acceptodds
Under review as a conference paper at ICLR 2027

Pseudo-Correspondence Guided Blind-Alignment Learning for Completely Unaligned Multi-View Clustering

Abstract

Multi-view clustering aims to exploit complementary and consistent information from heterogeneous views to discover a shared clustering structure without supervision. However, when cross-view instance correspondences are entirely unavailable, completely unaligned multi-view clustering faces three key challenges: unreliable local structures may hinder early correspondence discovery; correspondence estimation and representation learning are mutually dependent; and exact instance matching alone is insufficient to capture cluster-level semantics. To address these challenges, we propose Pseudo-Correspondence Guided Blind-Alignment Learning (PCGBAL), which treats cross-view alignment as an evolving latent-supervision process coupled with representation learning. PCGBAL first structurally bootstraps view-specific representations through masked topological reconstruction and then progressively discovers high-confidence pseudo-correspondences from the evolving representation space. These associations are fed back as latent supervision to refine heterogeneous representations, forming a mutually reinforcing correspondence–representation process. To further bridge the gap between instance-level correspondence and clustering semantics, instance-to-cluster contrastive alignment transfers cluster-level information across views beyond exact instance matches. Extensive experiments on eight benchmark datasets against seven competing methods demonstrate the effectiveness and robustness of PCGBAL in completely unaligned settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, cross-view instance, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eTtTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PCGBAL, a framework for completely unaligned multi-view clustering that couples pseudo-correspondence discovery with representation learning. Its main contribution is to use evolving cross-view matches as supervision for both instance-level and cluster-level learning. Masked graph reconstruction learns view-specific structural representations; reciprocal-neighbor matching and self-paced selection identify pseudo-correspondences; and feature synchronization with prototype-based instance-to-cluster alignment refines the representations used for subsequent matching. For final clustering, auxiliary-view representations are softly aligned to the template view, concatenated, and processed by k-means. Experiments on eight benchmark datasets against seven competing methods report competitive clustering performance, with component ablations providing further support for the design.

Critical strengths

The paper is well organized and reports strong clustering performance. The motivation, method, and experimental evaluation follow a clear logical progression, making the overall framework easy to understand. In the reported comparisons with seven state-of-the-art methods, PCGBAL achieves the highest ACC on all eight datasets. Component ablations across the eight datasets further support the main design choices, while the template-view and hyperparameter analyses provide useful information about the method’s behavior.

Critical weaknesses

The paper reports average results without showing the variability across runs. Section 4.2 states that each experiment is repeated ten times. Please include standard deviations or confidence intervals in the result tables to assess stability, particularly when improvements over baselines are small.

Questions for authors
  1. Why is the within-view graph neighborhood size fixed at k=5 across all datasets? Appendix E.2 studies κ for cross-view matching, but this differs from the k used to construct the MTMM graphs. Please explain the rationale for k=5 and whether clustering performance is sensitive to this choice.

  2. Could the cluster-balance regularizer harm clustering when the selected pairs are correct but unevenly distributed across classes? Self-paced selection may select more pairs from easier-to-match classes, so correct matching does not imply a uniform class distribution. Equation (8) still pushes the average cluster probabilities toward a uniform distribution. Please explain whether this could interfere with correct predictions and why this target is appropriate.

  3. Could the authors provide quantitative metrics to directly evaluate the accuracy of the discovered cross-view correspondences? Figure 2 presents clustering performance at different unaligned rates, but the final clustering results do not directly show whether these correspondences are accurate. Please report additional metrics, such as pairing precision or the proportion of same-class pairs, to better substantiate the reliability of the discovered correspondences.

1 comment
AuthorsOpenReview ↗

Response to Reviewer eTtT

Reviewer 3 (eTtT)

We thank the reviewer for the constructive comments.

W1. Mean and Standard Deviation of Experimental Results over Ten Runs

Each experiment was repeated ten times, and ACC, NMI, and Purity are reported as mean ±\pm standard deviation (%). Representative results on Handwritten are shown below.

MethodACC (%)NMI (%)Purity (%)
DAGF63.68 ±\pm 0.2964.21 ±\pm 0.2363.68 ±\pm 0.29
FUMC13.11 ±\pm 0.400.91 ±\pm 0.0813.40 ±\pm 0.25
BIN80.15 ±\pm 2.2269.10 ±\pm 0.8380.55 ±\pm 1.48
VITAL67.48 ±\pm 1.8055.77 ±\pm 0.9767.85 ±\pm 0.63
CANDY62.78 ±\pm 3.2159.29 ±\pm 2.8765.94 ±\pm 2.77
KNNDA42.78 ±\pm 1.5739.56 ±\pm 1.4745.53 ±\pm 2.85
AFUMC84.75 ±\pm 1.5874.80 ±\pm 0.8184.66 ±\pm 1.88
Ours93.14 ±\pm 0.1987.09 ±\pm 0.4993.14 ±\pm 0.19

The main text reports mean values, while the revised appendix provides means and standard deviations for all metrics across all datasets.

Q1. Rationale and Sensitivity of the Within-View Neighborhood Size

The parameter kk denotes the number of nearest neighbors used to construct each within-view kk-NN graph. We fix k=5k=5 to preserve local structure while limiting noisy edges. To validate this choice, we conduct a sensitivity analysis of k∈{3,5,10,15,20}k\in\{3,5,10,15,20\} on seven completely unaligned datasets. Representative results (%) on Handwritten and UCI3view are shown below.

DatasetMetrick=3k=3k=5k=5k=10k=10k=15k=15k=20k=20
HandwrittenACC92.1093.2093.0593.1092.15
NMI85.4287.0287.3887.3486.11
Purity92.1093.2093.0593.1092.15
UCI3viewACC84.4093.0587.1591.1589.40
NMI77.5287.2582.4184.8583.62
Purity84.4093.0587.1591.1589.40

The results show that k=5k=5 achieves the highest ACC and Purity with competitive NMI, supporting our fixed setting. The analysis has been added to the revised manuscript.

Q2. Cluster-Balance Regularization under Class Imbalance

The regularizer βKL⁡(pˉu∥uK)\beta\operatorname{KL}(\bar{\mathbf p}^{u}\|\mathbf u_K) acts as a soft constraint to prevent cluster collapse rather than enforcing equal class proportions. Since pˉu\bar{\mathbf p}^{u} is computed over all samples, unevenly selected pseudo-correspondences do not directly determine the regularization term. However, the uniform prior may affect clustering when the true class distribution is imbalanced.

To investigate this effect, we construct balanced (80,80,80,80,80)(80,80,80,80,80) and imbalanced (200,100,50,25,25)(200,100,50,25,25) subsets from Hdigit, each containing 400 samples. We vary β\beta while keeping other settings fixed. The comparison results (%) are shown below.

β\betaBalanced ACCBalanced NMIImbalanced ACCImbalanced NMI
079.0068.1366.0056.16
0.0179.0068.1348.7554.57
0.179.2568.5947.5049.83
162.0059.0269.2559.15
1088.0076.6647.0045.86

The results indicate that appropriately weighted cluster balancing can remain effective under unequal class proportions, although its effectiveness depends on the regularization strength. The experiment and analysis have been added to the revised ablation section.

Q3. Quantitative Evaluation of Cross-View Correspondences

We evaluate active cross-view candidates on the completely unaligned Handwritten dataset. A candidate is semantically correct when both samples belong to the same ground-truth class. Let P\mathcal P be the fixed candidate pool, S⊆P\mathcal S\subseteq\mathcal P its same-class candidates, and Ac\mathcal A_c the active candidates at epoch cc. Semantic recall and coverage are defined as SR⁡c=∣Ac∩S∣∣S∣\operatorname{SR}_c=\frac{|\mathcal A_c\cap\mathcal S|}{|\mathcal S|} and Coverage⁡c=∣Ac∣∣P∣\operatorname{Coverage}_c=\frac{|\mathcal A_c|}{|\mathcal P|}, respectively. The results are shown below, with the first 100 epochs serving as warm-up.

Epoch101125150175200
SR (%)13.2826.9539.8450.0059.38
Coverage (%)10.1619.8630.0239.9550.12

As training progresses, more same-class relations are recovered. The corresponding analysis has been added to the revised ablation study.

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

This paper addresses the problem that existing multi-view clustering methods typically assume a known one-to-one correspondence between samples across views, whereas in real-world applications data from different views may be completely unaligned, making cross-view information fusion difficult and limiting clustering performance. To this end, the authors propose an unaligned multi-view clustering framework based on pseudo-correspondence learning, which mines latent semantic associations between samples from different views to construct pseudo-correspondences and uses them to guide cross-view representation learning and clustering optimization, thereby achieving semantic alignment and information fusion across views without relying on ground-truth sample pairing. Experiments on multiple public datasets demonstrate the effectiveness of the framework, but the accuracy of the pseudo-correspondences and their independent contribution to final clustering performance still require more thorough validation.

Critical strengths
  1. Without known cross-view sample correspondences, the paper proposes using continuously updated pseudo-correspondences to guide representation learning, addressing the interdependence between correspondence estimation and representation learning; the research problem is clearly defined.
  2. The paper is well organized, with a clear correspondence among research motivation, method design, and experimental results; overall readability meets the basic requirements of a machine learning conference paper.
  3. The overall method design is coherent: MTMM, DSPPM, CFS, and I2C-CA focus respectively on initial representation, instance matching, and category-level consistency, rather than being an unrelated stack of modules.
  4. The method achieves good clustering performance on eight datasets, with notable improvements on several datasets, indicating that it is indeed effective under the adopted experimental settings.
Critical weaknesses
  1. The entire paper rests on the core assumption that "high-quality pseudo-correspondences can progressively improve cross-view representations," yet it does not directly or sufficiently demonstrate how the accuracy of pseudo-correspondences improves during training.
  2. There are many ablation studies, but the most critical validation of correspondence quality is missing.
  3. The method is a combination of multiple existing techniques, and its core novelty is not sufficiently prominent.
  4. The authors assume by default that pseudo-correspondences selected from high-similarity samples are highly reliable, but this assumption is not adequately validated. For example, two different instances may belong to the same category and thus have very similar features without being true corresponding samples; if the model mismatches them and continuously pulls their representations closer, erroneous correspondences may be reinforced.
  5. It remains unclear whether the performance gain comes from dynamic coupling itself or from a stronger encoder, additional training objectives, or gradually incorporating more samples. The existing ablations help to some extent but are insufficient to fully disentangle these factors.
  6. The experiments claim "completely unaligned" on the one hand, but on the other hand, how is the correctness of the final clustering evaluation guaranteed?
Questions for authors
  1. Supplement the changes in pseudo-correspondence accuracy and coverage during training.
  2. Supplement a direct comparison between Progressive Coupling and ordinary alternating optimization.
  3. Robustness experiment of MTMM against erroneous KNN edges
1 comment
AuthorsOpenReview ↗

Response to Reviewer EuSS

Reviewer 2 (EuSS)

We thank the reviewer for the constructive comments.

W1, W2, and Q1. Evolution of Pseudo-Correspondence Quality

We evaluate active cross-view candidates on the completely unaligned Handwritten dataset. A candidate is semantically correct when both samples belong to the same ground-truth class. Let P\mathcal P be the fixed candidate pool, S⊆P\mathcal S\subseteq\mathcal P its same-class candidates, and Ac\mathcal A_c the active candidates at epoch cc. Semantic recall and coverage are defined as SR⁡c=∣Ac∩S∣∣S∣\operatorname{SR}_c=\frac{|\mathcal A_c\cap\mathcal S|}{|\mathcal S|} and Coverage⁡c=∣Ac∣∣P∣\operatorname{Coverage}_c=\frac{|\mathcal A_c|}{|\mathcal P|}, respectively. The results are shown below, with the first 100 epochs serving as warm-up.

Epoch101125150175200
SR (%)13.2826.9539.8450.0059.38
Coverage (%)10.1619.8630.0239.9550.12

As training progresses, more same-class relations are recovered. The corresponding analysis has been added to the revised ablation study.

W3. Core Novelty Beyond Existing Techniques

PCGBAL addresses completely unaligned multi-view clustering by transforming geometric candidates into cluster-level relations without assuming exact instance correspondences. Its key novelty lies in the mutual reinforcement of correspondence discovery and representation learning: progressively refined relations guide representation learning, which in turn improves correspondence discovery. Ablation studies and the Static Matching comparison further validate this design.

W4. Impact of High-Similarity Cross-Cluster Mismatches

PCGBAL does not assume that high-similarity candidate pairs represent exact instance matches. Instead of directly pulling paired instances together, it uses candidate pairs to establish cluster-level relationships and aligns individual samples with shared cluster prototypes. Therefore, different instances from the same class can still provide meaningful semantic supervision.

W5. Disentangling the Contributions of Individual Components

To isolate component contributions, we expand the ablation study: w/o MTMM tests the encoder; w/o CFS and w/o I2C test the objectives; w/o MP tests candidate construction; and w/o SPS tests progressive activation. Static Matching further isolates progressive relation updates.

Additionally, we report the semantic precision of selected pseudo-correspondences throughout training to directly examine their reliability and evolution. The results support the contributions of individual components and the interaction between correspondence discovery and representation learning.

The analyses have been added to the revised ablation section.

W6. Correctness of Clustering Evaluation under Complete Unalignment

We select the template view using the unsupervised Neighborhood Stability Coefficient (NSC) and align auxiliary representations toward it without labels or known correspondences. The aligned representations are then fused and clustered using KK-means. Clustering metrics such as ACC and NMI are computed by comparing the predicted labels with the template view's ground-truth labels, which are used only for evaluation.

Q2. Progressive Coupling versus Ordinary Optimization

We introduce a degraded variant, Static Matching, which fixes the selected votes after warm-up, whereas Progressive Coupling progressively expands and updates them. The results are shown below.

DatasetStatic ACC (%)Static NMI (%)Progressive ACC (%)Progressive NMI (%)
Handwritten90.9084.9593.1487.09
UCI3view78.1080.1293.0987.48
Scene1548.4948.0252.8048.92
MNIST-USPS57.5858.9271.7165.27

Progressive Coupling outperforms Static Matching in ACC and NMI across all four datasets, confirming the benefit of progressive updates. Complete ablation results are provided in the revised manuscript.

Q3. MTMM Robustness to Erroneous k-NN Edges

To evaluate this concern, we randomly replace a specified fraction of within-view kk-NN edges in every view and conduct experiments on the completely unaligned Handwritten dataset. The results are shown below.

Replaced-edge fraction0.10.20.30.40.5
Full ACC (%)91.8590.1587.5583.9082.70
w/o MTMM ACC (%)89.5086.0568.6564.5054.50

As graph corruption increases, the performance gap between the full model and w/o MTMM widens. This demonstrates MTMM's robustness to kk-NN edge corruption. The results have been added to the revised ablation section.

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

This paper proposes PCGBAL, a framework for fully unaligned multi-view clustering that integrates masked topology reconstruction, progressive pseudo-correspondence mining, and cluster-level contrastive learning to achieve cross-view alignment and representation learning. Its effectiveness is demonstrated on eight benchmark datasets.

Critical strengths

This paper proposes a unified framework for fully unaligned multi-view clustering, integrating masked topology reconstruction, progressive pseudo-correspondence mining, and cluster-level contrastive learning. The overall framework is well designed and achieves promising clustering performance on eight benchmark datasets. Ablation studies, robustness analyses, and parameter sensitivity experiments further demonstrate its effectiveness.

Critical weaknesses
  1. The proposed method mainly combines existing techniques, including masked graph reconstruction, bidirectional nearest-neighbor matching, self-paced learning, and prototype contrastive learning, without substantial novel algorithmic designs. Compared with existing fully unaligned multi-view clustering methods, its core contributions are not sufficiently clear.

  2. The paper emphasizes the mutual reinforcement between correspondence mining and representation learning. However, it does not report how pseudo-correspondence accuracy, recall, and incorrect matching rates evolve during training, making it difficult to directly validate this core mechanism.

  3. Different template views may lead to significant performance variations. However, the paper always selects the first view as the template, without providing a reliable adaptive template view selection mechanism.

  4. A full grid search over λ1,λ2,p,κ\lambda_1,\lambda_2,p,\kappa requires 432 configurations. The paper does not adequately explain how the optimal hyperparameters are selected in a fully unsupervised setting, raising concerns about experimental fairness and reproducibility.

  5. Both cross-view similarity computation and soft alignment involve N×NN \times N matrices, resulting in quadratic computational and memory complexity, which limits the scalability of the proposed method to large-scale datasets.

Questions for authors

Please address the concerns raised in the Critical Weaknesses. I will consider revising my score based on the authors' responses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer pdMo

Reviewer pdMo

We thank the reviewer for the constructive comments. Our point-by-point responses are provided below.

Q1. Core Algorithmic Novelty and Contributions

PCGBAL addresses completely unaligned multi-view clustering by transforming geometric candidates into cluster-level relations without assuming exact instance correspondences. Its key novelty is the mutual reinforcement between correspondence discovery and representation learning, where progressively refined relations guide representation learning and vice versa. Ablation studies and the Static Matching comparison further validate this design.

Q2. Training Pseudo-Correspondence Quality

We evaluate active cross-view candidates on the completely unaligned Handwritten dataset. A candidate is semantically correct when its two samples share a ground-truth class. Let P\mathcal P be the fixed candidate pool, S⊆P\mathcal S\subseteq\mathcal P its same-class candidates, and Ac\mathcal A_c the active candidates at epoch cc. Semantic recall and coverage are defined as SR⁡c=∣Ac∩S∣∣S∣\operatorname{SR}_c=\frac{|\mathcal A_c\cap\mathcal S|}{|\mathcal S|} and Coverage⁡c=∣Ac∣∣P∣\operatorname{Coverage}_c=\frac{|\mathcal A_c|}{|\mathcal P|}, respectively. The corresponding results at different training epochs are presented in the following table, with the first 100 epochs serving as the warm-up stage.

Epoch101125150175200
SR (%)13.2826.9539.8450.0059.38
Coverage (%)10.1619.8630.0239.9550.12

As training progresses, more same-class relations are recovered. The corresponding analysis has been added to the revised ablation study.

Q3. Adaptive Template-View Selection

We introduce an unsupervised strategy based on the Neighborhood Stability Coefficient (NSC). The procedure is summarized below.

Algorithm: Unsupervised Template View Selection

Input: Multi-view data {Xv}v=1V\{\mathbf X^v\}_{v=1}^{V}, neighborhood size ℓ=10\ell=10, perturbation rounds B=5B=5, retention probability ζ=0.8\zeta=0.8, maximum sample size M=1000M=1000.

Output: Template view index tt.

  1. Set n=min⁡(N,M)n=\min(N,M).

  2. For v=1,…,Vv=1,\ldots,V:

    • Evenly sample nn rows from Xv\mathbf X^v to form Yv\mathbf Y^v and apply row-wise ℓ2\ell_2 normalization.
    • Construct an ℓ\ell-NN graph Av\mathbf A^v with neighborhoods Niv\mathcal N_i^v; set sv=0s_v=0.
    • For b=1,…,Bb=1,\ldots,B:
      • Generate a mask Mbv\mathbf M_b^v with retention probability ζ\zeta; set Y~bv=(Yv⊙Mbv)/ζ\widetilde{\mathbf Y}_b^v=(\mathbf Y^v\odot\mathbf M_b^v)/\zeta.
      • Construct the perturbed ℓ\ell-NN graph and obtain its neighborhoods N~i,bv\widetilde{\mathcal N}^{v}_{i,b}.
      • Update sv←sv+1Bnℓ∑i=1n∣Niv∩N~i,bv∣s_v\gets s_v+\frac{1}{Bn\ell}\sum_{i=1}^{n}|\mathcal N_i^v\cap\widetilde{\mathcal N}_{i,b}^v|.
  3. Return t=arg⁡max⁡vsvt=\arg\max_v s_v.

To evaluate NSC, we compare all template views on five datasets. The table below reports representative results, with NSC marking the selected view and the best results in bold.

DatasetTemplate ViewACCNMI
HandwrittenView 192.0184.57
View 271.1062.87
View 387.6081.26
View 466.0564.63
View 5 (NSC)93.1487.09
View 672.8066.99
UCI3viewView 1 (NSC)93.0987.48
View 272.2565.20
View 366.4565.77
Scene15View 1 (NSC)52.8048.92
View 237.3538.07
View 332.2632.02
MNIST-USPSView 1 (NSC)71.7165.27
View 258.3056.07
HdigitView 1 (NSC)72.4067.65
View 259.7361.99

The results demonstrate the effectiveness of the proposed template-selection strategy. We have updated the appendix with the strategy and complete results across all six metrics.

Q4. Unsupervised Hyperparameter Selection

Parameter-sensitivity experiments on seven completely unaligned datasets show that λ1∈[10−4,10−3]\lambda_1\in[10^{-4},10^{-3}], λ2∈[10−4,10−2]\lambda_2\in[10^{-4},10^{-2}], p∈[0.3,0.4]p\in[0.3,0.4], and k=5k=5 generally yield competitive performance, providing guidance for new unlabeled datasets. This analysis has been added to the revised manuscript.

Q5. Computational Scalability

Despite its O(N2)O(N^2) time and space complexity, we evaluate the practical efficiency of all compared models on eight benchmark datasets using total training and evaluation time (seconds). Representative results on three datasets are shown below.

DatasetFUMCDAGFAFUMCVITALCANDYKNNDABINOurs
Handwritten3.40111.3774.5013.1062.18183.8935.8310.15
UCI3view0.3716.7420.274.2993.38236.3992.415.67
Scene153.3894.8151.157.63160.55711.62114.7114.62

PCGBAL achieves competitive runtime. We highlight anchor-based learning toward linear complexity in the conclusion and provide complete runtime comparisons in the revised manuscript.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your reviews. The authors have posted their responses and updated the paper. If you have not already done so, please take a moment to review their rebuttal and provide any follow-up comments before the discussion deadline.

Thank you for your time and participation.

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.