acceptodds
Under review as a conference paper at ICLR 2027

Prototype‑Driven Reliability‑Aware Semantic Consensus Learning for Incomplete Multi‑View Clustering

Abstract

Incomplete multi-view clustering (IMVC) faces challenges from missing views, heterogeneous view distributions, and unequal view quality, making it difficult to learn reliable and consistent semantic consensus representations. To address these issues, we propose a Prototype-driven Reliability-aware Semantic Consensus Learning (PRSCL) framework, which centers on consensus semantic prototypes to align semantic representations across different views and ultimately achieves the fusion of reliable semantic representations. Specifically, PRSCL first employs a reliability-aware gating mechanism to adaptively evaluate the quality of available views and perform instance-level weighted fusion, thereby producing reliable consensus semantic representations and semantic prototypes. PRSCL then employs entropy-regularized optimal transport to derive reliable soft assignments between samples and consensus semantic prototypes, while enforcing cross-view consistency at the clustering level through bidirectional mutual supervision. By jointly modeling feature consistency, view reliability, and prototype-driven semantic structure, PRSCL learns robust consensus semantic representations without explicitly imputing missing views. Extensive experiments on benchmark datasets demonstrate its effectiveness under various missing-view conditions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering;Unsupervised learning, Prototype learning;Cross-view consistency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eBgyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Prototype-driven Reliability-aware Semantic Consensus Learning (PRSCL) framework for incomplete multi-view clustering (IMVC). Specifically, PRSCL first employs a gating mechanism to adaptively evaluate the quality of available views and perform instance-level weighted fusion, aiming at producing reliable consensus semantic representations and semantic prototypes. Then, PRSCL employs entropy-regularized optimal transport to derive soft assignments between samples and consensus semantic prototypes, and enforce cross-view consistency at the clustering level through bidirectional mutual supervision. Experimental results are provided to show the effectiveness of the proposed method.

Critical strengths
  1. Various types of experimental results are provided for empirical validation.
  2. The reliability of missing views is considered by an automatic weighting scheme.
Critical weaknesses
  1. The overall objective function comprises as many as 5 distinct loss terms. This inevitably raises the concern that the reported performance gains are largely achieved by heavily stacking loss components.
  2. The cross-view mutual supervision mechanism based on pseudo-labels is prone to be affected by incorrect clustering divisions (inaccurate prototypes and erroneous pseudo-labels) in the early stage of training, which will lead to unstable clustering performance, especially when the data missing rate is high.
  3. Only 4 datasets are used for experimental validation.
Questions for authors
  1. The proposed method seems to be a stacking of existing technologies. What is the essential contribution of this paper?
  2. How to overcome the problem of cumulative errors caused by the inaccuracy of the prototype and the inaccuracy of the pseudo-labels?
  3. Does the definition of cross-view semantic consistency rely on the assumption that each view is sufficient for clustering?
5 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q3

We sincerely thank you for the encouraging remarks and the valuable suggestions. Our responses are organized by each point below.

Response to Q1: We acknowledge that some of the fundamental techniques adopted in this work, such as entropy-regularized optimal transport, and bidirectional semantic supervision, have been explored in previous studies. Therefore, the contribution of this work does not lie in proposing these techniques individually, but in developing a task-oriented collaborative learning framework to address view reliability discrepancies and cross-view semantic consistency in incomplete multi-view clustering.

Specifically, the main improvements of our method lie in two aspects. First, unlike methods such as FreeCSL, which employ fixed completeness-based weights for view fusion, we introduce a learnable instance-level reliability-aware gating mechanism to adaptively adjust the fusion weights according to the feature information of different available views. This enables the learning of more reliable consensus representations and provides a more discriminative semantic foundation for subsequent shared prototype learning. Second, unlike methods that decouple view fusion from semantic prototype learning, PRSCL couples the two modules: the instance-level reliability-aware consensus representation GG is used to re-anchor the shared prototypes at every epoch, while the prototypes in turn provide cluster-level semantic supervision that further refines the consensus. In FreeCSL, by contrast, the shared prototypes are derived from a fixed completeness-weighted consensus. This reliability–semantic coupling, rather than shared prototypes or optimal transport per se, constitutes our core contribution.

Response to Q2: We agree that inaccurate prototypes and pseudo-labels in the early training stage may cause cumulative errors. PRSCL mitigates this issue through several designs. First, the encoder–decoder warm-up provides meaningful initial representations before prototype-based supervision begins. Second, instead of using hard pseudo-labels, entropy-regularized optimal transport produces balanced soft assignments, so that an uncertain sample distributes its mass over multiple prototypes rather than being locked to a single possibly wrong cluster. Third, the shared prototypes are re-estimated by K-means on the current reliability-weighted consensus representation GG at every epoch; early prototypes are therefore not fixed and can be progressively corrected as the representations improve. Finally, the reconstruction loss Lrec\mathcal{L}_{rec} and the feature-level cross-view contrastive loss Lccl\mathcal{L}_{ccl} provide learning signals that are independent of the cluster assignments. These designs reduce the propagation of early errors, although they cannot completely eliminate them. We will clarify this discussion in the revised paper.

Response to Q3: Our definition of cross-view semantic consistency does not assume that each individual view is independently sufficient for accurate clustering. Instead, it requires that the semantic representations of the same sample from different available views yield consistent or approximately consistent cluster assignments when evaluated against the shared semantic prototypes.

We acknowledge that different views may have varying discriminative capabilities, and some may contain noisy or insufficient semantic information. This is precisely why PRSCL incorporates a reliability-aware gating mechanism to adaptively assess the contributions of available views, together with prototype-based bidirectional supervision to promote cross-view semantic alignment. Therefore, semantic consistency refers to the alignment of clustering semantics across views, rather than requiring every view to independently achieve reliable clustering results.

Table 1. Cross-view semantic consistency and single-view clustering sufficiency.

ViewDim.CompletenessLearned wvw_vACC (learned zvz_v)
V07790.3%0.2220.822
V11390.4%0.1900.804
V26490.2%0.3350.897
V312590.0%0.3140.886
Multi-90.0%-0.914

As verified on ALOI-100, we examine per-view quality under our cross-view semantic consistency framework. From Table 1, the learned reliability weights reflect views' discriminative ability: views yielding higher single-view ACC obtain larger weights during consensus fusion. In this way, our method aggregates complementary information while tolerating views with inferior individual clustering capacity.

AuthorsOpenReview ↗

Rebuttal Part 2: Responses to W1–W2

Response to W1: We agree that using multiple loss terms may raise concerns about redundant constraints. However, the five constituent terms in PRSCL serve complementary purposes rather than simply stacking optimization objectives. Specifically, they address representation reconstruction Lrec\mathcal{L}_{rec}, feature-level cross-view alignment Lccl\mathcal{L}_{ccl}, semantic-level cross-view alignment Lcsa\mathcal{L}_{csa}, and reliability-aware fusion Lfusion\mathcal{L}_{fusion}. The latter jointly consists of semantic discrimination and entropy regularization, which coordinate consensus learning and adaptive view weighting. The ablation results in Table 2 of our paper further verify the complementary contributions of these objectives. The performance gains stem from their complementary interplay, rather than merely adding more loss terms.

Table 2. Ablation study on the BDGP dataset under different missing rates.

LcclL_{\mathrm{ccl}}LcsaL_{\mathrm{csa}}LfusionL_{\mathrm{fusion}}r=0.0 ACCr=0.0 NMIr=0.0 ARIr=0.1 ACCr=0.1 NMIr=0.1 ARIr=0.3 ACCr=0.3 NMIr=0.3 ARIr=0.5 ACCr=0.5 NMIr=0.5 ARI
✓33.529.034.8939.4417.7710.3434.0813.687.3535.219.386.52
✓94.0490.8791.2193.6490.3291.2392.1284.7486.5991.6183.1585.36
✓92.5284.1783.0394.2484.6086.3264.2852.5446.0662.0443.4636.83
✓✓96.2091.1190.9795.9290.2292.8794.1284.6387.2785.8468.9668.71
✓✓96.1691.0690.8895.0490.5492.1694.0484.3687.0886.4170.4369.93
✓✓96.8493.1693.7196.2890.9692.3791.1683.3585.6889.7279.3782.29
✓✓✓97.9694.3695.0097.1691.5093.0794.7685.5087.4793.2881.9284.05

Response to W2: We agree that inaccurate early prototypes may introduce noisy mutual supervision. PRSCL mitigates this issue through several designs. First, encoder–decoder warm-up provides meaningful initial representations before prototype-based training, while LrecL_{rec} and LcclL_{ccl} continue to provide learning signals independent of cluster assignments. Second, instead of using hard pseudo-labels, entropy-regularized OT generates balanced soft assignments, avoiding over-sharp supervision and reducing the risk of cluster collapse (Section 2.4). Third, the shared prototypes are periodically re-estimated by K-means on the evolving reliability-weighted consensus representations (Eq. (8)); thus, inaccurate early prototypes are not fixed but can be progressively revised. Meanwhile, the reliability-aware gate excludes missing views from consensus construction without explicit imputation, avoiding additional noise from unreliable completion. Therefore, although early assignment errors cannot be completely eliminated, PRSCL reduces their propagation and enables the prototypes and soft assignments to be refined jointly during training.

AuthorsOpenReview ↗

Rebuttal Part 3: Responses to W3

Response to W3: Evaluations over a broader spectrum of datasets are essential to validate the generalizability of incomplete multi-view clustering algorithms. Following your suggestion, we have supplemented two additional benchmark datasets, i.e., CUB (600 instances, 2 views) and YouTubeFace10 (38654 instances, 4 views), which cover fine-grained bird recognition and face identification tasks with distinct data distributions. The full quantitative results under four missing rates (r∈{0.0,0.1,0.3,0.5}r \in \{0.0, 0.1, 0.3, 0.5\}) are reported in Table 3, including ACC, NMI and ARI. As shown, our PRSCL achieves competitive or state-of-the-art performance across most settings, even under severe view missing. These results further demonstrate the effectiveness and robustness of our method.

Table 3. Performance comparison on CUB and YouTubeFace10. The best result of each metric under each missing rate is highlighted in bold.

DatasetMethodr=0.0 ACCr=0.0 NMIr=0.0 ARIr=0.1 ACCr=0.1 NMIr=0.1 ARIr=0.3 ACCr=0.3 NMIr=0.3 ARIr=0.5 ACCr=0.5 NMIr=0.5 ARI
CUBDCP53.2063.9040.0551.6059.2132.9437.4750.6022.6221.1721.1220.05
CPSPAN63.5069.0151.2363.8370.7853.7361.6764.5748.1664.6766.7953.01
ProImp81.6074.3365.8778.9372.9363.6574.8068.5257.4570.9064.9252.75
RPCIC74.3373.2461.7869.1769.4154.9359.5063.3847.0663.3363.6848.24
PMIMC57.8365.8049.8563.5069.2153.1675.8372.2860.2864.8366.5053.80
FreeCSL75.0168.9556.7766.1767.5154.1367.3368.5656.2756.6658.3541.97
HSACC62.0769.5149.7660.0768.1048.8361.2366.6545.1952.9055.8238.45
DCG77.1771.9760.5075.8372.3660.5874.0167.1854.9664.1765.9751.42
CAMERA79.6773.8163.9079.9473.5164.2378.5071.5861.0171.1165.8254.13
GLGC58.8356.4241.7854.4352.1034.9454.1749.7934.7044.3041.0722.75
OURS82.1774.4166.4980.1773.1465.0779.8372.1063.3176.0068.4157.83
YouTubeFace10DCP60.8565.7832.3862.0469.8163.4754.2863.3552.3356.5162.4358.22
CPSPAN69.8074.6959.3573.9479.2171.2768.1279.3464.1772.6880.0572.93
ProImp49.2654.6032.7353.4759.1152.2155.7163.9054.9253.8262.2452.02
RPCIC70.2377.0968.4174.4579.1473.0573.5875.3774.8674.2674.7673.64
PMIMC78.1582.6071.9171.2476.1563.8876.6880.5371.0175.8980.6970.49
FreeCSL77.4281.0770.4380.9981.7171.3079.9981.6171.5873.2878.3368.17
HSACC38.7741.197.4840.4943.488.3239.6044.8410.5038.1141.937.98
DCG82.8280.7373.0184.7783.3076.2677.8781.9074.6371.1476.5664.69
CAMERA74.3677.8967.7276.6579.0868.3073.2778.1865.2371.3177.7366.77
GLGC28.0430.3412.8232.1433.8715.1929.8831.7114.6327.5529.8311.32
OURS83.1882.7872.2784.1884.4374.0286.7284.6275.3583.4683.4772.79
Reviewer_4PUBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Prototype-driven Reliability-aware Semantic Consensus Learning (PRSCL) framework for incomplete multi-view clustering. The framework addresses challenges such as missing views, heterogeneous view distributions, and unequal view quality by centering on consensus semantic prototypes to align semantic representations across views. Specifically, PRSCL first employs a reliability-aware gating mechanism to evaluate the quality of available views and perform instance-level weighted fusion, thereby producing reliable consensus semantic representations and semantic prototypes. It then uses entropy-regularized optimal transport to derive soft assignments between samples and prototypes, while enforcing cross-view consistency at the clustering level through bidirectional mutual supervision. By jointly modeling feature consistency, view reliability, and prototype-driven semantic structure, PRSCL learns robust consensus representations without explicitly imputing missing views, and experiments verify its effectiveness.

Critical strengths
  1. A reliability-aware gating mechanism adaptively weighs available views, while entropy-regularized optimal transport and bidirectional mutual supervision enforce cross-view consistency, yielding reliable consensus representations without explicit imputation.

2.Joint modeling of multiple factors: By unifying feature consistency, view reliability, and prototype-driven semantic structure, PRSCL handles missing views, heterogeneous distributions, and unequal view quality, proving effective under various missing-view conditions.

Critical weaknesses

(1) Prototype-based missing data completion has been published in many papers and is not a novel concept, e.g., PMIMC and RPCIC. The authors did not analyze the differences from these works, and the number of experimental datasets is small, with experiments conducted only on simple datasets. (2) Prototype-based missing data completion suffers from the concept of prototype misalignment, and the authors did not discuss this key issue.

Questions for authors

1.Please analyze the time complexity of the proposed method and compare it with that of the competing methods. 2.How does the proposed prototype completion strategy differ from existing approaches? 3.How does the proposed method account for prototype misalignment?

4 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q3 and W1-W2

We sincerely thank you for the encouraging remarks and the valuable suggestions. Our responses are organized by each point below.

Response to Q1: We analyze the per-epoch computational complexity of PRSCL, which scales linearly with the number of samples, O(N)O(N). Its main pairwise term originates from InfoNCE and is common to most autoencoder-based baselines. Compared with FreeCSL, PRSCL removes expensive KNN-graph and spectral modules and runs K-means only once per epoch. It is also cheaper than DCG, GLGC and RPCIC. We will add this complexity analysis together with measured per-epoch training time to the revised paper.

Table 1. Per-epoch computational complexity comparison.

MethodDominant per-epoch complexityModules
DCPO(V2Nd2+VNDd)\mathcal{O}(V^2 N d^2 + V N D d)Instance contrast in feature space + dual prediction
CPSPANO(V2Nd2+VNDd)\mathcal{O}(V^2 N d^2 + V N D d)Prototype + instance alignment
ProImpO(NBd+N⋅num⋅d)\mathcal{O}(N B d + N \cdot \text{num} \cdot d)Instance contrast + cross-attention
PMIMCO(V2NBd+NPd)\mathcal{O}(V^2 N B d + N P d)Instance contrast + prototype completion
RPCICO(V2NBd+P2d)\mathcal{O}(V^2 N B d + P^2 d)Instance contrast + global prototype alignment over all prototypes
FreeCSLO(VNBD+VNBd+NKdTkm)\mathcal{O}(V N B D + V N B d + N K d T_{\text{km}})KNN graph construction + GCN spectral loss + two K-means runs per epoch
HSACCO(VNBd+VNd2)\mathcal{O}(V N B d + V N d^2)MMD with Gaussian kernel + instance contrast + dual inference
DCGO(TNU)\mathcal{O}(T N U)Diffusion denoising (TT denoising steps per batch per epoch; UU: cost of one UNet step)
CAMERAO(NBd+N⋅num⋅d)\mathcal{O}(N B d + N \cdot \text{num} \cdot d)Sample consistency + mutual attention
GLGCO(V2NBd+VNBd)\mathcal{O}(V^2 N B d + V N B d)Hypergraph (HGNN) + local / global graph contrast with multiple O(B2)\mathcal{O}(B^2) graphs
PRSCL (Ours)O(V2NBd+V2NKd+VNDd+NKdTkm)\mathcal{O}(V^2 N B d + V^2 N K d + V N D d + N K d T_{\text{km}})InfoNCE + Sinkhorn prototype constraint + one K-means run per epoch

Response to Q2 & W1: It is worth emphasizing that our method does not adopt prototypes for missing view completion. As summarized in Table 2, the compared methods differ in prototype definition, organization, role, and misalignment treatment. Our prototypes are KK shared cluster centroids, which are re-anchored per epoch via K-means on reliability-weighted consensus representations and perform semantic assignment with Sinkhorn-Knopp OT for clustering consensus. Missing samples are reconstructed by the autoencoder instead of being imputed via prototypes.

In contrast, PMIMC and RPCIC employ view-specific batch-wise representative samples as prototypes, whose primary function is missing value imputation; RPCIC further requires an extra greedy cross-view prototype matching step to relieve prototype misalignment. Besides, CPSPAN's called "prototypes" are actually feature dimension basis vectors for MSE-based alignment, without cluster centroids or OT assignment. Although our method shares similar prototype-anchoring and OT pipelines with FreeCSL, the main distinction lies in our learnable reliability-aware gating fusion rather than fixed completeness-based weights.

Table 2. Comparison of prototype definitions and roles in representative incomplete multi-view methods. PRSCL employs shared cluster-level prototypes for semantic supervision rather than prototype-based missing-data completion.

MethodPrototype definitionOrganizationPrimary roleMisalignment treatment
PMIMCRepresentative samplesView-specificMissing-feature completion and robust alignmentNo explicit prototype matching
RPCICRepresentative samplesView-specificCompletion after cross-view prototype alignmentExplicit greedy prototype matching
CPSPANFeature-dimensional basesView-specificFeature-level cross-view alignmentMSE-based basis alignment
FreeCSLCluster centroidsSharedCluster-level semantic assignmentShared references without view-specific matching
PRSCL (Ours)Cluster centroidsSharedSoft assignment and mutual supervision, not completionShared references and periodic K-means re-anchoring on reliability-weighted G

Response to Q3 & W2: PRSCL avoids the cross-view prototype misalignment problem by construction: it maintains a single set of shared cluster prototypes that are re-anchored on the reliability-weighted consensus representation GG at every epoch, so there are no independent view-specific prototype sets that require matching. The remaining prototype-representation drift is alleviated by the periodic K-means re-anchoring.

To improve readability, we add discussions regarding prototype misalignment and clarify the distinctions between our method and existing prototype-driven approaches in the related work section.

AuthorsOpenReview ↗

Addition experience for CUB and YouTubeFace10

Evaluations over a broader spectrum of datasets are essential to validate the generalizability of incomplete multi-view clustering algorithms. Following your suggestion, we have supplemented two additional benchmark datasets, i.e., CUB (600 instances, 2 views) and YouTubeFace10 (38654 instances, 4 views), which cover fine-grained bird recognition and face identification tasks with distinct data distributions. The full quantitative results under four missing rates (r∈{0.0,0.1,0.3,0.5}r \in \{0.0, 0.1, 0.3, 0.5\}) are reported in Table 3, including ACC, NMI and ARI. As shown, our PRSCL achieves competitive or state-of-the-art performance across most settings, even under severe view missing. These results further demonstrate the effectiveness and robustness of our method.

Table 3. Performance comparison on CUB and YouTubeFace10. The best result of each metric under each missing rate is highlighted in bold.

DatasetMethodr=0.0 ACCr=0.0 NMIr=0.0 ARIr=0.1 ACCr=0.1 NMIr=0.1 ARIr=0.3 ACCr=0.3 NMIr=0.3 ARIr=0.5 ACCr=0.5 NMIr=0.5 ARI
CUBDCP53.2063.9040.0551.6059.2132.9437.4750.6022.6221.1721.1220.05
CPSPAN63.5069.0151.2363.8370.7853.7361.6764.5748.1664.6766.7953.01
ProImp81.6074.3365.8778.9372.9363.6574.8068.5257.4570.9064.9252.75
RPCIC74.3373.2461.7869.1769.4154.9359.5063.3847.0663.3363.6848.24
PMIMC57.8365.8049.8563.5069.2153.1675.8372.2860.2864.8366.5053.80
FreeCSL75.0168.9556.7766.1767.5154.1367.3368.5656.2756.6658.3541.97
HSACC62.0769.5149.7660.0768.1048.8361.2366.6545.1952.9055.8238.45
DCG77.1771.9760.5075.8372.3660.5874.0167.1854.9664.1765.9751.42
CAMERA79.6773.8163.9079.9473.5164.2378.5071.5861.0171.1165.8254.13
GLGC58.8356.4241.7854.4352.1034.9454.1749.7934.7044.3041.0722.75
OURS82.1774.4166.4980.1773.1465.0779.8372.1063.3176.0068.4157.83
YouTubeFace10DCP60.8565.7832.3862.0469.8163.4754.2863.3552.3356.5162.4358.22
CPSPAN69.8074.6959.3573.9479.2171.2768.1279.3464.1772.6880.0572.93
ProImp49.2654.6032.7353.4759.1152.2155.7163.9054.9253.8262.2452.02
RPCIC70.2377.0968.4174.4579.1473.0573.5875.3774.8674.2674.7673.64
PMIMC78.1582.6071.9171.2476.1563.8876.6880.5371.0175.8980.6970.49
FreeCSL77.4281.0770.4380.9981.7171.3079.9981.6171.5873.2878.3368.17
HSACC38.7741.197.4840.4943.488.3239.6044.8410.5038.1141.937.98
DCG82.8280.7373.0184.7783.3076.2677.8781.9074.6371.1476.5664.69
CAMERA74.3677.8967.7276.6579.0868.3073.2778.1865.2371.3177.7366.77
GLGC28.0430.3412.8232.1433.8715.1929.8831.7114.6327.5529.8311.32
OURS83.1882.7872.2784.1884.4374.0286.7284.6275.3583.4683.4772.79

Thank you for the author's reply. The author still has not clearly explained the prototype misalignment issue. Why is the prototype filling algorithm used better than previous methods? Could the author add an experiment to demonstrate that their prototype is better than other methods?

AuthorsOpenReview ↗

Rebuttal: Prototype Misalignment and Additional Experiments

Thank you very much for your re-evaluation. Unlike PMIMC and RPCIC, which adopt view-specific prototypes for missing-view imputation and require explicit cross-view prototype matching during training, our method maintains a single set of shared prototypes in the consensus semantic space for all views. This design naturally eliminates the prototype matching burden.

To further verify that our prototypes are better aligned across views and compatible with evolving representations, we compare PRSCL with FreeCSL. We assign each view embedding to its nearest shared prototype and evaluate pairwise inter-view consistency and view-consensus consistency. In Table 1, PRSCL outperforms FreeCSL on all four metrics, demonstrating that our shared prototypes achieve stronger cross-view alignment than FreeCSL's prototypes derived from fixed completeness weighting.

Table 1. Cross-view prototype-assignment consistency on ALOI-100 (r=0.1r=0.1).

MetricPRSCL (Ours)FreeCSL
Mean pairwise NMI0.8540.750
Mean pairwise ARI0.6650.499
Mean view–consensus NMI0.9180.833
Mean view–consensus ARI0.8040.651

( (i) Pairwise consistency: the average NMI/ARI between assignments of every view pair, computed only on commonly observed samples. (ii) View-consensus consistency: the average NMI/ARI between each view's assignment and the consensus assignment, computed on samples visible to that view. Higher values indicate better cross-view prototype alignment. These metrics quantify how consistently different views and the consensus assign samples to the same shared prototypes.)

In addition, to further validate our reliability-aware gating against explicit imputation strategies, we replace its missing-view masking with four imputation strategies applied in the embedding space, while keeping all other components unchanged: mean imputation, KNN imputation (K=5K=5), learned weighted imputation, and GCN imputation (adopted from ICMVC [1], top-k=10k=10).

Among the imputation variants in Table 2, the structure-aware strategies clearly outperform the simple ones: GCN imputation attains ACC/NMI/ARI of 91.57/92.89/89.19 and KNN imputation 90.80/94.24/86.37, whereas mean imputation and weighted imputation degrade to 77.82 and 74.14, respectively. Our reliability-weighted fusion achieves 93.07/94.53/88.43, remaining competitive with the best imputation variants without explicitly reconstructing missing views, which further validates the effectiveness of the gating mechanism.

Table 2. Ablation on missing-view handling strategies on ALOI-100 (r=0.3r=0.3).

StrategyACCNMIARI
PRSCL93.0794.5388.43
+ KNN imputation90.8094.2486.37
+ Mean imputation77.8287.1771.32
+ Weighted imputation74.1484.8866.53
+ GCN imputation91.5792.8989.19

Although the reviewer specifically asked about prototype filling, our reliability-aware gating can be viewed as an implicit alternative to explicit prototype-based imputation: instead of reconstructing missing views, we down-weight unreliable view contributions during fusion.

Thank you again for your valuable feedback. We will incorporate this detailed discussion and the corresponding experimental analysis into the revised version of our manuscript.

References

[1] Chao, G., Jiang, Y., & Chu, D. (2024). Incomplete Contrastive Multi-View Clustering with High-Confidence Guiding. Proceedings of the AAAI Conference on Artificial Intelligence, 38(10), 11221–11229.

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

The paper proposes a Prototype-Driven Reliability-Aware Semantic Consensus Learning (PRSCL) method for incomplete multi-view clustering. The method first extracts view-specific features using autoencoders and cross-view contrastive learning, then adaptively assigns fusion weights to different views through a reliability-aware gating mechanism to construct a consensus semantic representation, from which shared semantic prototypes are obtained using K-Means. Furthermore, entropy-regularized optimal transport is employed to generate balanced pseudo-labels, and bidirectional mutual supervision is introduced to achieve cross-view semantic alignment.

Critical strengths

This paper proposes the prototype-driven cross-view semantic alignment that adaptively weights available views to establish a reliable consensus and leverages prototype-based soft assignments to promote cross-view clustering consistency, thereby enabling semantically consistent representation learning across views.

Critical weaknesses

1.This paper proposes a prototype-based clustering algorithm for incomplete multi-view data; however, it remains somewhat weak in terms of problem-solving and methodological innovation.

2.The use of only four benchmark datasets is insufficient to demonstrate the algorithm's effectiveness across diverse data.

Questions for authors

Q1. Could the authors explicitly map the differences from FreeCSL, especially its Section 2.3, and isolate their effects using a common encoder and evaluation protocol? Q2. How is the missing rate r defined and generated? Q3. Are the contrastive and KL losses restricted to the intersection of observed instance identities for each view pair? Eqs. (2), (15) and (16) do not specify this restriction, although different views can contain different numbers of observations. Q4. What evidence shows that the learned weights reflect reliability rather than merely task-dependent preference? Q5. Section 3.3 refers to detailed appendix settings, but Appendix B.2 mainly gives the hardware and architecture. Please explain the label-free selection rule for these hyperparameters and include training time with repeated K-means and OT computations.

5 comments
AuthorsOpenReview ↗

Rebuttal Part 1: Responses to Q1–Q3

We sincerely thank you for the encouraging remarks and the valuable suggestions. Our responses are organized by each point below.

Response to Q1: We agree that a clearer comparison with FreeCSL, particularly its Section 2.3, is necessary. Both FreeCSL and PRSCL employ consensus prototypes and entropy-regularized optimal transport for semantic assignment. However, their key difference lies in view fusion: FreeCSL uses fixed completeness-based weights determined by view availability, whereas PRSCL introduces a learnable reliability-aware gating mechanism to adaptively evaluate the semantic quality of each available view. Furthermore, PRSCL explicitly incorporates semantic constraints into the reliability-weighted consensus learning process.

In Table 1, we further compare PRSCL with representative prototype-based methods. This comparison highlights that PRSCL differs from prior work in both how the prototypes are constructed and how they are used.

Table 1. Comparison of prototype definitions and roles in representative incomplete multi-view methods. PRSCL employs shared cluster-level prototypes for semantic supervision rather than prototype-based missing-data completion.

MethodPrototype definitionOrganizationPrimary roleMisalignment treatment
PMIMCRepresentative samplesView-specificMissing-feature completion and robust alignmentNo explicit prototype matching
RPCICRepresentative samplesView-specificCompletion after cross-view prototype alignmentExplicit greedy prototype matching
CPSPANFeature-dimensional basesView-specificFeature-level cross-view alignmentMSE-based basis alignment
FreeCSLCluster centroidsSharedCluster-level semantic assignmentShared references without view-specific matching
PRSCL (Ours)Cluster centroidsSharedSoft assignment and mutual supervision, not completionShared references and periodic K-means re-anchoring on reliability-weighted G

In Table 2, we re-ran FreeCSL and PRSCL on BDGP under a common encoder (128-dim), shared masks, same batch size and epochs, and the same final-epoch K-means evaluation, over three seeds. PRSCL consistently outperforms FreeCSL. The key difference is that FreeCSL uses completeness-based mean fusion, whereas PRSCL learns instance-level reliability-aware gating. This supports that the fusion strategy, rather than the backbone or evaluation setting, drives the improvement.

Table 2. Comparison of PRSCL and FreeCSL on BDGP under a common encoder and evaluation protocol.

Methodr=0.1 · ACC(%)r=0.1 · NMI(%)r=0.1 · ARI(%)r=0.5 · ACC(%)r=0.5 · NMI(%)r=0.5 · ARI(%)
PRSCL (Ours)96.67 ± 0.7090.97 ± 1.3291.94 ± 1.6085.19 ± 8.5974.24 ± 6.1172.77 ± 9.47
FreeCSL82.73 ± 10.4872.65 ± 13.5967.29 ± 17.6463.25 ± 2.4549.02 ± 11.1143.69 ± 10.89

Response to Q2: Following our compared methods including CPSPAN, ProImp, RPCIC and PMIMC, the missing rate rr is defined as the fraction of missing instances within each individual view. For each view, we independently remove rr of its samples. Formally, for view vv,

r=NvmissN,r = \frac{N_v^{\text{miss}}}{N},

where NN is the total number of samples and NvmissN_v^{\text{miss}} is the count of missing samples in view vv. To exclude invalid configurations, we enforce that each sample is retained in at least one view.

Response to Q3: The contrastive and KL losses are restricted to the shared observed instances of each view pair; samples missing in either view are skipped. We will explicitly add this restriction to Eqs. (2), (15), (16) in the revision.

Taking Equation (2) as an example, we define a new set Iv,u\mathcal{I}_{v,u} to collect paired observed instances:

lcclvu=−1Iv,u∑i∈Iv,ulog⁡exp⁡(S(ziv,ziu)/τ)∑j∈Iv,uexp⁡(S(ziv,zju)/τ)l_{\mathrm{ccl}}^{vu} = -\frac{1}{I_{v,u}}\sum_{i\in I_{v,u}}\log \frac{\exp(S(z_i^v, z_i^u)/\tau)}{\sum_{j\in I_{v,u}} \exp(S(z_i^v, z_j^u)/\tau)}

where Iv,u=∣Iv,u∣I_{v,u}=|\mathcal{I}_{v,u}| denotes the index set of instances observed in both view vv and view uu, S(a,b)S(a,b) denotes the similarity between two samples (e.g., cosine similarity), and the temperature hyper-parameter τ\tau is set to 1.

AuthorsOpenReview ↗

Rebuttal Part 2: Responses to Q4 -Q5

Response to Q4: The weights are computed without relying on labels or downstream metrics including ACC. Empirical results further show that they are not arbitrary or static preferences, but can respond to changes in view quality and completeness.

First, the weight-generation process does not use class labels or any supervised downstream-task metric. For sample ii and view vv, the quality gate produces si(v)s_i^{(v)} from its latent representation. The normalized weight is computed as

αi(v)=Mi,vexp⁡(si(v)/τ)∑u=1VMi,uexp⁡(si(u)/τ),\alpha_i^{(v)} = \frac{M_{i,v}\exp(s_i^{(v)}/\tau)}{\sum_{u=1}^{V}M_{i,u}\exp(s_i^{(u)}/\tau)},

where Mi,vM_{i,v} indicates whether view vv is observed. Consequently, a missing view receives exactly zero weight, and the weights are normalized only over the available views. The gate is learned using unsupervised reconstruction and cross-view consistency/clustering objectives; neither ground-truth labels nor clustering ACC is supplied to the gate during training. Thus, the weights characterize the relative reliability of the available representations for constructing a cross-view consensus.

Second, the view-wise results in Table 3 show that the learned weights are consistent with post-hoc single-view clustering quality. The four views have nearly identical completeness rates (90.0%–90.4%), but substantially different learned weights. Nevertheless, their weights exhibit perfect rank agreement with single-view ACC in this experiment: V2 and V3 obtain higher ACC values (0.897 and 0.886) and receive larger weights (0.335 and 0.314), whereas V0 and V1 obtain lower ACC values (0.822 and 0.804) and receive smaller weights (0.222 and 0.190). Since the completeness rates are almost identical, this ordering cannot be explained by availability alone. Moreover, the fused representation achieves an ACC of 0.914, exceeding the best single-view result of 0.897 by 1.7 percentage points, suggesting that the learned weights facilitate effective integration of complementary information rather than merely selecting a preferred view.

Table 3. Cross-view semantic consistency and single-view clustering sufficiency.

ViewDim.CompletenessLearned wvw_vACC (learned zvz_v)
V07790.3%0.2220.822
V11390.4%0.1900.804
V26490.2%0.3350.897
V312590.0%0.3140.886
Multi-90.0%-0.914

More direct evidence comes from our controlled asymmetric-missingness experiment. We gradually vary the completeness of a single target view from 100% to 80%, 60%, and finally 40%, while keeping all remaining views fully observed, and retrain the model separately for each completeness setting. As summarized in Table 4, for each target view, its average learned weight decreases monotonically with declining completeness: V0 drops from 0.199 to 0.179, 0.155 and 0.154; V1 from 0.176 to 0.141, 0.116 and 0.111; V2 from 0.327 to 0.283, 0.269 and 0.249; V3 from 0.311 to 0.271, 0.237 and 0.224. Meanwhile, the relative weights allocated to the other available views generally rise, demonstrating adaptive weight redistribution instead of a fixed static preference over views.

Table 4. Learned weights under varying target-view completeness (ALOI-100).

View100%80%60%40%
V00.1990.1790.1550.154
V10.1760.1410.1160.111
V20.3270.2830.2690.249
V30.3110.2710.2370.224

These results support interpreting the learned weights as task-relevant reliability estimates for consensus learning, rather than merely task-dependent preference.

Response to Q5: Label-free selection rule: Our hyperparameters are selected via a label-free validation strategy relying only on embedding quality metrics. We will add this rule to Appendix in revision.

Table 5. Hyperparameter settings for each dataset.

Datasetλβγ
ALOI-100110.01
BDGP111
CUB111
MNIST0.10.010.01
Scene-150.10.10.001
YouTubeFace10110.1

Training time: The overall per-epoch complexity of our method is O(V2NBd+V2NKd+VNDd+NKdTkm)O(V^2 N B d + V^2 N K d + V N D d + N K d T_{\text{km}}). Among these terms, the K-means and Sinkhorn–Knopp optimal transport subroutines scale linearly with NN, captured by O(NKdTkm)O(N K d T_{\text{km}}). We have added a complexity comparison against competing methods in the revised appendix. Taking ALOI-100 as an example, one full run on a single NVIDIA GPU 4060 takes approximately 254 seconds: 22 seconds for 50 pretraining epochs and 232 seconds for 120 main training epochs, averaging 1.9 seconds per epoch. This measurement includes one K-means prototype re-anchoring per epoch (0.9 s/epoch) and all Sinkhorn–Knopp optimal-transport computations: there are 77 assignments with three iterations each, costing roughly 5.5 ms per assignment and summing to about 0.4 s per epoch.

AuthorsOpenReview ↗

Rebuttal Part 3: Responses to W1–W2

Response to W1: We agree that shared prototypes and optimal transport are not individually novel. Our contribution is how they are reorganized to jointly solve view-reliability inconsistency and cross-view semantic misalignment without explicit missing-view completion. Specifically, we introduce instance-level reliability-aware gating and couple it with semantic prototype learning: the reliability-weighted consensus re-anchors the shared prototypes, which in turn supervise the consensus. Furthermore, extensive experiments on multiple benchmark datasets consistently demonstrate strong clustering performance across various missing rates, validating the effectiveness and generality of our method.

Response to W2: Evaluations over a broader spectrum of datasets are essential to validate the generalizability of incomplete multi-view clustering algorithms. Following your suggestion, we have supplemented two additional benchmark datasets, i.e., CUB (600 instances, 2 views) and YouTubeFace10 (38654 instances, 4 views), which cover fine-grained bird recognition and face identification tasks with distinct data distributions. The full quantitative results under four missing rates (r∈{0.0,0.1,0.3,0.5}r \in \{0.0, 0.1, 0.3, 0.5\}) are reported in Table 6, including ACC, NMI and ARI. As shown, our PRSCL achieves competitive or state-of-the-art performance across most settings, even under severe view missing. These results further demonstrate the effectiveness and robustness of our method.

Table 6. Performance comparison on CUB and YouTubeFace10. The best result of each metric under each missing rate is highlighted in bold.

DatasetMethodr=0.0 ACCr=0.0 NMIr=0.0 ARIr=0.1 ACCr=0.1 NMIr=0.1 ARIr=0.3 ACCr=0.3 NMIr=0.3 ARIr=0.5 ACCr=0.5 NMIr=0.5 ARI
CUBDCP53.2063.9040.0551.6059.2132.9437.4750.6022.6221.1721.1220.05
CPSPAN63.5069.0151.2363.8370.7853.7361.6764.5748.1664.6766.7953.01
ProImp81.6074.3365.8778.9372.9363.6574.8068.5257.4570.9064.9252.75
RPCIC74.3373.2461.7869.1769.4154.9359.5063.3847.0663.3363.6848.24
PMIMC57.8365.8049.8563.5069.2153.1675.8372.2860.2864.8366.5053.80
FreeCSL75.0168.9556.7766.1767.5154.1367.3368.5656.2756.6658.3541.97
HSACC62.0769.5149.7660.0768.1048.8361.2366.6545.1952.9055.8238.45
DCG77.1771.9760.5075.8372.3660.5874.0167.1854.9664.1765.9751.42
CAMERA79.6773.8163.9079.9473.5164.2378.5071.5861.0171.1165.8254.13
GLGC58.8356.4241.7854.4352.1034.9454.1749.7934.7044.3041.0722.75
OURS82.1774.4166.4980.1773.1465.0779.8372.1063.3176.0068.4157.83
YouTubeFace10DCP60.8565.7832.3862.0469.8163.4754.2863.3552.3356.5162.4358.22
CPSPAN69.8074.6959.3573.9479.2171.2768.1279.3464.1772.6880.0572.93
ProImp49.2654.6032.7353.4759.1152.2155.7163.9054.9253.8262.2452.02
RPCIC70.2377.0968.4174.4579.1473.0573.5875.3774.8674.2674.7673.64
PMIMC78.1582.6071.9171.2476.1563.8876.6880.5371.0175.8980.6970.49
FreeCSL77.4281.0770.4380.9981.7171.3079.9981.6171.5873.2878.3368.17
HSACC38.7741.197.4840.4943.488.3239.6044.8410.5038.1141.937.98
DCG82.8280.7373.0184.7783.3076.2677.8781.9074.6371.1476.5664.69
CAMERA74.3677.8967.7276.6579.0868.3073.2778.1865.2371.3177.7366.77
GLGC28.0430.3412.8232.1433.8715.1929.8831.7114.6327.5529.8311.32
OURS83.1882.7872.2784.1884.4374.0286.7284.6275.3583.4683.4772.79

Comments on the paper

1 comment

Dear Reviewers and Authors,

As a quick reminder, the author-reviewer discussion period will close on November 18.

@Reviewers: Please make sure to read the authors' rebuttals, acknowledge their responses, and engage in further discussion if any points remain unclear. While final decisions are not made at this stage, we ask that you carefully re-evaluate your ratings in light of the new clarifications and experimental results:

Increase your score if the authors have satisfactorily resolved your concerns and strengthened the paper.

Decrease your score if new, unaddressed issues have emerged during the discussion.

Maintain your score if critical concerns remain unresolved.

Please note that you are under no obligation to change your rating, provided your final score accurately reflects the current state of the manuscript. To ensure a fair evaluation, please ground your assessment in concrete evidence rather than personal preference, and consider whether the work offers meaningful scientific value to the ICLR community.

Thank you very much for your time, thoughtful engagement, and contributions to the review process.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.