acceptodds
Under review as a conference paper at ICLR 2027

Beyond View Heterogeneity: Federated Incomplete Multi-View Clustering with Arbitrary View Subsets

Abstract

Federated multi-view clustering (FedMVC) learns cluster structures from multi-view data distributed across clients without centralizing raw data. Beyond differences in view-specific feature spaces and client view subsets, individual samples may lack views available at their client. Client-level view heterogeneity and sample-level view missingness jointly determine the cross-view supervision and relational evidence available for learning common representations. In this paper, we propose Federated Incomplete Multi-view Clustering via Prediction and Alignment (FIMCPA) for this more general setting. Specifically, masked cross-view predictive pretraining exploits partial joint observations to capture shared information for common representation initialization. Relation-verified contrastive alignment balances sample contributions and uses shared-view agreement to mitigate potential false-negative repulsion, promoting consistency among common and observed-view representations. Observation-weighted aggregation further accounts for differences in available data during model aggregation. These designs allow clients with different observation patterns to contribute to common representation learning without imputing genuinely missing views. Theoretical analysis provides insight into relation verification and its effects on the contrastive objective. Extensive experiments demonstrate superior clustering performance, with further analyses supporting the effectiveness of FIMCPA under varying observation conditions and client compositions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
federated incomplete multi-view clustering, arbitrary view subsets, masked cross-view prediction, contrastive learning, relation verification

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Ui4jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Federated Incomplete Multi-view Clustering via Prediction and Alignment (FIMCPA), addressing client-level view heterogeneity and sample-level view missingness. The framework combines masked cross-view predictive pretraining, sample-balanced contrastive alignment with shared-view relation verification, and observation-weighted model aggregation. A federated Lloyd refinement procedure is further employed for global clustering. Experiments on eight benchmark configurations demonstrate competitive clustering performance under varying missing rates and client compositions.

Critical strengths
  • The masked cross-view predictive pretraining strategy learns shared information from available observations without reconstructing genuinely missing views, providing a reasonable initialization for subsequent representation learning.

  • The framework jointly considers client-level view heterogeneity and sample-level missingness. The relation-verification mechanism further addresses potential false negatives in contrastive learning using shared-view neighborhood information.

  • The paper provides detailed theoretical and implementation analyses. Experiments across eight benchmarks, together with ablation studies, missing-rate evaluations, and client-composition analyses, demonstrate competitive clustering performance.

Critical weaknesses
  • The relation verifier determines neighborhood size based on an approximately balanced-cluster assumption in Eq. (5), while the experiments use class-stratified client partitions. Its effectiveness under label-skewed, naturally non-IID client distributions and imbalanced clusters remains unverified.

  • All evaluated client configurations retain full-view clients. The proposed framework is not tested when only partially overlapping view subsets are available across clients, leaving its effectiveness in the claimed arbitrary-view-subset setting insufficiently established.

  • FIMCPA employs global federated Lloyd refinement, whereas FMCSC and AFMCC retain different client-wise K-means procedures. Without evaluating all learned representations under the same clustering protocol, it is unclear how much of the improvement comes from representation learning rather than clustering refinement.

  • The ablation study in Table 2 does not independently evaluate sample-balanced contrastive alignment or observation-weighted aggregation, although both are central components of FIMCPA. Their individual contributions to the reported improvements remain unclear.

Questions for authors
  • Can the authors evaluate FIMCPA under label-skewed or naturally non-IID client partitions with imbalanced cluster distributions?

  • How does FIMCPA perform without full-view clients when the cross-client view co-observation graph remains connected?

  • Do the reported improvements persist when all methods use the same global clustering initialization and Lloyd refinement procedure?

  • Can the authors provide separate ablations comparing sample-balanced alignment with standard contrastive weighting and observation-weighted aggregation with conventional FedAvg?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Ui4j W1/Q1: Non-IID Client Distributions

Thank you for your careful review and constructive comments. We report experiments on non-IID client distributions and learning without full-view clients, comparisons under matched global clustering, and independent weighting ablations. Below, we explain how the findings support FIMCPA's ability to learn from heterogeneous observations and the quality of its learned common representations.

W1/Q1: Non-IID client distributions. To evaluate learning under local class imbalance, we test Dirichlet non-IID partitions with α=1\alpha=1 on all eight datasets. We retain the original neighborhood and verification rules, together with the existing view heterogeneity and missing observations.

We report IID and non-IID ACC (%) as mean ± sample SD over five runs with different random seeds. ACC change denotes non-IID mean ACC minus IID mean ACC, in percentage points.

SettingUCIHWCaltech-5VOutdoorSceneFashionCCVCIFAR10-MVMNIST-3V
IID ACC85.92 ± 2.2190.20 ± 2.1279.51 ± 3.2173.90 ± 1.7989.81 ± 0.4924.45 ± 0.7566.70 ± 2.6299.00 ± 0.06
Non-IID ACC83.05 ± 6.5084.58 ± 4.4873.46 ± 3.3668.26 ± 2.9285.37 ± 0.8023.40 ± 0.8662.74 ± 3.1698.19 ± 0.10
ACC change-2.87-5.62-6.05-5.64-4.44-1.05-3.96-0.81

Dirichlet non-IID partitioning adds class-distribution heterogeneity to the existing view heterogeneity and missing observations. At the same missing-sample rate, locally underrepresented classes have fewer co-observed samples for prediction and fewer same-class pairs eligible for verification. The method must therefore learn from incomplete views with unequal numbers of observed samples across classes.

Caltech-5V shows the largest decrease (6.05 percentage points). Its 1,400 samples are distributed across clients with heterogeneous subsets of five views. Non-IID class proportions and missing observations further restrict the local samples available for prediction and relation verification, making learning more demanding with this limited sample pool. FIMCPA nevertheless achieves 73.46% ACC in this setting.

Equation (5) controls the candidate-neighborhood size; it does not assign samples to the same class. Under non-IID partitions, candidates must still satisfy mutual-neighbor agreement in every shared observed view, with at least two such views, before triggering negative removal. Even a verified relation is not added as a cross-sample positive. Positive pairs remain representations of the same sample, so their correspondence does not depend on local class proportions.

Candidate generation, negative removal and positive-pair construction therefore have distinct roles: neighborhood size defines candidates, cross-view verification determines negative removal, and sample identity defines positives. A change in neighborhood composition does not directly create cross-sample attraction. MCPR first learns from reconstruction and observed prediction targets; the subsequent contrastive stage combines reconstruction with relation-verified alignment.

Non-IID client distributions further increase the learning difficulty under view heterogeneity and missing observations. Under the Dirichlet setting with α=1\alpha=1, FIMCPA retains the original neighborhood and verification rules and preserves most of its IID performance, with an average ACC decrease of 3.81 percentage points across the eight datasets. Given the additional challenges posed by these combined sources of heterogeneity, we consider this average decrease acceptable in the evaluated setting. These results support FIMCPA's ability to adapt to the non-IID conditions tested here.

AuthorsOpenReview ↗

Response to Reviewer Ui4j W2/Q2: View Availability

W2/Q2: Learning without full-view clients. The original configurations retain full-view clients. To evaluate learning from partial overlap alone, we remove all full-view clients in two additional settings while preserving a connected view co-observation graph.

We evaluate P and P+S on all eight datasets using five paired random seeds for each setting, yielding 80 additional experiments. Ref is the default mixture (full:partial:single=2:1:1); P uses only partial-view clients (0:1:0); P+S mixes partial- and single-view clients equally (0:1:1). Partial-view clients acquire 2 to V−1 views. We retain class-stratified partitions, training budgets and an incomplete-sample rate of r=0.5 on multi-view clients. In every run, the view co-observation graph is connected: views are nodes, and an edge means that at least one sample observes both views on a client. All views are linked by such overlap paths, even though no client observes them all. ACC, NMI and ARI (%) are listed in rows for each dataset, reported as mean ± sample SD over five runs with different random seeds.

DatasetMetricRefPP+S
UCIACC85.92 ± 2.2184.70 ± 1.1279.74 ± 2.17
NMI76.58 ± 1.9874.76 ± 1.3168.21 ± 2.10
ARI72.96 ± 3.3270.77 ± 1.7563.08 ± 3.28
HWACC90.20 ± 2.1290.21 ± 0.8085.62 ± 2.74
NMI81.72 ± 2.7181.41 ± 1.1975.38 ± 1.96
ARI80.19 ± 3.5279.93 ± 1.4572.60 ± 2.92
FashionACC89.81 ± 0.4983.63 ± 1.7778.41 ± 1.20
NMI82.33 ± 0.6576.35 ± 0.7869.66 ± 0.69
ARI79.78 ± 0.8971.29 ± 1.4462.36 ± 1.18
Caltech-5VACC79.51 ± 3.2179.06 ± 4.6274.84 ± 2.21
NMI66.94 ± 2.8068.76 ± 1.8058.84 ± 2.42
ARI63.65 ± 3.6864.54 ± 2.6755.51 ± 2.99
OutdoorSceneACC73.90 ± 1.7970.73 ± 0.8765.77 ± 1.20
NMI57.42 ± 1.7854.26 ± 0.9747.58 ± 1.41
ARI50.85 ± 2.4546.75 ± 1.2639.22 ± 1.80
CIFAR10-MVACC66.70 ± 2.6266.57 ± 2.6165.96 ± 1.24
NMI54.61 ± 0.9053.67 ± 1.4652.50 ± 0.66
ARI45.01 ± 1.4044.37 ± 2.5343.42 ± 1.04
CCVACC24.45 ± 0.7522.58 ± 0.9919.95 ± 0.76
NMI18.78 ± 0.1817.34 ± 0.6514.03 ± 0.43
ARI8.77 ± 0.308.07 ± 0.316.12 ± 0.33
MNIST-3VACC99.00 ± 0.0698.57 ± 0.0298.23 ± 0.06
NMI96.93 ± 0.1495.80 ± 0.0594.97 ± 0.12
ARI97.78 ± 0.1396.84 ± 0.0596.11 ± 0.12

Without any full-view clients, P retains 97.32% of Ref ACC on average, showing that partial-view clients alone support effective clustering under the tested connected view-overlap patterns. P+S additionally includes single-view clients. Its results show that FIMCPA can still learn collaboratively from partial observations on the remaining clients when some clients cannot supply local cross-view prediction targets.

This capability stems from how MCPR exploits local co-observations. Each prediction task uses an actually observed view as its target and other observed views of the same sample as input; neither the sample nor its client needs all views. Removing full-view clients therefore does not remove cross-view predictive supervision. For example, clients with view subsets {1,2} and {2,3} construct their own local prediction tasks. Aggregation of updates to the shared view-2 encoder and fusion parameters connects these local learning tasks, allowing cross-view learning to exploit complementary observations distributed across clients.

Single-view clients do not construct cross-view prediction targets, but update the model through reconstruction and same-sample alignment, and receive and use parameters jointly learned by multi-view clients. Clients in P+S thus contribute to different objectives according to their available observations and collaborate through parameter sharing. Together, the P and P+S results show that full-view clients are not a prerequisite under the tested connected view-overlap patterns: distributed local co-observations still provide effective learning signals for FIMCPA.

AuthorsOpenReview ↗

Response to Reviewer Ui4j W3/Q3: Common Global Clustering

W3/Q3: Common global clustering. To assess representation quality under a matched final clustering procedure, we give FMCSC and AFMCC the same global initialization rule and Lloyd refinement as FIMCPA, and extend this evaluation to the remaining five baselines. The table reports ACC (%) as mean ± sample SD over five random seeds.

We retain training, features, default data partitions and masks. Local K-means provides centers and cluster sizes; count-weighted server-side K-means initializes global centers in each representation space (K-means++: 10 local/100 server restarts). All methods use at most 20 federated Lloyd updates, a convergence tolerance of 10−510^{-5}, the same distance metric, tie-breaking rule and empty-cluster handling, and a single global Hungarian matching. Centers are updated using feature sums and sample counts aggregated across clients for each cluster.

FIMCPA uses HH; FMCSC/AFMCC use learned representations with local clustering replaced. EFDMVC/FreeCSL/HSACC retain their K-means inputs (FreeCSL: observed-view consensus ZZ). DSIMVC uses probabilities averaged over observed views; H2-FedMVC uses final MoE logits. The latter two serve as supplementary diagnostics of prediction-head outputs; H2-FedMVC retains label-assisted training.

DatasetFIMCPAFMCSCAFMCCEFDMVCFreeCSLHSACCDSIMVCH2-FedMVC
UCI85.92 ± 2.2173.07 ± 1.8277.58 ± 0.9276.09 ± 2.6075.89 ± 3.9668.51 ± 5.4153.50 ± 3.3677.86 ± 4.62
HW90.20 ± 2.1277.15 ± 1.9278.34 ± 0.7880.98 ± 4.1482.77 ± 1.8763.59 ± 5.7850.40 ± 3.7680.32 ± 2.54
Fashion89.81 ± 0.4984.80 ± 0.6589.06 ± 0.1784.26 ± 2.0486.95 ± 4.6881.55 ± 5.3764.57 ± 1.1285.97 ± 1.56
Caltech-5V79.51 ± 3.2171.99 ± 1.7974.00 ± 4.3065.64 ± 1.5878.83 ± 3.7633.09 ± 11.4641.59 ± 3.4272.09 ± 6.15
OutdoorScene73.90 ± 1.7962.40 ± 3.3468.90 ± 2.1861.38 ± 2.0953.56 ± 4.0145.42 ± 4.2847.00 ± 2.9057.48 ± 3.01
CIFAR10-MV66.70 ± 2.6253.03 ± 2.4555.02 ± 2.8751.55 ± 3.1162.05 ± 4.4634.85 ± 4.2650.30 ± 1.0041.18 ± 3.93
CCV24.45 ± 0.7519.50 ± 1.0321.57 ± 0.4720.19 ± 0.5821.39 ± 0.9014.95 ± 0.4515.83 ± 0.2115.92 ± 0.77
MNIST-3V99.00 ± 0.0694.77 ± 1.0098.48 ± 0.0298.27 ± 0.0798.53 ± 0.0972.06 ± 2.6596.43 ± 3.9395.17 ± 2.77

FMCSC and AFMCC now receive the same global coordination, yet FIMCPA exceeds them on all eight datasets, by 9.10 and 5.82 percentage points in ACC on average. FIMCPA has the highest mean ACC on all eight datasets. The mean margin over the strongest alternative per dataset is 3.74 points.

Under this protocol, FMCSC, AFMCC and all other baselines update centers using cluster statistics pooled across clients, receiving the same final global coordination. FIMCPA remains ahead, supporting the clustering quality of its learned common representations. Lloyd refinement changes centers and assignments on fixed representations without updating encoders, so this comparison controls the final clustering procedure.

During representation learning, MCPR removes the target view from both input paths and predicts it from other observed views, preventing direct copying of the target. Subsequent alignment connects same-sample view-specific and common representations, while relation verification uses shared-view neighborhood agreement to remove potential false negatives from the contrastive objective. These constraints shape the representation space before final clustering.

Global refinement updates common centers using cluster statistics aggregated across all clients, allowing each update to use information across clients rather than only samples from a single client. The complementary control uses the same FIMCPA representations and initial centers: client-wise refinement retains similar performance, while global refinement provides a further average improvement. For additional evidence on the effect of global refinement, please see our response to Reviewer PBoh, “Fixed-Representation Control.” Together, the comparisons show that FIMCPA's advantage cannot be explained solely by final global coordination; global refinement adds an average benefit on top of the learned common representations and shared initialization.

Reviewer_VU5NOpenReview ↗
Rating
4
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

FIMCPA targets the more general federated multi-view clustering setting in which client-level view heterogeneity and sample-level view missingness coexist, and proposes a "predict-then-align" framework: it first employs masked cross-view predictive pretraining to capture shared information from partial joint observations for initializing the common representation, and then applies relation-verified contrastive alignment to balance sample contributions and to verify alignment quality via shared-view agreement, thereby learning a clustering-friendly common representation without centralizing raw data.

Critical strengths

(1)The paper has a clear and well-defined motivation, consistently incorporating actually observed data. It holds out only observed prediction targets (without predicting them), avoiding reliance on imputed supervision; meanwhile, sample-balanced alignment alleviates the problem of excessive contribution from samples with more observed view

(2)The experimental validation is thorough. FIMCPA achieves the highest average ACC, NMI, and ARI across eight benchmark configurations, with significant improvements on HW and CIFAR10. The ablation studies, visualization experiments, and hyperparameter analyses are complete.

Critical weaknesses

1.The rationale for requiring at least two views could be clearer. The verifier requires at least two shared observed views, but the main text provides limited explanation of why single-view evidence is excluded. 2.The treatment of single-view clients needs clarification. Single-view clients cannot provide cross-view predictive supervision or participate in relation verification. Their contribution across the two training stages could be explained more explicitly.

Questions for authors

1.Could the authors explain the rationale for requiring at least two shared views for relation verification, and the trade-off this introduces between reliability and coverage? 2. Could the authors clarify which model components single-view clients update during each training stage, and how these clients benefit from federated collaboration?

6 comments
AuthorsOpenReview ↗

Response to Reviewer VU5N W1/Q1: Relation Verification

Thank you for your careful reading and questions about relation verification and single-view clients. We report relation-precision and coverage diagnostics and explain single-view updates and parameter-sharing paths in both training stages. These analyses clarify how FIMCPA uses cross-view evidence to screen relations while including clients with limited observations in collaborative learning.

W1/Q1: Two-view verification and the reliability–coverage trade-off. The two-view requirement determines which sample pairs can supply cross-view evidence for negative removal. We examine its rationale and compare the relation precision and coverage of one-view and two-view rules using the same learned representations.

Neighbors in one view can reflect view-specific factors. We require mutual-neighbor agreement in every shared observed view, with at least t=2 such views. This is the least restrictive count threshold permitting cross-view corroboration; a larger threshold requires more observations.

Verification removes selected cross-sample negatives from the denominator without adding cross-sample positives. A pair supported by only one shared view remains a negative because it lacks cross-view corroboration. Its samples still contribute reconstruction and same-sample alignment during contrastive training; samples with multiple observed views also supply prediction targets during pretraining. Verification thus controls which relations trigger negative removal, rather than which samples can learn: stricter relation screening does not reduce training-sample coverage in the same proportion. This allows FIMCPA to select more reliable relations while continuing to learn from observations outside the verified pairs.

Using representations learned with t=2, we compare t=1 and t=2 offline, keeping representations, masks, samples, neighborhood sizes and diagnostic minibatches fixed. Labels are used only for this offline evaluation. Precision P is the proportion of verified pairs from the same class; coverage C is the proportion of verified pairs among all unordered pairs of distinct samples, pooled across minibatches within clients. Default Full:Partial:Single=2:1:1, r=0.5; entries are percentages reported as mean ± sample SD over five random seeds.

DatasetP (t=1)P (t=2)C (t=1)C (t=2)
UCI75.56 ± 1.2292.61 ± 0.806.764 ± 0.0731.539 ± 0.063
HW84.09 ± 1.8195.68 ± 0.546.592 ± 0.0632.450 ± 0.046
Fashion83.94 ± 0.6496.48 ± 0.516.967 ± 0.0461.999 ± 0.046
Caltech-5V71.74 ± 1.6583.48 ± 1.049.989 ± 0.0883.771 ± 0.167
OutdoorScene59.77 ± 2.0872.62 ± 1.558.282 ± 0.1022.664 ± 0.068
CIFAR10-MV51.36 ± 0.6955.44 ± 0.697.351 ± 0.0142.126 ± 0.008
CCV16.94 ± 0.4928.08 ± 1.452.409 ± 0.0240.175 ± 0.009
MNIST-3V94.55 ± 0.2797.06 ± 0.247.505 ± 0.0102.363 ± 0.011

On UCI, retained relations achieve a precision of 92.61±0.80%, compared with 70.56±1.57% for excluded one-shared-view relations. The fraction of different-class pairs incorrectly removed falls from 1.829±0.085% to 0.126±0.017%, while same-class recall falls from 53.03±1.21% to 14.79±0.54%.

Requiring two shared views retains relations with a higher same-class proportion than the excluded subset. Fewer incorrect removals preserve different-class negatives, at the cost of leaving more same-class pairs as negatives.

The offline comparison shows higher precision and lower coverage for all 120 trained models: eight datasets × five seeds × three settings (default, partial-only, equal partial/single). The latter two settings contain no full-view clients. The precision improvement also holds within matched full- and partial-view client groups. Verification remains imperfect (CCV: 28.08% precision), but this consistent trade-off at fixed representations supports two shared views as the minimum for cross-view corroboration: it selects more reliable relations for negative removal while leaving other samples available for reconstruction and same-sample alignment.

AuthorsOpenReview ↗

Response to Reviewer VU5N W2/Q2: Single-View Clients

W2/Q2: Single-view updates and federated transfer.

Single-view clients cannot construct local cross-view prediction targets or verify relations across views. Their training combines observed-view reconstruction, same-sample alignment and federated parameter sharing, with different model components involved in the two stages.

During Masked Cross-view Predictive Representation (MCPR) pretraining, a single-view client has no cross-view prediction target. Observed-view reconstruction updates its encoder and decoder; the fusion network receives no local pretraining gradient.

During relation-verified contrastive alignment, its common and observed-view representations form a same-sample positive pair. Sample identity supplies a valid correspondence even without a second view. Alignment updates the observed-view encoder and fusion network, while reconstruction trains the encoder and decoder. Cross-sample verification is unavailable, so cross-sample negatives remain in the denominator: Lprec=LRc,Lc=LRc+αLCc.\mathcal L_{\mathrm{pre}}^c=\mathcal L_{\mathrm R}^c,\qquad\mathcal L^c=\mathcal L_{\mathrm R}^c+\alpha\mathcal L_{\mathrm C}^c.

With support fraction ρi=1/V\rho_i=1/V, fusion contributes to the common representation. Gradients reach observed-view input columns and shared fusion parameters. Unobserved-view encoders, decoders and input columns receive no local data gradient or view-specific aggregation weight.

Observation-weighted aggregation weights view-specific parameter blocks by the number of observations for that view and shared fusion parameters by the total number of observed sample–view pairs. During pretraining, unchanged single-view fusion parameters still enter aggregation because weights depend on observations, not nonzero gradients. Each client uses the received encoder and fusion parameters to compute its common representations. Parameter sharing thus allows predictive supervision from multi-view clients to influence these representations.

Single-view clients thus contribute observed-view reconstruction and same-sample alignment updates, and receive encoder and fusion parameters jointly learned with multi-view clients. They use these parameters to compute common representations and the global centers to obtain cluster assignments. Parameter sharing therefore transfers information learned from multi-view observations to single-view clients without requiring a local cross-view prediction loss.

We report relation-precision and coverage diagnostics and clarify the updates and parameter-sharing paths of single-view clients in both training stages. These additions explain the evidence behind the two-view rule and how clients with limited observations participate in federated learning. Thank you again for your careful reading and helpful questions. We welcome further feedback and are happy to clarify any remaining details during the discussion.

Did the author consider how to solve the situation where some clients have only one view while others have multiple views?

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

The paper proposes FIMCPA for federated incomplete multi-view clustering under both client-level view heterogeneity and sample-level missingness. The method first learns common representations through masked cross-view prediction, using held-out observed views as targets. It then applies sample-balanced contrastive alignment with relation verification to screen potential false negatives based on agreement across shared views. Observation-weighted aggregation accounts for the data supporting different model updates, and a federated clustering procedure produces shared global cluster centers. The evaluation covers eight benchmark configurations, with additional ablations, missingness analyses, and client-composition experiments.

Critical strengths

1-Masked cross-view prediction uses only actually observed targets and removes each target from both paths used to construct its input representation. This provides a well-motivated way to exploit partial joint observations without treating genuinely missing views as supervision. Relation verification also uses the evidence available to each sample pair: it requires mutual neighborhood agreement across shared views and removes potential false negatives without introducing additional cross-sample positives. Together with sample-balanced alignment and observation-weighted aggregation, these choices form a coherent response to the joint heterogeneity and missingness setting.

2-FIMCPA achieves the highest reported mean ACC, NMI, and ARI across the eight benchmark configurations. The methods share partitions, observation masks, and five paired seeds, with standard deviations provided in the appendix. The four-configuration ablation study examines prediction, target masking, alignment, and verification, including controls that retain reconstruction pretraining and its schedule. The missingness experiments further show consistently strong mean performance across the tested incomplete-sample rates.

These complementary experiments support the usefulness of the proposed components. Although some margins over the strongest baseline are small, the overall empirical coverage makes a convincing case for the framework’s practical value.

Critical weaknesses

1-The default full/partial/single-view client ratio is 2:1:1, and the alternative ratios in Table 3 are 1:2:1 and 1:1:2. All of these settings retain full-view clients. The missingness sweep changes sample observations while keeping the client acquisition configuration fixed, so it does not establish performance when no client acquires the complete set of views.

Section 3 appropriately acknowledges that recovering common semantics requires informative cross-view associations. A useful additional evaluation would therefore remove full-view clients while retaining connected, partially overlapping view supports. This would test an important part of the stated setting and clarify how much the current results depend on the cross-view evidence supplied by full-view clients.

2-FIMCPA jointly refines global cluster centers through federated Lloyd iterations. In contrast, Appendix B.3 states that FMCSC and AFMCC perform client-wise K-means after shared initialization. Since predictions are concatenated across clients and evaluated using a single global label matching, these different procedures can affect both partition quality and consistency across clients.

The current results therefore compare complete pipelines with different clustering stages. The within-method ablations provide useful evidence for FIMCPA’s representation-learning components, but they do not determine how much of its advantage over other methods comes from the learned representations and how much comes from global clustering coordination. Comparisons under a matched clustering procedure, where the representations permit it, would make this distinction clearer.

Questions for authors

1-Can the authors evaluate a configuration with zero full-view clients while preserving connected cross-view overlap, such as overlapping two-view subsets in a three-view setting? Reporting the amount of co-observed predictive supervision and the fraction of sample pairs eligible for relation verification would help explain performance as overlap becomes sparse.

2-Can compatible learned representations be compared using the same clustering procedure, initialization strategy, and refinement budget? An additional comparison of client-wise versus globally refined clustering on FIMCPA’s fixed representations would help isolate the contribution of the final clustering stage.

6 comments
AuthorsOpenReview ↗

Response to Reviewer PBoh W1/Q1: No Full-View Clients

Thank you for your questions about learning from partial observations and comparability of the clustering procedures. We report experiments and supervision-coverage measurements without full-view clients, matched clustering comparisons, and a fixed-representation control. These evaluations examine how partial overlap supports learning and distinguish the roles of learned representations and final refinement.

W1/Q1: Learning without full-view clients. To examine the dependence on observations supplied by full-view clients, we evaluate two configurations with no full-view clients and a connected view co-observation graph: partial-view clients only, and an equal mixture of partial- and single-view clients.

We evaluate P and P+S on all eight datasets using five paired random seeds for each setting, yielding 80 additional experiments. Ref is the default mixture (full:partial:single=2:1:1); P uses only partial-view clients (0:1:0); P+S mixes partial- and single-view clients equally (0:1:1). Partial-view clients acquire 2 to V−1 views. We use class-stratified partitions, matched training budgets and an incomplete-sample rate of r=0.5 on multi-view clients. In every run, the view co-observation graph is connected: views are nodes, and an edge means that at least one sample observes both views on a client. All views are linked by such overlap paths, even though no client observes them all. ACC, NMI and ARI (%) are listed in rows for each dataset, reported as mean ± sample SD over five runs with different random seeds.

DatasetMetricRefPP+S
UCIACC85.92 ± 2.2184.70 ± 1.1279.74 ± 2.17
NMI76.58 ± 1.9874.76 ± 1.3168.21 ± 2.10
ARI72.96 ± 3.3270.77 ± 1.7563.08 ± 3.28
HWACC90.20 ± 2.1290.21 ± 0.8085.62 ± 2.74
NMI81.72 ± 2.7181.41 ± 1.1975.38 ± 1.96
ARI80.19 ± 3.5279.93 ± 1.4572.60 ± 2.92
FashionACC89.81 ± 0.4983.63 ± 1.7778.41 ± 1.20
NMI82.33 ± 0.6576.35 ± 0.7869.66 ± 0.69
ARI79.78 ± 0.8971.29 ± 1.4462.36 ± 1.18
Caltech-5VACC79.51 ± 3.2179.06 ± 4.6274.84 ± 2.21
NMI66.94 ± 2.8068.76 ± 1.8058.84 ± 2.42
ARI63.65 ± 3.6864.54 ± 2.6755.51 ± 2.99
OutdoorSceneACC73.90 ± 1.7970.73 ± 0.8765.77 ± 1.20
NMI57.42 ± 1.7854.26 ± 0.9747.58 ± 1.41
ARI50.85 ± 2.4546.75 ± 1.2639.22 ± 1.80
CIFAR10-MVACC66.70 ± 2.6266.57 ± 2.6165.96 ± 1.24
NMI54.61 ± 0.9053.67 ± 1.4652.50 ± 0.66
ARI45.01 ± 1.4044.37 ± 2.5343.42 ± 1.04
CCVACC24.45 ± 0.7522.58 ± 0.9919.95 ± 0.76
NMI18.78 ± 0.1817.34 ± 0.6514.03 ± 0.43
ARI8.77 ± 0.308.07 ± 0.316.12 ± 0.33
MNIST-3VACC99.00 ± 0.0698.57 ± 0.0298.23 ± 0.06
NMI96.93 ± 0.1495.80 ± 0.0594.97 ± 0.12
ARI97.78 ± 0.1396.84 ± 0.0596.11 ± 0.12

Without any full-view clients, P retains 97.32% of Ref ACC on average, showing that partial-view clients alone support effective clustering under the tested connected view-overlap patterns. P+S additionally includes single-view clients. Its results show that FIMCPA can still learn collaboratively from partial observations on the remaining clients when some clients cannot supply local cross-view prediction targets.

This capability stems from how MCPR exploits local co-observations. Each prediction task uses an actually observed view as its target and other observed views of the same sample as input; neither the sample nor its client needs all views. Removing full-view clients therefore does not remove cross-view predictive supervision. For example, clients with view subsets {1,2} and {2,3} construct their own local prediction tasks. Aggregation of updates to the shared view-2 encoder and fusion parameters connects these local learning tasks, allowing cross-view learning to exploit complementary observations distributed across clients.

Single-view clients do not construct cross-view prediction targets, but update the model through reconstruction and same-sample alignment, and receive and use parameters jointly learned by multi-view clients. Clients in P+S thus contribute to different objectives according to their available observations and collaborate through parameter sharing. Together, the P and P+S results show that full-view clients are not a prerequisite under the tested connected view-overlap patterns: distributed local co-observations still provide effective learning signals for FIMCPA.

See our next response, “Supervision and Coverage,” for predictive-supervision quantities and relation-verification coverage.

AuthorsOpenReview ↗

Response to Reviewer PBoh W1/Q1: Supervision and Coverage

W1/Q1: Remaining supervision without full-view clients. To quantify the evidence available under partial overlap, we report observed prediction targets and distinguish pairs eligible for relation verification from those that pass it. These measurements accompany the clustering results without full-view clients.

Ref denotes the default 2:1:1 full/partial/single mixture; P contains partial-view clients only; P+S mixes partial- and single-view clients equally. We retain class-stratified partitions, matched training budgets and an incomplete-sample rate of r=0.5 on multi-view clients.

Eligible and Verified use the same denominator: all unordered pairs of distinct samples pooled across client-local training minibatches, including single-view clients. Eligible pairs share at least two observed views; Verified pairs additionally satisfy mutual-neighbor agreement in every shared view. Co-observation retention and pair coverage are expressed as percentages. Entries are ordered as P/P+S, except targets/pass, which follows Ref/P/P+S. All values are means over five random seeds.

DatasetCo-obs.Targets/passEligibleVerified
UCI50.14 / 25.072488.8 / 2000 / 100024.90 / 12.451.526 / 0.717
HW60.70 / 30.285145 / 4755.6 / 2359.653.81 / 26.353.009 / 1.617
Fashion50.05 / 25.0312503.4 / 10016 / 5008.425.01 / 12.521.915 / 0.909
Caltech-5V52.09 / 23.673048 / 2577.2 / 1196.845.34 / 20.814.695 / 2.028
OutdoorScene53.20 / 26.614812.4 / 4030.4 / 201638.46 / 19.223.379 / 1.721
CIFAR10-MV50.02 / 25.0162454.8 / 50000 / 2500024.98 / 12.491.692 / 0.828
CCV49.96 / 24.988490.8 / 6784 / 339225.01 / 12.510.285 / 0.153
MNIST-3V50.00 / 25.0075010 / 60000 / 3000024.98 / 12.491.815 / 0.907

Targets/pass denotes the number of observed prediction targets per data pass. Co-observation retention is the mean paired-seed ratio to Ref of co-observed unordered view-pair counts, summed over samples. Ref training coverage is unavailable in the logs.

Predictive supervision under partial overlap. A sample with s observed views contributes s(s−1)/2 view pairs but s predictive targets when s≥2. The number of co-observed view pairs grows quadratically with the number of observed views, whereas target count grows linearly while another input view is available. View-subset reduction can therefore remove view pairs faster than prediction targets. On UCI, P retains 50.14% of co-observations but 80.36% of targets (2000/2488.8). MCPR exploits these remaining observed targets without requiring all view pairs, allowing partial-view clients to continue supplying cross-view supervision.

Eligible and verified coverage measure different conditions. The former records whether observations permit a cross-view check; the latter also requires agreement in the learned neighborhoods. Verification filters negatives, whereas MCPR learns from observed targets. Low verified coverage therefore does not mean that the same fraction of samples participates in training. Reconstruction and same-sample alignment remain active, and samples with sufficient co-observation can still supply predictive targets even when they form no verified cross-sample pair.

In P+S, half the clients are single-view clients, further reducing local opportunities for prediction and verification. On CIFAR10-MV, predictive targets per pass fall from 50,000 in P to 25,000 in P+S, and eligible-pair coverage falls from 24.98% to 12.49%, while ACC changes from 66.57 to 65.96, a decrease of 0.61 points. On MNIST-3V, predictive targets also halve, from 60,000 to 30,000, while ACC decreases by 0.34 points to 98.23. These examples show that substantial reductions in cross-view supervision can coexist with small accuracy changes.

These measurements distinguish prediction supervision from relation coverage: partial-view clients retain prediction targets even when few sample pairs pass verification, while single-view clients contribute reconstruction and same-sample alignment updates. Aggregation connects these local contributions. The results therefore support learning without full-view clients through the remaining observations; low verification coverage limits which relations alter the negative set, rather than excluding other samples from training.

AuthorsOpenReview ↗

Response to Reviewer PBoh W2/Q2: Common Global Clustering

The original pipelines differ in their final clustering stages. We therefore compare the learned representations under a common global initialization rule and refinement budget, allowing each method to use cluster statistics aggregated across clients.

W2/Q2: Common clustering across methods. We retain training, features, default data partitions and masks. Local K-means provides centers and cluster sizes; count-weighted server-side K-means initializes global centers in each method's space. K-means++ uses 10 local/100 server restarts. All methods use at most 20 federated Lloyd updates, a convergence tolerance of 10−510^{-5}, the same distance metric, tie-breaking rule and empty-cluster handling, and a single global Hungarian matching. Clients assign samples to global centers; the server updates them using aggregated within-cluster feature sums and sample counts.

FIMCPA uses HH; FMCSC/AFMCC supply learned representations, replacing local K-means with the global procedure. EFDMVC/FreeCSL/HSACC retain their K-means inputs, including observed-view consensus ZZ for FreeCSL. DSIMVC uses probabilities averaged over observed views, and H2-FedMVC uses final MoE logits. The latter two are supplementary diagnostics of prediction-head outputs; H2-FedMVC retains label-assisted training. ACC (%) is reported as mean ± sample SD over five random seeds.

DatasetFIMCPAFMCSCAFMCCEFDMVCFreeCSLHSACCDSIMVCH2-FedMVC
UCI85.92 ± 2.2173.07 ± 1.8277.58 ± 0.9276.09 ± 2.6075.89 ± 3.9668.51 ± 5.4153.50 ± 3.3677.86 ± 4.62
HW90.20 ± 2.1277.15 ± 1.9278.34 ± 0.7880.98 ± 4.1482.77 ± 1.8763.59 ± 5.7850.40 ± 3.7680.32 ± 2.54
Fashion89.81 ± 0.4984.80 ± 0.6589.06 ± 0.1784.26 ± 2.0486.95 ± 4.6881.55 ± 5.3764.57 ± 1.1285.97 ± 1.56
Caltech-5V79.51 ± 3.2171.99 ± 1.7974.00 ± 4.3065.64 ± 1.5878.83 ± 3.7633.09 ± 11.4641.59 ± 3.4272.09 ± 6.15
OutdoorScene73.90 ± 1.7962.40 ± 3.3468.90 ± 2.1861.38 ± 2.0953.56 ± 4.0145.42 ± 4.2847.00 ± 2.9057.48 ± 3.01
CIFAR10-MV66.70 ± 2.6253.03 ± 2.4555.02 ± 2.8751.55 ± 3.1162.05 ± 4.4634.85 ± 4.2650.30 ± 1.0041.18 ± 3.93
CCV24.45 ± 0.7519.50 ± 1.0321.57 ± 0.4720.19 ± 0.5821.39 ± 0.9014.95 ± 0.4515.83 ± 0.2115.92 ± 0.77
MNIST-3V99.00 ± 0.0694.77 ± 1.0098.48 ± 0.0298.27 ± 0.0798.53 ± 0.0972.06 ± 2.6596.43 ± 3.9395.17 ± 2.77

MCPR removes the target view from both input paths and predicts it from other observed views, preventing direct copying of the target. Subsequent alignment connects same-sample view-specific and common representations, while relation verification uses shared-view neighborhood agreement to remove potential false negatives. These training constraints shape the representation space before final clustering; Lloyd refinement only updates centers and assignments on fixed representations.

After matching global initialization and Lloyd refinement, FIMCPA exceeds FMCSC and AFMCC by 9.10 and 5.82 percentage points in ACC on average and has the highest mean ACC on all eight datasets. Thus, its advantage persists when final global coordination is also available to every baseline. The complementary fixed-representation control holds FIMCPA representations and initial centers fixed to isolate subsequent refinement. Matching the procedure across methods and isolating refinement within FIMCPA provide two distinct controls that jointly support the quality of the learned common representations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.