 (38,654 samples, 4 views, 10 clusters). We compared ASFMVC with both unsupervised federated MVC methods and centralized semi-supervised MVC methods.

Table 6. Performance and efficiency comparison with unsupervised federated MVC methods on YouTubeFace10.

MethodACC (%)NMI (%)Purity (%)Time (s)Comm. (MB)
FedCMv61.53±1.7865.02±1.0766.20±1.1020.50326.56
FedFCMv26.45±0.1016.67±0.1026.45±0.10394.46630.83
FedMVLOOMOOMOOMOOMOOM
CeFMC76.86±1.6074.86±1.9779.81±1.496.8016.71
ASFMVC80.02±1.4679.25±1.8182.43±1.7020.6215.47

Table 7. Performance and running time comparison with centralized semi-supervised MVC methods on YouTubeFace10.

MethodACC (%)NMI (%)Purity (%)Time (s)
AMGLOOMOOMOOMOOM
MLANOOMOOMOOMOOM
SMGCOOMOOMOOMOOM
ASFMVC80.02±1.4679.25±1.8182.43±1.7020.62

OOM (Out of Memory) indicates that the method failed to complete the experiment due to insufficient memory.

As shown in Table 6, ASFMVC achieves the highest ACC, NMI, and Purity among the compared methods, outperforming all unsupervised federated MVC baselines, while maintaining comparable communication overhead. ASFMVC also maintains a reasonable running time compared with the other federated methods, demonstrating its practical computational efficiency on large-scale datasets.

Furthermore, Tables 6 and 7 show that ASFMVC successfully processes YouTubeFace10, whereas FedMVL, AMGL, MLAN, and SMGC fail to complete the experiments due to insufficient memory. These results demonstrate the practical scalability of ASFMVC and its ability to handle large-scale multi-view clustering while maintaining competitive clustering performance and low communication overhead. We will include these additional results in the revised manuscript.

Dear Reviewer,

Thank you again for your valuable comments and suggestions.

As the discussion period is approaching its end, we hope that our detailed responses and additional experimental results have adequately addressed your concerns. If there are any remaining questions or issues that require further clarification, please let us know. We would be happy to provide additional explanations.

Thank you for your time and effort in reviewing our paper !

Reviewer_jKyTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ASFMVC, a semi-supervised framework for federated multi-view clustering. Clients learn label-guided representations and transmit pseudo-labels and centroids to the server. The server combines adaptive client weighting with a normalized signed graph encoding must-link and cannot-link constraints to refine the consensus clustering. The method uses alternating optimization, with reported linear complexity in the number of samples and reduced communication costs.

Critical strengths

1.The manuscript is clearly organized, with a notation table, detailed optimization derivations, explicit stopping criteria, and ten repetitions per experiment. 2.The communication analysis is concrete and verifiable, showing how transmitting assignments and centroids reduces costs relative to representation-sharing methods.

Critical weaknesses
  1. The main comparisons confound methodological gains with access to supervision. In Table 2, ASFMVC uses 20% labeled data, whereas the federated baselines are unsupervised. These results therefore do not isolate the contribution of the proposed framework. Against centralized semi-supervised methods in Table 3, ASFMVC performs worse on five of six datasets, with substantial gaps on Scene-15 (38.20 vs. 50.54) and Flower17 (40.90 vs. 63.39). Such comparisons provide useful context, but neither comparison establishes the method’s advantage under matched conditions.
  2. A semi-supervised federated baseline is missing. The most informative comparison would give competing methods the same supervision and federated constraints. Suitable controls could include a federated constrained-clustering method or a simplified version of ASFMVC that uses labels without the signed graph and adaptive weighting. Without this comparison, it remains unclear how much the proposed design adds beyond access to labels.
  3. The privacy claims exceed the evidence provided. Avoiding raw-feature transmission does not by itself establish privacy preservation. Clients transmit per-sample cluster assignments at every round, potentially exposing partition structure and its evolution. The paper should specify a threat model and assess potential leakage, or narrow its claims to the concrete property that raw features remain local.
  4. Several component gains are small relative to the reported variability.
  5. The novelty needs sharper positioning. Label-guided representation learning, inverse-deviation weighting, and signed-graph constraints are established techniques. The contribution appears to lie primarily in their integration into a communication-efficient federated framework. The paper should clarify what this integration enables beyond existing formulations and discuss whether the weighting rule can allow a single client to dominate.
Questions for authors

see above

3 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q3

We thank Reviewer jKyT for the constructive review. We provide point-by-point responses to the concerns raised as follows:

Q1: Unmatched supervision conditions and the absence of a semi-supervised federated baseline.

A1: To the best of our knowledge, existing federated multi-view clustering methods predominantly focus on unsupervised learning, and we have not identified an existing semi-supervised federated MVC method directly applicable to our vertical federated setting. Although there is no suitable matched baseline identified in the existing literature, we additionally construct CeFMC+S, a semi-supervised federated baseline that incorporates must-link and cannot-link constraints into the original CeFMC framework, using the same 20% labeled-data budget.

Table 1. ACC comparison under a 20% label ratio.

DatasetCeFMCCeFMC+SASFMVC
BBC77.24±0.0581.43±1.2084.26±0.50
Reuters47.27±0.3049.23±0.4053.00±2.27
HW94.17±0.2195.44±0.2395.59±0.26
Scene-1530.31±4.8038.01±4.9438.20±4.84
Flower1733.93±2.0540.08±1.9640.90±2.08
ORL71.88±2.6172.92±2.4973.72±2.59

CeFMC+S consistently improves upon CeFMC, demonstrating the benefit of incorporating supervision into federated clustering. Besides, ASFMVC achieves higher ACC than CeFMC+S, indicating that the proposed framework offers benefits beyond supervision alone. We will include this additional baseline in the revised manuscript.

Q2: Privacy threat model and potential leakage.

A2: We consider an honest-but-curious server that faithfully executes the prescribed protocol but may analyze received information to infer private properties of client data. Specifically, the server can observe low-dimensional representations, pseudo-labels, and server-side signed pairwise constraints. It's potential inference targets include client-specific feature properties, sample relationships, and local clustering structures. However, raw feature matrices and observed sample labels remain at their respective clients and are not directly transmitted to the server. We assume that clients follow the protocol and exclude active attacks, client compromise, and collusion from the considered threat model. Under this threat model, intermediate information may reveal certain structural properties of local data, although the server cannot directly access raw features.

Q3: Small ablation gains relative to variability.

A3: We acknowledge that the improvements from LGR and AW are relatively modest under some experimental settings. However, their contributions vary with the amount of available supervision. To further investigate this behavior, we conducted additional ablation experiments on Reuters and leave experiments on other datasets for future work due to the tight time limit during the rebuttal period.

Table 2. Ablation study on Reuters under different label ratios (ACC).

Ratiow/o Sw/o LGRw/o AWASFMVC
1%47.84±0.1347.94±0.4347.39±0.1748.04±0.30
5%47.94±0.1048.03±0.5547.72±0.3448.70±0.70
10%48.04±0.1950.00±0.3648.50±0.5150.05±0.80
20%48.18±0.2752.97±2.2449.43±0.4153.00±2.27
30%48.22±0.1053.03±3.0850.17±0.5255.47±3.20
40%48.39±0.1753.78±3.2950.83±3.4055.77±3.54

At 20%, 30%, and 40%, removing AW lowers ACC by 3.57, 5.30, and 4.94 points; removing LGR lowers it by 2.44 and 1.99 points at 30% and 40%. The three components serve different roles, so their effects on final ACC don't need to be equally large. These additional experiments indicate that LGR and AW can provide meaningful improvements under certain supervision conditions.

Due to the character limit, our responses to Q4 and Q5 are provided in a separate follow-up comment below.

AuthorsOpenReview ↗

Rebuttal Part 2: Responses to Q4–Q5

Due to the character limit, we provide our responses to Q4 and Q5 in this follow-up comment.

Q4: What does the integration enable beyond existing formulations.

A4: Federated MVC is largely unsupervised, whereas centralized semi-supervised MVC assumes joint access to all views. The challenge is to use supervision across separate views at both local and global levels, without repeatedly transmitting full representations. Rather than directly combining existing techniques, we adapt three mechanisms to the federated alternating optimization.

First, LGR integrates the observed-label matrix into the orthogonality-constrained local representation objective. Accordingly, we modify the client-side SVD update by introducing the supervision-dependent term ηYo(Yo⊤Hv)\eta Y_o(Y_o^\top H_v), enabling local representations to exploit labels without transmitting them to the server.

Second, AW is incorporated into the original consensus alignment through discrepancy-dependent coefficients αv\sqrt{\alpha_v}. This modifies both local pseudo-label optimization and server-side consensus construction, allowing client contributions to be adjusted using existing clustering information without additional high-dimensional communication.

Third, signed constraints are adapted to the server-side discrete consensus update through a balanced and normalized signed graph. The resulting label-update rule jointly considers weighted client agreement and must-link/cannot-link relationships.

Together, these modifications establish a optimization process in which supervision influences both local representations and global consensus, while the updated consensus guides subsequent client optimization, forming an iterative supervision-feedback loop rather than three isolated additions.

Q5: Whether the weighting rule can allow a single client to dominate.

A5: We examine the normalized fusion coefficient qj=αj/∑v=1Vαvq_j=\sqrt{\alpha_j}/\sum_{v=1}^{V}\sqrt{\alpha_v} during optimization. Table 3 reports the actual fusion weights of all three views at each iteration on ORL. Due to the character limit, we present only the ORL results here; the per-iteration fusion weights of all views on the other datasets will be provided in the appendix of the revised manuscript.

Table 3. Evolution of client fusion weights on ORL (%).

IterationView 1View 2View 3
133.33333.33333.333
233.93932.56533.496
333.72832.64033.631
433.60832.71833.674
533.57632.72333.701
633.51732.73933.744

The weights remain close to uniform allocation (33.33%) throughout optimization, with a maximum of 33.939%. These results indicate that no single view dominates the clustering process in this run.

Thank you for your response. I believe the score I assigned is reasonable at this stage, given that the reviewers’ scores are generally low overall. While Reviewer 1 recommended clear acceptance, I would lean toward a more moderate rating based on my own reading of the paper.

Reviewer_cboKOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an Adaptive Semi-supervised Federated Multi-view Clustering (ASFMVC) method. The approach leverages limited labeled data to guide local representation learning on clients and employs an adaptive weighting strategy to adjust the contribution of different views to the global clustering process. Furthermore, the authors incorporate "must-link" and "cannot-link" constraints to refine the global clustering results. Experimental evaluations conducted on datasets demonstrate that the proposed method outperforms the compared unsupervised federated multi-view clustering methods.

Critical strengths

This paper incorporates supervisory information into both local representation learning and global consensus clustering, while employing an adaptive weighting strategy to adjust the contribution of different views to the final clustering results. These components are integrated into a unified optimization framework.

During the iterative optimization process, this method does not require the transmission of complete local representations; instead, it achieves collaborative optimization by exchanging cluster labels and centroid information. The paper also provides a theoretical analysis of communication overhead along with experimental comparisons, demonstrating that the method offers advantages in terms of communication efficiency.

Critical weaknesses

The paper constructs must-link and cannot-link constraints using the sample pair set (E) in Eq. (8), but it does not clearly explain how these pairs are selected. In particular, it is unclear whether all labeled sample pairs are used, or how the numbers and proportions of the two types of constraints are determined. The authors are encouraged to clarify the constraint construction process.

The paper adjusts the adaptive weights based on the discrepancies between individual views and the global consensus, but lacks an evaluation of this strategy under different view qualities. For example, when some views contain substantial noise, can the proposed strategy effectively reduce the influence of low-quality views?

As shown in Table 4, removing the LGR and AW modules leads to only minor performance drops on some datasets. The authors are encouraged to further explain these results to better clarify the actual contributions of these modules.

Questions for authors
  1. How are the labeled samples selected in the experiments? Do different random selections of labeled samples have a significant impact on the clustering results?

  2. When the label ratio is low, the labeled samples may not cover all classes. How does the proposed method handle this situation? Would it affect local representation learning and the final clustering results?

  3. The paper assumes that the known labels are correct. If some labeled samples have incorrect labels, can the proposed method still achieve good clustering performance?

  4. Table 5 reports the results under different label ratios, but only on BBC, Reuters, and Flower17. In addition, the parameter sensitivity analysis is only conducted on BBC. The authors are encouraged to provide results on other datasets to further evaluate the stability of the proposed method.

2 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q4

We thank Reviewer cboK for the constructive review. We provide point-by-point responses to the concerns raised as follows:

Q1: How are labeled samples and must-link/cannot-link constraints selected? How sensitive are the results to random selections?

A1: In our implementation, labeled samples are randomly selected from the dataset. Given a label ratio ρ\rho and NN samples, we select Nl=ρNN_l=\rho N labeled samples. The same supervision is then transformed into pairwise constraints among the selected labeled samples. For each queried pair (i,j)(i,j), samples from the same class form a must-link constraint (Sij=+1S_{ij}=+1), while samples from different classes form a cannot-link constraint (Sij=−1S_{ij}=-1). The implementation considers up to 10510^5 pairs. When the number of candidate pairs exceeds this limit, 10510^5 pairs are randomly selected. Consequently, the numbers of must-link and cannot-link constraints are determined by the sampled pairs rather than a predefined ratio. To prevent the more frequent constraint type from dominating optimization, we apply the type-balancing coefficients in Eq. (9), followed by symmetric absolute-degree normalization in Eq. (11).

Regarding the impact of random label selection, our reported results are averaged over ten independent runs with different randomly selected labeled subsets. The reported standard deviations reflect the empirical variability across these runs. For example, ASFMVC achieves 84.26±0.50% ACC on BBC and 95.59±0.26% on HW, indicating relatively stable performance under the evaluated random selections. We will clarify the sampling protocol and provide more detailed information about the constraint counts in the revised manuscript.

Q2: What happens when the labeled samples do not cover all classes?

A2: Our existing low-label-ratio experiments provide empirical evidence that the framework remains functional under scarce supervision. At a 1% label ratio, ASFMVC achieves 78.28±1.10% ACC on BBC, 48.04±0.30% on Reuters, and 33.99±2.10% on Flower17. For reference, the corresponding unsupervised CeFMC results are 77.24±0.05%, 47.27±0.30%, and 33.93±2.05%, respectively. These results indicate that ASFMVC can still produce meaningful clustering results without full class coverage. Nevertheless, the performance advantages are modest at very low supervision ratios, and the absence of labels for certain classes may reduce the effectiveness of supervision-guided optimization.

Q3: If some labeled samples have incorrect labels, can the proposed method still achieve good clustering performance?

A3: We will conduct additional experiments by randomly assigning incorrect class labels to different proportions of the labeled samples while keeping the total number of labeled samples unchanged. We will evaluate the clustering performance under different label corruption rates and compare the results with those obtained using correct labels. The corresponding experimental results will be included in the appendix of the revised manuscript.

Q4: Additional label-ratio and parameter-sensitivity evaluations.

A4: We have conducted additional experiments under different label ratios on the remaining three datasets: HW, Scene-15, and ORL. The corresponding results are reported in Table 1.

Table 1. Clustering performance under different label ratios (mean ± std, %).

DatasetMetric1%5%10%20%30%
HWACC94.23±0.2194.61±0.2195.10±0.2795.59±0.2696.08±0.26
NMI87.57±0.2688.35±0.3789.21±0.4790.12±0.5191.13±0.55
Purity94.23±0.2194.61±0.2195.10±0.2795.59±0.2696.08±0.26
Scene-15ACC30.34±5.0031.61±4.9633.86±4.5938.20±4.8442.37±4.20
NMI27.03±5.2228.37±5.1030.16±5.0033.93±4.9437.94±4.75
Purity32.66±5.3533.93±5.1735.84±4.9339.97±4.9243.90±4.31
ORLACC71.98±2.7572.48±2.8272.78±2.7573.73±2.5974.83±2.30
NMI85.45±1.7085.61±1.8485.67±1.6386.13±1.7886.87±1.56
Purity75.55±2.3475.93±2.4376.05±2.2976.73±2.5077.73±2.16

Due to the character limit, we present only these additional label-ratio results here. The complete label-ratio results and parameter-sensitivity analysis across all six datasets will be provided in the appendix of the revised manuscript.

Our responses to Q5 and Q6 are provided in a separate follow-up comment below.

AuthorsOpenReview ↗

Rebuttal Part 2: Responses to Q5–Q6

Q5: Can adaptive weighting reduce the influence of low-quality views?

A5: We conducted additional noise experiments on Reuters. Specifically, we injected Gaussian noise into the fifth view (View 5) at different noise levels and compared ASFMVC with its variant without adaptive weighting (w/o AW).

Table 2. Noise experiments on Reuters.

Noise level σ\sigmaView discrepancyFusion weight (%)ASFMVC ACC (%)w/o AW ACC (%)
016.3422.8153.00±2.2749.43±0.41
0.519.4520.6952.44±1.2248.88±0.57
1.020.2920.5252.04±2.1848.55±2.61

As the noise level increases, the discrepancy between the corrupted view and the global consensus increases from 16.34 to 20.29, while its normalized fusion weight decreases from 22.81% to 20.52%. Moreover, ASFMVC consistently achieves higher mean ACC than w/o AW under the evaluated noise levels. These results indicate that the adaptive weighting mechanism can respond to the degradation of a view by reducing its relative contribution.

Q6: Removing the LGR and AW modules leads to only minor performance drops on some datasets.

A6: We acknowledge that the improvements from LGR and AW are relatively modest under some experimental settings. However, their contributions vary with the amount of available supervision. To further investigate this behavior, we conducted additional ablation experiments on Reuters and leave experiments on other datasets for future work due to the tight time limit during the rebuttal period.

Table 3. Ablation study on Reuters under different label ratios (ACC).

Ratiow/o Sw/o LGRw/o AWASFMVC
1%47.84±0.1347.94±0.4347.39±0.1748.04±0.30
5%47.94±0.1048.03±0.5547.72±0.3448.70±0.70
10%48.04±0.1950.00±0.3648.50±0.5150.05±0.80
20%48.18±0.2752.97±2.2449.43±0.4153.00±2.27
30%48.22±0.1053.03±3.0850.17±0.5255.47±3.20
40%48.39±0.1753.78±3.2950.83±3.4055.77±3.54

At 20%, 30%, and 40%, removing AW lowers ACC by 3.57, 5.30, and 4.94 points; removing LGR lowers it by 2.44 and 1.99 points at 30% and 40%. The three components serve different roles, so their effects on final ACC don't need to be equally large. These additional experiments indicate that LGR and AW can provide meaningful improvements under certain supervision conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.