acceptodds
Under review as a conference paper at ICLR 2027

Towards Discriminative Incomplete Multi-View Clustering via Prototype Diffusion

Abstract

Incomplete multi-view clustering (IMVC) aims to learn unified representations from multi-source data with missing views to realize unsupervised clustering, which has broad application value in various real-world scenarios. Most existing IMVC methods only rely on limited paired complete samples to implement instance-level cross-view alignment and lack global semantic anchors. This drawback easily leads to cross-view semantic inconsistency and latent feature collapse, which further degrades the discriminability of representations. To tackle these issues, this paper proposes a Discriminative Prototype Diffusion framework named DPD. We adopt learnable shared prototypes as global semantic anchors to unify multi-view latent distributions through prototype-guided distribution alignment. Global category constraints are further introduced to mitigate latent feature collapse and restrain feature space compression. Meanwhile, the framework is equipped with mask-aware InfoNCE loss to achieve basic cross-view feature alignment by maximizing multi-view mutual information. We utilize the prototype alignment module to regularize latent diffusion inference for recovering missing views in the semantically aligned latent space. A confidence-aware prototype fusion module is constructed to adaptively fuse complementary information of all views and generate highly discriminative latent representations for clustering. Comparative experiments on multiple standard benchmark datasets demonstrate that the proposed DPD achieves superior clustering performance on incomplete multi-view clustering tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-View Clustering, Diffusion

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

This paper addresses the problem of cross-view semantic alignment in incomplete multi-view clustering and proposes DPD, a framework that combines prototype diffusion with distribution alignment. It enhances semantic consistency across different views through shared prototypes and generates unified representations using a mask-aware fusion module. Experiments on five benchmark datasets under different missing rates demonstrate the competitive performance of the proposed method.

Critical strengths

The paper focuses on the problem of missing views in incomplete multi-view clustering. The proposed method has some novelty in combining prototype diffusion with alignment mechanisms to leverage information across different views and improve clustering performance. Overall, the paper presents a relatively clear and understandable approach.

Critical weaknesses
  1. The prototype distribution alignment, instance-level contrastive learning, and latent diffusion regularization techniques used in the paper are all combinations of existing mechanisms. Ablation studies that remove individual loss terms alone are insufficient to clarify whether these mechanisms have indispensable technical connections, leaving it unclear whether the core framework offers substantive innovation that distinguishes it from existing methods.

  2. The effectiveness of the prototype distribution alignment mechanism is difficult to assess. MMD primarily constrains the overall distribution discrepancy between different views and the shared prototype distribution, but does not explicitly establish semantic correspondences between samples and prototypes. In addition, the effects of the number of prototypes and their initialization strategy on model performance remain to be verified.

  3. The training and fusion strategies have certain limitations. The paper primarily models the problem in a two-view setting, while missing views are mainly handled through the final mask-weighted fusion. This design does not fully exploit the large number of unpaired samples, and the method's ability to scale effectively to more complex scenarios involving a larger number of views has yet to be demonstrated.

  4. The paper does not sufficiently distinguish the individual contributions and interaction effects of prototype alignment and diffusion regularization through experiments. The ablation study only evaluates the contribution of each component by removing individual loss terms, which is insufficient to verify whether combining the two components provides additional benefits.

  5. The paper lacks parameter sensitivity analysis, convergence analysis, and feature visualization on additional datasets, making it difficult to fully demonstrate the effectiveness of the proposed method across different datasets.

Questions for authors
  1. The proposed diffusion module does not explicitly use prototype information for conditional guidance, but instead establishes an indirect connection through shared latent representations. What unique advantages does this design offer compared with conventional latent diffusion regularization?

  2. The paper lacks sufficient experiments to demonstrate the actual runtime, GPU memory consumption, and scalability trends of the proposed method on larger datasets. It is recommended that the authors provide a clearer complexity analysis and conduct experiments on larger-scale datasets.

  3. The authors do not report the standard deviations across different runs or conduct statistical significance tests. For example, on the LandUse-21 dataset under the 0.7 missing-rate setting, the NMI is 29.37, while COMP achieves 29.11. Without variance information, it is difficult to determine whether the reported performance difference is statistically significant.

4 comments

Q1. Advantages of implicit prototype-diffusion coupling

Thank you for the question. Implicit coupling lets shared prototypes directly constrain the final clustering representations while keeping the denoiser simple.

Compared with standalone latent diffusion regularization, standard view-wise noise prediction denoises each latent distribution but does not itself require both views to align with a common distribution. View-specific distribution differences can therefore remain after denoising. DPD aligns both views with shared prototypes (Eq. 12) and denoises these representations (Eqs. 15-16), so distribution alignment and denoising jointly shape the clustering space.

Why implicit rather than explicit prototype conditioning? Conditioning on a selected prototype requires estimating sample-prototype associations. Early inaccuracies can enter noise prediction directly. Soft attention over all prototypes reduces dependence on a single assignment but requires condition construction and injection. DPD places the prototype constraint in distribution alignment, while the denoiser receives only noisy latents and timesteps. This retains a common distribution reference without sample-prototype selection or extra conditioning operations during denoising.

Proto-MMD updates shared prototypes and view encoders, changing clean representations and their noisy training distributions. After warm-up, diffusion loss also backpropagates through the same encoders. Prototype information thus influences denoising through the learned latent space, placing cross-view distribution constraints directly on the final clustering representations.

Q2. Complexity, runtime, GPU memory, and scalability

Thank you for raising these concerns. We provide a complexity analysis and measured runtime and GPU memory usage.

Let BB, dd, and KK denote batch size, latent dimension, and prototype count. With fixed network architectures and a fixed number of kernel scales, per-batch InfoNCE costs O(B2d)O(B^2d) and Proto-MMD costs O((B+K)2d)O((B+K)^2d). Encoding, reconstruction, and noise prediction scale linearly with batch size. Diffusion training samples one timestep per latent representation and performs one noise prediction. With fixed batch size, model configuration, and epochs, training costs O(Nc)O(N_c), where NcN_c is the number of complete pairs. Clustering costs O(N)O(N) with fixed model, cluster count, and K-means iterations, where NN is the total sample count. Since complete pairs form a subset of all samples, overall time complexity under these fixed settings is O(N)O(N).

We performed five runs on Hdigit (MR=0.5) and LandUse-21 (MR=0.7), measuring total training-plus-evaluation time and peak GPU allocated and reserved memory. Values are means ± sample standard deviations.

SettingTime (s)Peak allocated (MiB)Peak reserved (MiB)
Hdigit, MR=0.5340.20 ± 17.631318.46 ± 0.911422.00 ± 0.00
LandUse-21, MR=0.756.86 ± 1.14226.1160 ± 0.0055272.00 ± 0.00

These measurements quantify resource requirements under the tested settings and show stable memory usage across runs. Table 1 already includes Hdigit (10,000 samples) and NoisyMNIST (50,000 samples), relatively large benchmarks in this field. Following your suggestion, we will evaluate performance at larger sample sizes in future work.

Analyses already in the submitted manuscript. Parameter sensitivity is presented in Fig. 3 and Sec. 3.3.1; convergence in Fig. 5 and Sec. 3.3.3; and feature visualization in Fig. 4 and Sec. 3.3.2. Thus, all three analyses mentioned in the review are already provided and discussed.

Q3. Variation across runs

Thank you for highlighting run-to-run variation. The original main table reports mean results. Related work also uses single-value main-table entries, e.g., Incomplete Multi-view Clustering via Hierarchical Semantic Alignment and Cooperative Completion (NeurIPS 2025). To address your concern, we performed five additional runs on each of Hdigit (MR=0.5) and LandUse-21 (MR=0.7), reporting means ± sample standard deviations.

SettingACC (%)NMI (%)ARI (%)
Hdigit, MR=0.597.61 ± 0.1393.40 ± 0.2894.75 ± 0.28
LandUse-21, MR=0.727.27 ± 0.6528.89 ± 0.8112.36 ± 0.16

These results quantify run-to-run variation under both settings.

Dear Reviewer pwBF ,

Thank you for your time and effort in reviewing our paper. Your comments have been valuable in helping us improve its quality.

As the author–reviewer discussion period is scheduled to end in approximately two days, we kindly encourage you to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards, The Authors

Thank you for providing the additional experiments and analyses. However, the authors' current response does not directly address several key issues raised in my Weaknesses, particularly the methodological novelty and the effectiveness of prototype alignment. Therefore, although I acknowledge the improvements made by the authors, I believe that the paper still has shortcomings.

Thank you for your further assessment. We would like to further elaborate on these two aspects.

1. Methodological Novelty

The methodological contribution of DPD lies in their complementary integration into a multi-level representation learning framework for incomplete multi-view clustering.

Specifically, paired-sample InfoNCE establishes instance-level correspondence between available cross-view pairs, whereas Proto-MMD introduces shared learnable prototypes as global distributional anchors, aligning view-specific latent distributions toward a common prototype distribution. This provides a complementary global constraint beyond pairwise instance alignment. Furthermore, random-timestep latent diffusion regularizes local variations in the same prototype-aligned latent space through noise prediction. Importantly, the prototypes are not explicitly injected into the denoising networks; instead, prototype alignment and diffusion regularization jointly optimize the shared latent representations through the encoders. At the clustering stage, the same prototypes are further employed to enhance available-view representations through prototype cross-attention and mask-aware aggregation.

Therefore, the contribution is not simply the combination of independent modules, but a multi-level learning strategy that integrates instance-level correspondence, prototype-guided global distribution alignment, and latent distribution regularization, followed by prototype-aware fusion for incomplete-view representation construction.

2. Effectiveness of Prototype Alignment

We further clarify the specific role of Proto-MMD. Unlike instance-level InfoNCE, which relies on paired observations to establish sample-wise correspondence, Proto-MMD aligns the latent distributions of different views toward a shared learnable prototype distribution. It thus provides a global structural constraint that complements instance-level supervision.

To isolate its contribution, our ablation study removes only the Proto-MMD objective while retaining the learnable prototypes and prototype-based fusion module. On Hdigit at a missing rate of 0.5, removing Proto-MMD decreases ACC/NMI/ARI from 97.71/93.68/94.97 to 95.78/90.39/90.88, respectively. This corresponds to consistent improvements of 1.93, 3.29, and 4.09 percentage points when prototype alignment is incorporated.

Furthermore, our additional factorized ablation on Hdigit and LandUse-21 demonstrates that Proto-MMD consistently improves clustering performance under different denoising configurations. For example, when random-timestep diffusion is employed, incorporating Proto-MMD increases ACC from 95.78% to 97.71% on Hdigit and from 24.41% to 27.90% on LandUse-21.

These results demonstrate that prototype alignment provides a measurable and complementary contribution to the overall framework. We have further clarified the distinction between instance-level alignment and distribution-level prototype alignment, as well as the interaction between prototype alignment and latent diffusion regularization, in our revised explanation.

We sincerely appreciate the reviewer's comments, which have helped us better articulate the methodological motivation and the individual contributions of the proposed framework.

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

This paper proposes Discriminative Prototype Diffusion (DPD) for incomplete multi-view clustering, which combines view-specific reconstruction, paired-sample contrastive learning, prototype-guided MMD alignment, and latent diffusion regularization. After that, a mask-aware fusion module is used to construct missing views based on the observed views. Extensive experiments on five datasets at various missing rates show competitive performance, with comprehensive ablation studies.

Critical strengths
  1. The paper is logically organized and addresses an important problem in multi-view clustering, i.e., part of the views are missing.

  2. Comparisons with existing baselines across five datasets show the competitive performance, with ablation studies support the contribution of each objective.

Critical weaknesses
  1. My major concern is the clarification of the prototype attention module, i.e., the details need clarification. Specifically, it is unclear how to get the prototypes and whether these prototypes are normalized or not before computing attention.

  2. I believe that the details for the diffusion configuration are underspecified, e.g., the parameters TT, βt\beta_t, e0e_0, the denoiser architecture, and the timestep sampling. I think that the authors should provide these details.

  3. Besides, the authors should report the computational overhead to support the claim of a lightweight regularizer.

  4. The paper does not cite or discuss two closely related works, Incomplete Multi-view Clustering via Prototype-based Imputation, IJCAI 2023, and Community-aware Multi-view Representation Learning with Incomplete Information, TPAMI 2026. Both of them are closely relevant to prototype-based representation learning with incomplete multi-view information.

Questions for authors

See the weaknesses.

2 comments

Q1. Prototype construction and normalization

Thank you for requesting clarification. The shared prototypes form a learnable matrix P with K vectors of latent dimension d. They are randomly initialized and jointly optimized with the view encoders through Proto-MMD (Eqs. 9-12). The same learned matrix is used for both views and for prototype attention at clustering time.

In Eq. 17, attention is computed as A = softmax(z P^T), with softmax applied over the K prototypes. Neither P nor z receives an additional L2 normalization before this attention computation. Softmax normalizes attention weights, not prototype vectors. The context AP is added to z (Eqs. 18-19); observed-view representations are then aggregated using the binary mask, and the resulting unified embedding is L2-normalized (Eqs. 21-22). Thus, "spherical fusion" refers to normalization of the final fused embedding. The normalization used for InfoNCE in Eq. 6 is a separate operation.

Q2. Diffusion configuration and denoiser

Thank you for pointing out the missing implementation details. Each view has its own noise-prediction MLP with two hidden layers: (d+1) -> H -> H -> d, with ReLU after the first two linear layers. A normalized scalar timestep is concatenated with the noisy d-dimensional latent; the output predicts d-dimensional Gaussian noise.

For each training sample, a timestep is sampled uniformly from the T discrete noise levels, independently of the standard Gaussian noise. The noisy latent is constructed directly using Eq. 13, followed by one denoiser evaluation and the noise-prediction MSE in Eq. 15. Noise variance follows a linear schedule. Diffusion is activated at epoch e0, as specified by Eq. 24; afterward its gradients also update the encoders.

The original experimental configuration is: T=[T], beta_start=[beta_start], beta_end=[beta_end], e0=[e0], and hidden width H=[H].

Q3. Computational overhead

Thank you for requesting quantitative evidence. The diffusion term operates in the low-dimensional latent space. Each sample uses one sampled noise level and one denoiser evaluation per view, without executing a T-step reverse chain. For batch size B and hidden width H, its additional per-batch computation is O(B(dH+H^2)), up to the constant factor for two views. At clustering time, representations are constructed with the encoders and prototype fusion; the denoisers are not invoked.

We measured total training-plus-evaluation time and peak GPU memory over five runs. The table reports full-model means ± sample standard deviations.

SettingTime (s)Peak allocated (MiB)Peak reserved (MiB)
Hdigit, MR=0.5340.20 ± 17.631318.46 ± 0.911422.00 ± 0.00
LandUse-21, MR=0.756.86 ± 1.14226.1160 ± 0.0055272.00 ± 0.00

These measurements document end-to-end resource requirements and stable memory usage under the two tested settings. The structural basis for describing the regularizer as lightweight is its latent-space MLP, one randomly sampled timestep per sample, and absence of denoiser computation during clustering.

Q4. Relation to ProImp and CAMERA

Thank you for pointing out these two related works. We acknowledge their relevance and have cited both in the revised manuscript. We provide a comparison below.

ProImp learns view-specific prototypes and transfers sample-prototype relationships from the observed view to the missing view to impute missing representations. CAMERA employs bidirectional sample-community attention and community commonality and versatility objectives to support missing-sample imputation and view alignment.

DPD learns a shared prototype distribution and aligns both view distributions to this common reference through Proto-MMD, while jointly regularizing the same latent space through view-specific noise prediction. Its prototype attention enhances observed representations at clustering time, and the mask excludes missing views from aggregation. The differences therefore lie in the shared distributional role of prototypes, their joint optimization with latent denoising, and the observed-view fusion strategy.

Thank you for the detailed response. The clarifications have addressed my concerns, and I am raising my rating from 3 to 4.

Reviewer_QyVwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the issues of insufficient paired samples and cross-view semantic inconsistency in incomplete multi-view clustering, and proposes the DPD framework. The method combines paired contrastive learning, shared prototype MMD alignment, and latent diffusion regularization. It further constructs a unified representation through prototype enhancement and mask-based fusion, followed by K-means clustering. Experiments are conducted on five benchmark datasets with three missing ratios, comparing with ten methods and evaluating by ACC, NMI, and ARI. The model is also verified through ablation studies, parameter analysis, visualization, and convergence curves. The results show that DPD achieves competitive performance in most settings.

Critical strengths

1.The method design is of certain value. The paper combines shared prototype MMD alignment, instance-level contrastive learning, and latent diffusion regularization to constrain multi-view representations from both distribution-level and instance-level perspectives, offering a meaningful exploration for incomplete multi-view clustering.

2.The experimental results are competitive. Under five datasets and three missing ratios, DPD outperforms the baselines in most settings. For example, at MR = 0.5, the ACC on Scene-15 and Hdigit exceeds the best baseline by 5.41 and 1.58 percentage points, respectively, providing empirical support for the effectiveness of the method.

Critical weaknesses

1.Insufficient utilization of observed information undermines the consistency between the research motivation and the methodological design. Section 2.1 states that the model is trained only on the complete paired subset Dc, yet the reconstruction, prototype MMD, and diffusion losses can all be computed from single-view samples. The paper does not explain why these observations are excluded, nor does it provide relevant control experiments. Consequently, its claim of alleviating the insufficiency of paired samples, especially under high missing ratios, lacks sufficient support.

2.The evidence for the contribution of the diffusion module is insufficient. In Table 2, removing the diffusion module reduces ACC by only 1.93 percentage points, whereas removing the prototype MMD leads to a substantial performance drop. The diffusion ablation covers only one dataset and one missing ratio, and lacks both variance reporting and comparisons with alternative approaches such as plain denoising and consistency regularization. This makes it difficult to demonstrate the necessity of the diffusion design and its synergy with prototype alignment.

Questions for authors
  1. Use of incomplete samples: Section 2.1 says training uses only the paired subset Dc. Do incomplete samples contribute to any loss? Since reconstruction, Proto-MMD, and diffusion can use observed views independently, please explain this restriction and report a comparison using all observed views for these losses while reserving paired samples for InfoNCE, across missing rates.

2.Diffusion contribution: On Hdigit at MR=0.5, removing diffusion lowers ACC by 1.93 points, while removing Proto-MMD causes a much larger drop. How are prototypes handled in the Proto-MMD ablation, given their role in final fusion? Results across more datasets and missing rates, with run-to-run variation and comparisons to simpler denoising or consistency regularization, would clarify diffusion’s independent contribution and its interaction with prototype alignment.

2 comments

Q1. Use of incomplete samples.

Thank you for noting that reconstruction, Proto-MMD, and diffusion can use observed views separately. In the submitted model, all four losses share complete-pair minibatches (Sec. 2.1). Single-view samples contribute at clustering time through mask-aware fusion (Sec. 2.6).

Paired batches coordinate instance alignment and view-wise learning. Both views represent the same objects and receive all four constraints. Their matched sample composition also helps Proto-MMD focus distribution alignment on view-specific representation differences.

Following your suggestion, we tested all-observed training on Hdigit at three missing rates. InfoNCE uses complete pairs, while the other losses use every observed view. Below, P denotes paired-only results from Table 1 and A denotes the new all-observed means. Scores are percentages.

MRP ACCP NMIP ARIA ACCA NMIA ARI
0.398.2495.6096.1790.6892.3788.20
0.597.7193.6894.9785.5987.0581.69
0.794.0886.0587.3076.7173.0465.59

All-observed training scores lower on every metric, with ACC gaps of 7.56, 12.12, and 17.37 points. A possible explanation is the changed balance among objectives. Extra single-view observations receive reconstruction, distribution alignment, and denoising, but lack same-instance cross-view positive-pair constraints. Averaging reconstruction and denoising over the enlarged sets reduces paired samples' relative contributions, while InfoNCE still uses paired samples. Finite samples and random masks can also yield different sample compositions across views, adding variability to shared-prototype learning.

As missingness increases, more samples contribute only to view-wise losses, widening this difference in objective coverage. This agrees with the increasing performance gaps. Under the current loss weights and training strategy, paired batches perform better. Using additional single-view observations requires coordinating view-wise learning with instance alignment. Shared prototypes supplement limited paired supervision with distribution-level constraints, and this comparison provides empirical support for the paired training protocol.

Q2. Diffusion contribution and prototype handling.

We apologize for the misplaced ablation labels in Table 2 and Sec. 3.2.2. The correct Hdigit ACC at MR=0.5 is 97.71% for the full model, 91.82% without RSDM (a 5.89-point drop), 95.78% without Proto-MMD (a 1.93-point drop), and 32.95% without InfoNCE. We have corrected both locations in the revision. All core conclusions remain unchanged.

For Proto-MMD ablation, we remove only the prototype alignment loss, so latent embeddings are no longer explicitly aligned to shared prototypes. To preserve architectural comparability, learnable prototypes and prototype-based fusion remain intact and are updated normally with the remaining objectives.

We further compare random-timestep diffusion (RSDM) with fixed-timestep diffusion (FSDM) on Hdigit (MR=0.5) and LandUse-21 (MR=0.7). Both use identical denoisers and noise-prediction objectives, while FSDM trains at one fixed timestep. All rows retain reconstruction and InfoNCE. Values below are ACC/NMI/ARI (%).

DenoisingProto-MMDHdigit, MR=0.5LandUse-21, MR=0.7
NoneOff83.68/85.66/79.7523.85/25.97/11.61
FSDMOff84.76/87.33/80.9322.38/25.69/10.75
RSDMOff95.78/90.39/90.8824.41/26.89/10.59
NoneOn91.82/88.07/84.8426.01/28.71/11.92
FSDMOn91.31/91.04/88.6725.29/27.97/11.45
RSDMOn97.71/93.68/94.9727.90/29.76/12.51

RSDM significantly outperforms FSDM and no denoising on both datasets. With Proto-MMD, its ACC gains over FSDM and no denoising are 6.40 and 5.89 points on Hdigit, and 2.61 and 1.89 points on LandUse-21. FSDM scores below no denoising in both settings, demonstrating that a single noise scale cannot handle the complex noise induced by view missingness, whereas random timestep sampling enables effective multi-scale denoising. Comparisons with and without Proto-MMD also confirm their synergy. RSDM smooths the latent manifold and prevents missingness-induced noise from corrupting prototype estimation, while accurate shared prototypes anchor denoised features into compact, semantically aligned clusters.

Dear Reviewer QyVw ,

Thank you for your time and effort in reviewing our paper. Your comments have been valuable in helping us improve its quality.

As the author–reviewer discussion period is scheduled to end in approximately two days, we kindly encourage you to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.