acceptodds
Under review as a conference paper at ICLR 2027

Learning Observation Robust Representations for Unbalanced Incomplete Multi-View Clustering

Abstract

Multi-view clustering is commonly studied under the assumption that every sample is observed in all views. In practice, however, some views are missing, and the amount of observed data can vary substantially across views, producing strong views with dense observations and weak views with sparse ones. Existing incomplete multi-view clustering methods mainly recover missing information through completion or alignment, while unbalanced methods adjust the contributions of strong and weak views. These strategies do not directly constrain the consensus representation of the same sample when its observed-view set changes. As a result, the representation may vary across different missing patterns and may still correlate with frequently occurring observation patterns. We propose COIN, a content–observation invariance framework for unbalanced incomplete multi-view clustering. COIN uses an observation factor to encode the missing pattern, while the content factor for clustering is computed from mask-gated view features. Nested content invariance encourages the same sample to maintain a consistent representation under coarser missing patterns, and rebalanced decorrelation reduces linear correlation between the content factor and the missing pattern after rare patterns are upweighted. Experiments on five unbalanced incomplete multi-view clustering benchmarks show that COIN consistently outperforms representative incomplete and unbalanced methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-View Clustering, Incomplete Multi-View Clustering, Unbalanced Multi-View Clustering, Representation Invariance, Observation Bias.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dGQvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COIN, a deep representation-learning method for unbalanced incomplete multi-view clustering, where views are observed at substantially different rates. The method aims to reduce two effects of missingness: changes in a sample’s representation when an observed view is removed, and correlation between the learned representation and observation patterns.

COIN learns a content factor from mask-gated view features and an observation factor from the missingness mask, with reconstruction objectives supporting representation learning. It introduces two regularizers. Nested content invariance aligns the normalized content representations before and after removing one observed view, with sparsely observed views more likely to be removed. Rebalanced decorrelation penalizes linear content–mask cross-covariance using truncated inverse-frequency weights to increase the contribution of rare observation patterns. Final clustering applies K-means to the normalized content factors. The theoretical analysis shows that nested invariance and zero content–mask covariance are distinct properties and clarifies the relationship between these criteria and the implemented losses.

The authors evaluate COIN against twelve baselines on five benchmarks, using one synthetically generated, feature- and label-independent unbalanced mask per dataset at target mean missing rates from 0.1 to 0.5. They report the best ACC, NMI, and ARI in all 75 dataset–rate–metric cells. Supporting experiments include ablations of the two regularizers, hyperparameter sensitivity analyses, and diagnostics measuring representation drift under view removal and residual predictability of observation masks.

Critical strengths
  1. The paper distinguishes two properties: the stability of a sample's content representation when an observed view is removed, and the linear correlation between content and observation patterns under a rebalanced distribution. Proposition 1, while elementary, shows that neither property implies the other, which gives a clear rationale for addressing both. COIN translates this distinction into explicit representation objectives, with sampling and weighting that account for unequal view availability and pattern frequencies. Masking consistency has precedents such as CIMLN, but this joint formulation is still a useful conceptual contribution: it specifies which aspects of observation robustness the method targets and makes those claims measurable.

  2. Beyond clustering accuracy, the appendix reports representation drift, content–mask covariance, and frozen linear mask-probe results across ablation variants at two missing rates. These ablations provide concrete evidence for the targeted stability effect. The authors also report substantial residual mask predictability in the full model and explicitly distinguish decorrelation from independence. These diagnostics connect the method to observable representation behavior and make the scope of the empirical support clear.

Critical weaknesses
  1. Appendix C.5 states that six loss coefficients are selected independently for each dataset-missing-rate combination by maximizing ground-truth ACC, followed by NMI and ARI. Although labels do not enter the training loss, they directly determine which models are reported, including the ablation results in Table 3. Meanwhile, BURG and CIMLN use their official recommended hyperparameters, and an equivalent selection budget is not established for any baseline. The comparison therefore does not clearly separate the benefit of the proposed method from a tuning advantage. Moreover, each dataset–rate combination uses one fixed mask, and averaging over ten kmeans initializations does not quantify variability across mask realizations or representation-training runs. Several baseline results also behave erratically across missing rates (e.g., BURG on MSRCV1 drops from 87.2 to 25.3 ACC, and CIMLN on HandWritten rises from 79.6 to 90.9 between r = 0.1 and 0.2), suggesting run-to-run instability that may inflate the reported margins.

  2. Although the diagnostics confirm reduced representation drift and content–mask covariance, they provide limited evidence that these improvements translate into reduced reliance on observation patterns while preserving complementary view information. At 50% mean missingness, the full model’s mask-probe AUC is 0.963 on Washington, compared with 0.678 for reconstruction alone, and remains 1.000 on Synthetic3d. The probes use raw content, whereas the covariance diagnostic and clustering use normalized content, leaving residual mask predictability in the actual clustering representation unmeasured. Moreover, a near-zero covariance norm on normalized content can coexist with strong linear separability. These observations do not contradict the covariance-only claim in Section 3.4, but leave a gap between the measured criterion and the broader motivation of preventing the representation space from being organized around observation patterns. The invariance diagnostic also leaves an important tradeoff unresolved. Because scarce views are dropped more frequently, low drift could partly reflect reduced use of information unique to those views. The reported reconstruction and clustering results do not directly establish that complementary scarce-view information is preserved. A controlled experiment in which a scarce view carries complementary cluster information, with clustering performance reported separately when that view is available and absent, would help assess this tradeoff.

Questions for authors
  1. Apart from BURG and CIMLN, which use official defaults, how were hyperparameters set for the other ten baselines? Did any baseline receive label-based or per-cell selection comparable to COIN’s?
  2. What are COIN’s results, including the ablations, when one coefficient setting is used per dataset or across all datasets, chosen without ground-truth labels? Examples would be a fixed default or a label-free internal criterion. How large is the gap between these results and the per-cell oracle-selected results in Tables 2 and 3?
  3. Can you report means and standard deviations over multiple independently generated masks and multiple representation-training seeds per dataset and missing rate? Several baselines behave erratically across missing rates
  4. The mask probe uses raw content, while the covariance diagnostic and clustering use normalized content. Can you report probe results on the normalized content actually used for clustering? Can you add a scale-invariant dependence measure, such as canonical correlation between content and mask?
  5. Can you show that low drift is not achieved by down-weighting the views most often dropped?
  6. CIMLN already enforces consistency across observation patterns by masking views, but it masks densely observed views, while COIN preferentially drops scarce ones. How does nested invariance differ conceptually and empirically from CIMLN’s masking module?
  7. All masks are generated independently of features and labels. How does COIN behave when missingness depends on the class or the content?
7 comments
AuthorsOpenReview ↗

W1/Q1: Baseline Hyperparameter Configuration and Selection Fairness

We examined the hyperparameter configurations of all ten baselines, excluding BURG and CIMLN. Each baseline was assessed using a single recorded configuration for each method, adhering to the respective released implementation or its established experimental setup.

The principal configurations were as follows:

BaselineHyperparameter configuration
APADC500 epochs, latent dimension 16, learning rate 10−410^{-4}
COMPLETER500 epochs, latent dimension 128, learning rate 10−410^{-4}
CPSPAN200 pretraining and 50 alignment epochs
DIMVCEmbedding dimension 10, 500 pretraining epochs
RecFormerDimension 512, 50 reconstruction epochs, learning rate 10−310^{-3}
ProImp50 pretraining epochs, latent dimension 256, learning rate 10−310^{-3}
SMILE150 epochs, 50 warm-up epochs, learning rate 10−310^{-3}
SURE80 epochs, latent dimension 10
BGIMVSCView-weight exponent r=7r=7, λ=0.1\lambda=0.1, γ=10\gamma=10, initial μ=10−5\mu=10^{-5}, ρ=1.05\rho=1.05, binary 5-NN graph and maximum 100 iterations
SIMVC-SAλ=1\lambda=1, μ=1\mu=1, and 2K2K anchors

The recorded baseline evaluations used these configurations without performing a ground-truth-guided search over the dataset–missing-rate cells.

In contrast to the historical COIN evaluation, which selected configurations independently for each cell using the ACC, followed by the NMI and ARI, we conducted a different evaluation. This evaluation employed a single fixed coefficient configuration across all 25 cells and included repeated comparisons with BURG and CIMLN under shared-observation masks.

AuthorsOpenReview ↗

W1 / Q2. Fixed-Configuration COIN and Ablation Results

We conducted a rerun of REC, REC+INV, REC+IND, and Full COIN, employing a consistent global code-default tuple across all five datasets and five missing rates. This evaluation utilized three training seeds per cell, without implementing per-cell label-based selection:

α=1,β=0.5,λview=λrender=1,λcov=0.01,λobs=0.1,\alpha=1,\quad\beta=0.5,\quad\lambda_{\rm view}=\lambda_{\rm render}=1,\quad\lambda_{\rm cov}=0.01,\quad\lambda_{\rm obs}=0.1, γ=0.5,γs=1,wmax⁡=8.\gamma=0.5,\quad\gamma_s=1,\quad w_{\max}=8.

The macro-averages of the 25 dataset-rate cell means are:

Fixed variantACCNMIARIDobsD_{\rm obs}ACC/NMI/ARI gain over REC
REC0.55900.40960.36680.3542--
REC+INV0.62940.49920.46860.0036+0.0703 / +0.0896 / +0.1018
REC+IND0.62690.49230.45790.2820+0.0678 / +0.0827 / +0.0911
Full COIN0.64230.50810.47860.0046+0.0832 / +0.0985 / +0.1118

Fixed-track ablation adheres to the intended module logic. The INV variant significantly reduces nested observation drift, whereas IND enhances clustering performance beyond mere reconstruction. Full COIN achieves the highest average values for ACC, NMI, and ARI among the four variants, while maintaining a near-zero DobsD_{\rm obs}. Consequently, the joint objective integrates clustering improvements with observation robustness under a single, globally fixed coefficient setting.

Gap to the oracle-selected tables. In Table 2 Full COIN, the fixed configuration exhibits a lower accuracy (ACC) in 24 out of 25 cells, with a mean fixed-minus-oracle difference of −0.0577-0.0577 and a median difference of −0.0212-0.0212. This indicates a nonuniform gap across datasets and rates. Regarding the Table 3 ablations, the archived like-for-like comparison is available at r∈{0.3,0.5}r\in\{0.3,0.5\}. Across 40 method-dataset-rate cells, the mean fixed-minus-oracle differences are −0.0076-0.0076, −0.0067-0.0067, and −0.0094-0.0094 for ACC, NMI, and ARI, respectively. The fixed results are lower in 17, 19, and 18 out of 40 cells, respectively. Consequently, per-cell oracle selection significantly enhances the historical Full-COIN headline ACC, whereas the average Table 3 ablation gap is considerably smaller and not consistently negative.

AuthorsOpenReview ↗

W1 / Q3. Multi-Mask and Training-Seed Stability

We combined three independently generated mask seeds with three representation-training seeds at r∈{0.3,0.5}r\in\{0.3,0.5\}. Each model execution averages ten K-means initializations internally; these initializations are not considered independent representation-training runs. Consequently, the nine values per dataset-rate cell comprise three mask realizations and three training seeds per mask, rather than nine independent missingness mechanisms.

The corrected fixed external comparison is as follows. Each entry is the mean$\pm$standard deviation over the same 33 mask seeds ×\times 33 training seeds, in the order of ACC/NMI/ARI.

Dataset, rrFixed COINBURGCIMLN
HandWritten, 0.3.8438$\pm$.0715 / .7834$\pm$.0460 / .7373$\pm$.0795.4764$\pm$.0753 / .4663$\pm$.0560 / .2947$\pm$.0659.7972$\pm$.0337 / .7856$\pm$.0185 / .7175$\pm$.0231
HandWritten, 0.5.8099$\pm$.0696 / .7387$\pm$.0305 / .6847$\pm$.0610.2593$\pm$.0208 / .2150$\pm$.0253 / .0823$\pm$.0206.3586$\pm$.1577 / .3138$\pm$.1908 / .2301$\pm$.1742
MSRCV1, 0.3.8454$\pm$.0201 / .7393$\pm$.0296 / .6765$\pm$.0355.5789$\pm$.0623 / .4659$\pm$.0606 / .3411$\pm$.0686.2246$\pm$.0243 / .0569$\pm$.0215 / .0039$\pm$.0132
MSRCV1, 0.5.7385$\pm$.0329 / .6046$\pm$.0401 / .5061$\pm$.0621.2704$\pm$.0276 / .1219$\pm$.0336 / .0323$\pm$.0152.2190$\pm$.0073 / .0491$\pm$.0113 / .0003$\pm$.0047
Synthetic3d, 0.3.9390$\pm$.0071 / .7780$\pm$.0189 / .8264$\pm$.0190.7271$\pm$.0842 / .4369$\pm$.0988 / .4208$\pm$.1169.7415$\pm$.0763 / .4613$\pm$.0726 / .4318$\pm$.0903
Synthetic3d, 0.5.8945$\pm$.0262 / .6783$\pm$.0616 / .7153$\pm$.0649.5126$\pm$.0559 / .1419$\pm$.0969 / .1365$\pm$.0882.4954$\pm$.0906 / .1700$\pm$.1147 / .1200$\pm$.0835
TinyImageNet, 0.3.1408$\pm$.0051 / .0939$\pm$.0023 / .0302$\pm$.0021.1019$\pm$.0017 / .0585$\pm$.0026 / .0137$\pm$.0008.1222$\pm$.0057 / .0866$\pm$.0034 / .0244$\pm$.0019
TinyImageNet, 0.5.1288$\pm$.0078 / .0852$\pm$.0046 / .0254$\pm$.0029.0923$\pm$.0018 / .0438$\pm$.0041 / .0089$\pm$.0010.1048$\pm$.0034 / .0568$\pm$.0023 / .0142$\pm$.0008
Washington, 0.3.6143$\pm$.0212 / .2448$\pm$.0162 / .2966$\pm$.0212.5771$\pm$.0211 / .2323$\pm$.0156 / .2674$\pm$.0186.4975$\pm$.0787 / .2477$\pm$.0429 / .2501$\pm$.0753
Washington, 0.5.5494$\pm$.0386 / .1799$\pm$.0375 / .2136$\pm$.0414.4941$\pm$.0510 / .1432$\pm$.0282 / .1378$\pm$.0443.3443$\pm$.0676 / .1005$\pm$.0603 / .0793$\pm$.0741

The standard deviations presented here encompass both between-mask and within-mask training variations. The fixed COIN model surpassed the BURG model across all ten cell means for the ACC, NMI, and ARI metrics. Additionally, it outperformed the CIMLN model in ten instances for the ACC, eight instances for the NMI, and ten instances for the ARI. Notably, the CIMLN model demonstrated superior NMI performance on the Handwritten and Washington datasets at r=0.3r=0.3.

We examined the two anomalies highlighted by the reviewer using one frozen mask and three training seeds. For BURG on MSRCV1, the accuracy is 0.8467±0.05210.8467\pm0.0521 at r=0.1r=0.1 and 0.2506±0.00510.2506\pm0.0051 at r=0.5r=0.5. The significant rate-dependent decline remains evident despite minor seed variations at r=0.5r=0.5, aligning with its reliance on complete or co-observed samples under conditions of severe missingness. For CIMLN on HandWritten, the accuracy is 0.7990±0.02100.7990\pm0.0210 at r=0.1r=0.1 and 0.9018±0.01880.9018\pm0.0188 at r=0.2r=0.2. This increase is consistent across training seeds and is not attributable to a single-seed anomaly. In the multi-mask experiment, CIMLN exhibits considerable variability in combined mask and training effects on certain cells, including HandWritten at r=0.5r=0.5.

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

This paper proposes COIN for unbalanced incomplete multi-view clustering. The method constructs a content representation from mask-gated view features and an observation representation from the missingness mask. Nested content invariance encourages a sample’s representation to remain stable when one observed view is removed, with scarce views sampled more frequently for removal. A rebalanced decorrelation objective penalizes covariance between content and observation masks after accounting for pattern frequencies. The model combines these constraints with reconstruction objectives and applies K-means to normalized content representations. Experiments on five benchmarks include clustering comparisons, ablations, sensitivity analyses, and diagnostics of representation drift and observation dependence.

Critical strengths

1-Removing an observed view can change a sample’s fused representation even though its underlying identity remains unchanged. The proposed nested-observation construction expresses this desired stability directly and is naturally applicable to incomplete multi-view data. Separating this instance-level objective from content–mask decorrelation also gives the framework a clear motivation.

2-The evaluation covers five datasets and five missing rates, and the reported clustering results are encouraging. More specifically, Appendix E measures representation changes after removing one observed view. Adding the invariance term reduces the normalized representation drift substantially across the evaluated datasets; for example, on Synthetic3d at (r=0.5), the drift decreases from 0.5164 under reconstruction alone to 0.0026 with the invariance term. This is direct evidence supporting that component. Reporting the mask-probe results, including cases where missingness remains highly predictable, is also valuable for assessing the method’s limitations.

Critical weaknesses

1-Appendix C.5 states that six coefficients—(\alpha), (\beta), (\lambda_{\mathrm{cov}}), (\lambda_{\mathrm{obs}}), (\lambda_{\mathrm{view}}), and (\lambda_{\mathrm{render}})—are selected separately for each dataset–missing-rate setting by maximizing ACC, followed by NMI and ARI. Meanwhile, Appendix C states that CIMLN and BURG use their officially recommended hyperparameters. Although labels do not enter the training loss, they still determine which trained configuration is reported, giving COIN a selection advantage that is not matched by these baselines.

This is my main concern because the reported ranking combines the contribution of the method with the benefit of tuning against the ground-truth clustering labels. It also leaves unclear how the reported performance would be obtained in an application where those labels are unavailable. The sensitivity analyses do not resolve this issue: Appendix D fixes the coefficients outside each sweep at their already selected dataset-specific values. A primary comparison using label-free parameter selection or independently fixed defaults, together with comparable tuning budgets, is needed. The complete search spaces, trial budgets, and selected configurations should also be reported.

2-Tables 9 and 10 show that the full model achieves a mask-prediction ROC-AUC of 1.000 on Synthetic3d at both (r=0.3) and (r=0.5), despite substantially reduced covariance. At (r=0.5), for example, the reported squared covariance norm decreases from 0.0595 under reconstruction alone to (5.50\times10^{-5}) for the full model.

The authors acknowledge this behavior and appropriately avoid claiming statistical independence. Nevertheless, the practical interpretation remains incomplete: what aspect of observation dependence relevant to clustering is reduced when the missingness bits remain so readily recoverable by a linear probe? The training penalty uses reweighted covariance, while Appendix E reports unweighted covariance on normalized content and describes the probe using frozen content factors. Diagnostics with matched normalization and weighting would help clarify the relationship. In particular, evidence that the penalty reduces the influence of mask-related directions on clustering geometry would make the mechanism more convincing.

3-Appendix C.1 uses one generated observation mask for each dataset–missing-rate setting. Seed 42 determines the view-scarcity configuration, and the subsequent quota sampling follows a fixed seed schedule. Averaging over ten K-means initializations assesses clustering-initialization variability, but does not establish robustness to different missing samples or different choices of scarce views.

This limitation is especially relevant because COIN explicitly uses view availability and observation-pattern frequencies for its sampling and reweighting. The five missing rates provide useful coverage of missingness severity, but the results could still depend on the particular configurations generated. Comparisons over multiple independently sampled masks and scarcity configurations, with each mask shared across methods and uncertainty reported across runs, would strengthen the empirical case.

Questions for authors

1-How much of COIN’s advantage remains under a label-free selection protocol and comparable tuning budgets? Please provide the full coefficient-search spaces, number of trials, and selected configurations, and report results using a selection rule or defaults established without access to the evaluated clustering labels.

2-What does the decorrelation penalty change about the clustering representation when missingness remains perfectly predictable by the reported linear probe? Could the covariance and probe analyses use matched normalization and weighting, and could the authors examine the influence of mask-related directions on clustering distances or assignments?

3-Do the conclusions persist across independently generated observation masks and different scarce-view configurations? Repeating the comparisons with each sampled mask shared across methods, and reporting variation across masks and training runs, would help establish the stability of the gains.

4 comments
AuthorsOpenReview ↗

W1 / Q1. Fixed-Configuration Competitiveness and Search Summary

We used one unchanged tuple in every dataset-rate cell: α=1\alpha=1, β=0.5\beta=0.5, λcov=0.01\lambda_{\rm cov}=0.01, λobs=0.1\lambda_{\rm obs}=0.1, λview=λrender=1\lambda_{\rm view}=\lambda_{\rm render}=1, γ=0.5\gamma=0.5, γs=1\gamma_s=1, and wmax⁡=8w_{\max}=8. It was not selected separately for any cell in these new runs. This is a fixed-configuration evaluation without per-cell label-based selection; it does not establish independence from all prior benchmark development.

The corrected COIN, BURG, and CIMLN runs use the same missing mask for the same canonical sample; preprocessing and optimization remain method-specific. This is a fixed shared-mask comparison, not a fully budget-matched comparison.

DatasetrMethodACCNMIARI
HandWritten0.3BURG0.4764 +/- 0.07530.4663 +/- 0.05600.2947 +/- 0.0659
HandWritten0.3CIMLN0.7972 +/- 0.03370.7856 +/- 0.01850.7175 +/- 0.0231
HandWritten0.3Fixed COIN0.8438 +/- 0.07150.7834 +/- 0.04600.7373 +/- 0.0795
HandWritten0.5BURG0.2593 +/- 0.02080.2150 +/- 0.02530.0823 +/- 0.0206
HandWritten0.5CIMLN0.3586 +/- 0.15770.3138 +/- 0.19080.2301 +/- 0.1742
HandWritten0.5Fixed COIN0.8099 +/- 0.06960.7387 +/- 0.03050.6847 +/- 0.0610
MSRCV10.3BURG0.5789 +/- 0.06230.4659 +/- 0.06060.3411 +/- 0.0686
MSRCV10.3CIMLN0.2246 +/- 0.02430.0569 +/- 0.02150.0039 +/- 0.0132
MSRCV10.3Fixed COIN0.8454 +/- 0.02010.7393 +/- 0.02960.6765 +/- 0.0355
MSRCV10.5BURG0.2704 +/- 0.02760.1219 +/- 0.03360.0323 +/- 0.0152
MSRCV10.5CIMLN0.2190 +/- 0.00730.0491 +/- 0.01130.0003 +/- 0.0047
MSRCV10.5Fixed COIN0.7385 +/- 0.03290.6046 +/- 0.04010.5061 +/- 0.0621
Synthetic3d0.3BURG0.7271 +/- 0.08420.4369 +/- 0.09880.4208 +/- 0.1169
Synthetic3d0.3CIMLN0.7415 +/- 0.07630.4613 +/- 0.07260.4318 +/- 0.0903
Synthetic3d0.3Fixed COIN0.9390 +/- 0.00710.7780 +/- 0.01890.8264 +/- 0.0190
Synthetic3d0.5BURG0.5126 +/- 0.05590.1419 +/- 0.09690.1365 +/- 0.0882
Synthetic3d0.5CIMLN0.4954 +/- 0.09060.1700 +/- 0.11470.1200 +/- 0.0835
Synthetic3d0.5Fixed COIN0.8945 +/- 0.02620.6783 +/- 0.06160.7153 +/- 0.0649
TinyImageNet0.3BURG0.1019 +/- 0.00170.0585 +/- 0.00260.0137 +/- 0.0008
TinyImageNet0.3CIMLN0.1222 +/- 0.00570.0866 +/- 0.00340.0244 +/- 0.0019
TinyImageNet0.3Fixed COIN0.1408 +/- 0.00510.0939 +/- 0.00230.0302 +/- 0.0021
TinyImageNet0.5BURG0.0923 +/- 0.00180.0438 +/- 0.00410.0089 +/- 0.0010
TinyImageNet0.5CIMLN0.1048 +/- 0.00340.0568 +/- 0.00230.0142 +/- 0.0008
TinyImageNet0.5Fixed COIN0.1288 +/- 0.00780.0852 +/- 0.00460.0254 +/- 0.0029
Washington0.3BURG0.5771 +/- 0.02110.2323 +/- 0.01560.2674 +/- 0.0186
Washington0.3CIMLN0.4975 +/- 0.07870.2477 +/- 0.04290.2501 +/- 0.0753
Washington0.3Fixed COIN0.6143 +/- 0.02120.2448 +/- 0.01620.2966 +/- 0.0212
Washington0.5BURG0.4941 +/- 0.05100.1432 +/- 0.02820.1378 +/- 0.0443
Washington0.5CIMLN0.3443 +/- 0.06760.1005 +/- 0.06030.0793 +/- 0.0741
Washington0.5Fixed COIN0.5494 +/- 0.03860.1799 +/- 0.03750.2136 +/- 0.0414

The corrected win counts are COIN vs BURG: 10/10 ACC, 10/10 NMI, and 10/10 ARI; COIN vs CIMLN: 10/10 ACC, 8/10 NMI, and 10/10 ARI. CIMLN retains higher NMI on HandWritten at r=0.3r=0.3 and Washington at r=0.3r=0.3.

AuthorsOpenReview ↗

W2 / Q2. Geometry Validation

We investigate how the decorrelation objective affects clustering geometry by examining the influence of mask-predictive directions on content representations.

Four variants (REC, REC+INV, REC+IND, and Full COIN) are evaluated using a shared 80/20 diagnostic split of frozen embeddings. Per-view logistic probes identify mask-predictive directions on the training partition. Their coefficient subspace is extracted through SVD and removed from held-out representations, followed by L2 normalization. Cluster assignments are evaluated using fixed K-means centers fitted on the original training embeddings.

We examine fixed-center assignment changes, Cluster–Mask NMI, and mask separability to characterize the relationship between observation patterns and clustering geometry.

Decorrelation reduces assignment sensitivity to mask-predictive directions. Compared with REC, REC+IND reduces fixed-center assignment changes in 9/10 dataset-rate settings, with an average reduction of 0.0500 (5.0 percentage points). Cluster–Mask NMI decreases in 9/10 settings (−0.1080), while held-out mask separability decreases in 8/10 (−0.1502).

These findings demonstrate that IND modifies clustering-relevant content geometry. Cluster assignments become less sensitive to the removal of mask-predictive directions, accompanied by weaker statistical association between clustering assignments and observation patterns.

Mechanistic interpretation. IND constrains the association between content representations and observation patterns, reducing the influence of mask-related directions on the learned clustering structure. Nested Invariance complements this objective by promoting content consistency across nested observation subsets.

Together, these mechanisms address observation dependence and representation instability at different levels, explaining their complementary roles in COIN.

AuthorsOpenReview ↗

W3 / Q3. Robustness Across Observation Masks and Scarcity Configurations

We evaluated COIN across three independently generated observation masks and three training seeds per mask, covering five datasets at r=0.3r=0.3 and r=0.5r=0.5. Each mask seed generates a new scarcity vector, changing both the sample-level observation patterns and the relative availability of the individual views.

Variation in scarcity configurations. MSRCV1, Washington, HandWritten, and TinyImageNet exhibit three distinct view-availability rankings at both missing rates, while Synthetic3d exhibits two. For example, at r=0.5r=0.5, Washington produces availability vectors (0.6391,0.3000,0.7739,0.2913)(0.6391,0.3000,0.7739,0.2913), (0.3696,0.6783,0.4304,0.5174)(0.3696,0.6783,0.4304,0.5174), and (0.8000,0.2913,0.5826,0.3217)(0.8000,0.2913,0.5826,0.3217) under seeds 101, 102, and 103, respectively. The scarcest view consequently changes from view 4 to view 1 and then view 2.

Robustness across masks and training seeds. The evaluation comprises 180 model-training runs and 90 paired Full-minus-REC comparisons. Each dataset-rate-mask setting includes three independently trained models, with ten K-means initializations averaged within each run. Across the 90 paired comparisons, Full COIN improves ACC in 67, NMI in 73, and ARI in 71, while reducing DobsD_{\rm obs} in all 90. At the dataset-rate level, ACC/NMI/ARI improve in 6/10, 8/10, and 8/10 settings, respectively, with lower DobsD_{\rm obs} in all 10.

Overall finding. COIN consistently reduces nested representation drift across independently sampled observation masks and varying view-scarcity configurations. The clustering results further demonstrate improvements across most paired runs, with the magnitude of improvement varying across datasets.

AuthorsOpenReview ↗

Historical Hyperparameter Search and Selection Protocol (W1)

The historical COIN evaluation employed a two-stage search. Stage A evaluated nine combinations of α∈{0.5,1,2}\alpha\in\{0.5,1,2\} and β∈{0.2,0.5,1}\beta\in\{0.2,0.5,1\}. Stage B fixed the selected pair and evaluated 16 combinations of λview,λrender∈{0.5,1}\lambda_{\rm view},\lambda_{\rm render}\in\{0.5,1\}, λcov∈{0.01,0.1}\lambda_{\rm cov}\in\{0.01,0.1\}, and λobs∈{0.05,0.1}\lambda_{\rm obs}\in\{0.05,0.1\}. This yielded 25 executions and 24 unique configurations per dataset-rate setting, with configurations selected by ground-truth ACC, followed by NMI and ARI.

The original Table 3 ablations inherited the cell-specific coefficients selected for Full COIN, with α\alpha and/or β\beta disabled to isolate individual objectives. The new fixed-configuration experiments reported in Part 1 provide the corresponding evaluation under one shared parameter setting.

The complete 25-cell historical configuration record is available in the supplementary search table.

Reviewer_7t3YOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes COIN for unbalanced incomplete multi-view clustering. The method learns a shared content representation and introduces two regularization objectives: nested content invariance to maintain representation consistency under view removal, and rebalanced decorrelation to reduce dependence on observation patterns.

Critical strengths
  1. The paper addresses a relevant representation-level problem in incomplete multi-view clustering, with a clearly presented motivation.
  2. The method is straightforward and achieves consistent improvements across five benchmarks, supported by ablation studies and representation-level diagnostics.
Critical weaknesses
  1. The main concern is the assumption that representations should remain consistent when observed views are removed. Different views may contain substantially different discriminative information, and losing an informative view can reasonably change the representation. The paper does not sufficiently justify this assumption or examine whether enforcing invariance may suppress useful view-specific information.
  2. The rationale for preferentially dropping sparsely observed views is unclear. Since these views are already underrepresented, this strategy may further discourage the model from exploiting their unique information. Comparisons with uniform dropout or preferentially dropping frequently observed views are missing.
  3. The experiments provide limited evidence that observation dependence is effectively reduced. The proposed decorrelation only suppresses linear covariance, while Table 9 shows that observation masks remain highly predictable from the learned content representations.
  4. Appendix C.5 states that hyperparameters are selected separately for each dataset and missing rate using ground-truth labels to maximize clustering accuracy. This raises concerns about selection bias and the fairness of comparisons with baseline methods.
  5. The robustness evaluation is limited. The experiments use artificially generated missing patterns independent of features and labels, with only a single mask seed per setting. More realistic missingness mechanisms and differences in view importance are not examined. Moreover, the absolute clustering performance on TinyImageNet remains low.
Questions for authors
  1. How is the invariance assumption justified when views contain different amounts of discriminative information? Can the authors evaluate such settings?
  2. Why are sparsely observed views preferentially dropped? How does this compare with uniform dropout or preferentially dropping frequently observed views?
  3. How do the authors explain the high mask-prediction ROC-AUC values in Table 9 despite the decorrelation objective?
  4. Can the authors report results with label-free hyperparameter selection, comparable tuning budgets for baselines, and multiple missingness seeds?
  5. How does COIN perform under realistic, feature-dependent missingness mechanisms?
4 comments
AuthorsOpenReview ↗

W1_Q1_invariance_with_unequal_view_information.md

Nested Invariance ensures that content shared across observable view subsets remains stable when views are removed. As a soft regularization term used in conjunction with reconstruction, it fosters consistency in representation while enabling the model to utilize discriminative information from available views. This interaction is analyzed through fixed-configuration ablations and a controlled experiment, where complementary information is concentrated in the least common view.

Fixed-configuration ablation: By employing a single coefficient tuple across five datasets and five missing rates (resulting in 25 dataset-rate configurations with three training seeds each), we derived the following macro-averaged outcomes:

VariantACCNMIARIDobsD_{\rm obs}
LREC\mathcal{L}_{\rm REC}0.55900.40960.36680.3542
LREC+LINV\mathcal{L}_{\rm REC}+\mathcal{L}_{\rm INV}0.62940.49920.46860.0036
LREC+LIND\mathcal{L}_{\rm REC}+\mathcal{L}_{\rm IND}0.62690.49230.45790.2820
Full COIN0.64230.50810.47860.0046

Incorporating INV into the reconstruction process enhances the average ACC, NMI, and ARI by 0.0703, 0.0896, and 0.1018, respectively, while decreasing DobsD_{\rm obs} from 0.3542 to 0.0036. The complete COIN method attained the highest macro-averaged clustering scores while sustaining a minimal representation drift. These findings indicate that nested representation consistency can be achieved with improved clustering performance across the assessed benchmarks.

Controlled unequal-information experiment. In the Synthetic3d dataset, at r=0.5r=0.5, complementary class information is positioned within the rarest view and subsequently compared to a matched permutation control. Each variant undergoes training with three distinct seeds.

VariantControl ACCInformative-view ACCGain
LREC\mathcal{L}_{\rm REC}0.442±0.0880.442\pm0.0880.451±0.0830.451\pm0.083+0.010
LREC+LINV\mathcal{L}_{\rm REC}+\mathcal{L}_{\rm INV}0.558±0.0360.558\pm0.0360.883±0.0160.883\pm0.016+0.324
Full COIN0.618±0.0230.618\pm0.0230.894±0.0110.894\pm0.011+0.276

The integration of complementary information into the rarest view results in an increase in ACC by 0.324 for REC+INV and 0.276 for Full COIN, in contrast to an increase of 0.010 for REC alone. Both invariance variants sustain a normalized rare-view ablation drift of approximately 0.004.

The analysis reveals that nested invariance plays a crucial role in maintaining the consistency of shared content. Simultaneously, it allows for the effective extraction of discriminative information from perspectives that are not frequently observed.

AuthorsOpenReview ↗

W2 / Q2. Scarce-view drop policy

The COIN-Scarce policy represents an availability-aware sampling strategy for the nested invariance loss, distinct from merely reducing the contribution of the scarce view to the overall representation. Given that a scarce view appears in fewer eligible samples, uniform sampling results in fewer nested pairs that assess its removal. COIN-Scarce employs inverse availability to enhance the coverage of these underrepresented transitions.

We conducted a comparison with Uniform and Frequent dropping at r=0.5r=0.5, utilizing the same recorded mask and three training seeds. Our verification confirmed that this design alters the realized training coverage across all datasets. The second column denotes a fixed property of the input mask, specifically the fraction of samples in which the scarcest view is observed. The subsequent columns indicate the frequency with which the observed view is selected for nested removal. Consequently, the COIN-Scarce removal rate reflects the outcome of our sampling policy, whereas Uniform and Frequent serve as comparison policies, with means calculated over three training seeds.

DatasetScarcest-view availability (fixed mask)COIN-Scarce removal rate (ours)Uniform removal rateFrequent removal rate
MSRCV123.823.8%46.246.2%27.227.2%13.013.0%
Washington26.526.5%62.362.3%45.245.2%28.928.9%
Synthetic3d42.342.3%55.055.0%49.049.0%43.043.0%
HandWritten25.925.9%52.152.1%33.733.7%18.518.5%
TinyImageNet26.526.5%58.658.6%39.139.1%21.621.6%

COIN-Scarce exhibits superior performance in scarce-view removal coverage across all five datasets, achieving an increase of 1.121.12--1.70×1.70\times compared to uniform weighting and 1.281.28--3.55×3.55\times compared to frequent weighting. This policy effectively addresses availability issues by preventing frequent views from dominating nested pairs. Such targeted coverage enhances both clustering and view removal robustness. Across all datasets, the average policy-induced maximum-minimum spreads are 0.0208/0.0177/0.02490.0208/0.0177/0.0249 for ACC/NMI/ARI, with DobsD_{\rm obs} ranging from 0.00030.0003 to 0.00390.0039 across all 15 settings. Consequently, COIN-Scarce consistently provides a coverage advantage for the least observed views while maintaining COIN's clustering competitiveness and low-drift characteristics of COIN.

AuthorsOpenReview ↗

W3 / Q3. High mask-prediction AUC despite decorrelation

The objective focuses on Cov⁡μ(C,M)\operatorname{Cov}_{\mu}(C,M) rather than independence. We maintain Full's original fixed-80/20 held-out Mask AUCs, as presented in Table 9: 0.8460.846, 0.9630.963, 1.0001.000, 0.9190.919, and 0.9240.924, indicating that the mask information remains recoverable. A distinct cross-validated statistic, which averages per-view max⁡(AUCv,1−AUCv)\max(\mathrm{AUC}_v,1-\mathrm{AUC}_v), serves as a measure of descriptive separability and does not substitute the held-out probe. On L2-normalized content, Full reduces rebalanced covariance and unweighted regularized CCA in all 10/1010/10 settings compared to REC; cross-validated separability decreases in 8/108/10 instances, whereas fixed-80/20 geometry protocol separability decreases in only 4/104/10 instances. The covariance penalty mitigates the weighted first-order association between content and observation patterns. In contrast, the ROC-AUC assesses the ranking separability of individual mask bits. These metrics capture distinct aspects of dependence, which elucidates why a lower covariance can coexist with high mask-prediction accuracy.

AuthorsOpenReview ↗

W4 / Q4. Fixed configuration, tuning fairness, and multiple masks

Direct answer. To exclude the per-cell ground-truth-guided selection from the new evaluation, we applied a uniform global code-default tuple across all five datasets and varying missing rates. The parameters were established as follows: α=1\alpha=1, β=0.5\beta=0.5, λcov=0.01\lambda_{\rm cov}=0.01, λobs=0.1\lambda_{\rm obs}=0.1, λview=λrender=1\lambda_{\rm view}=\lambda_{\rm render}=1, γ=0.5\gamma=0.5, γs=1\gamma_s=1, and wmax⁡=8w_{\max}=8. This tuple was predetermined and remained unchanged throughout the runs, independent of the labels or outcomes.

Under this single setting, the macro-averaged results over all 25 dataset-rate cells are as follows:

VariantACCNMIARIDobsD_{\rm obs}
LREC\mathcal{L}_{\rm REC}0.55900.40960.36680.3542
LREC+LINV\mathcal{L}_{\rm REC}+\mathcal{L}_{\rm INV}0.62940.49920.46860.0036
LREC+LIND\mathcal{L}_{\rm REC}+\mathcal{L}_{\rm IND}0.62690.49230.45790.2820
Full COIN0.64230.50810.47860.0046

In the absence of per-cell selection in the new track, Full COIN demonstrated the highest average ACC, NMI, and ARI among the four variants. It surpasses the REC by +0.0833+0.0833, +0.0985+0.0985, and +0.1118+0.1118, respectively, while sustaining a minimal observation-induced drift.

We conducted separate tests to assess the sensitivity to both missingness and representation-training randomness at r∈{0.3,0.5}r\in\{0.3,0.5\}. For each of the five datasets, three independently generated masks and three training seeds per mask were employed, with each mask being utilized by the methods under comparison. In the 90 matched full-versus-REC runs, the mean differences in ACC, NMI, and ARI were +0.1190+0.1190, +0.1187+0.1187, and +0.1473+0.1473, respectively, while the mean reduction in DobsD_{\rm obs} was 0.37570.3757; DobsD_{\rm obs} decreased in all 90/9090/90 matched runs. The complete results include the adverse HandWritten clustering cases and the nearly unchanged, low absolute ACC for TinyImageNet.

In the external comparison, each dataset-rate-method cell used three mask seeds and three training seeds. Following the identification of two Washington views that were inadvertently swapped in the original baseline adapter during a sample/view audit, we rectified the semantic view order prior to masking and subsequently reran all 36 affected tasks. The final masks were consistent across methods at both the original sample and semantic view levels. The corrected COIN algorithm demonstrates superior performance compared to BURG in terms of ACC, NMI, and ARI across all 10/1010/10 evaluated dataset-rate cells. Additionally, it surpassed CIMLN in 10/1010/10, 8/108/10, and 10/1010/10 cells, respectively. However, CIMLN exhibits a higher NMI on the Handwritten and Washington datasets when r=0.3r=0.3.

For transparency, historical Tables 2 and 3 track 9 α/β\alpha/\beta executions followed by 16 combinations of (λview,λrender,λcov,λobs)(\lambda_{\rm view},\lambda_{\rm render},\lambda_{\rm cov},\lambda_{\rm obs}) for the selected pair, yielding 24 unique tuples and 25 successful executions in the available archive. The selection used ground-truth ACC with NMI/ARI tie-breakers. Fixed Full COIN is below the historical Table 2 ACC in 24/2524/25 cells, with a mean difference of −0.0577-0.0577, which directly reports the oracle-to-fixed gap.

Comments on the paper

1 comment
AuthorsOpenReview ↗

W5 / Q5. Feature-dependent missingness and robustness scope

Direct answer. We evaluated COIN under two content-dependent missingness mechanisms at r=0.5r=0.5, using all five datasets and three independently generated masks per dataset-mechanism cell. Across the resulting 30 paired Full-versus-REC evaluations, Full COIN lowers DobsD_{\rm obs} in all 30 comparisons. Averaged over the ten dataset-mechanism cells, the Full-minus-REC changes are +0.0596+0.0596 ACC, +0.0443+0.0443 NMI, and +0.0532+0.0532 ARI, together with a 0.40310.4031 reduction in DobsD_{\rm obs}.

The two mechanisms have distinct clustering effects:

Missingness mechanismFull −- REC: ACC / NMI / ARIDobsD_{\rm obs}: REC →\rightarrow Full
Self-content-dependent (MNAR-style)+0.1061 / +0.1090 / +0.13580.3801 →\rightarrow 0.0012
Cross-view content-dependent (MAR-inspired)+0.0131 / -0.0203 / -0.02950.4287 →\rightarrow 0.0013

In the self-content mechanism, the missingness-selection score for a view depends on the magnitude of that view's own pre-mask PC1 score. Full COIN improves the three metrics on average and reduces DobsD_{\rm obs} from 0.38010.3801 to 0.00120.0012. In the cross-view mechanism, the score for a view depends on another view's pre-mask PC1 score with Gumbel jitter; DobsD_{\rm obs} falls from 0.42870.4287 to 0.00130.0013, while the average NMI and ARI changes are −0.0203-0.0203 and −0.0295-0.0295. We call the first mechanism MNAR-style and the second MAR-inspired, rather than strict observed-data MAR, because the predictor view in the cross-view construction may itself be missing. Both generators preserve the original unequal per-view quotas, ensure that at least one view remains observed per sample, and do not use class labels. Within each dataset, the same coefficient setting was used for both mechanisms, with no mechanism-specific coefficient choice; each dataset-mechanism cell used mask seeds 101/102/103.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.