acceptodds
Under review as a conference paper at ICLR 2027

Decoupled Conditional Alignment for Partially View-Aligned Clustering

Abstract

Partially View-aligned Clustering (PVC) aims to handle complex multi-view data where only a small subset of samples exhibits strict cross-view correspondence. However, existing methods often overlook intra-view geometric topology during the feature extraction process. Furthermore, blindly aligning massive unpaired samples risks false-negative repulsion, leading to severe semantic confusion and feature manifold collapse. To address these critical issues, we propose a divide-and-conquer framework termed Decoupled Conditional Alignment (DeCA). DeCA first employs a dual-decoder Graph Autoencoder (GAE) to preserve the native intra-view neighborhood connectivity for extracting structure-preserved latent representations. At the sample level, it constructs stable global semantic anchors via strong cross-view contrast on aligned samples. For unaligned samples, pseudo-label-guided conditional Maximum Mean Discrepancy (MMD) and intra-view neighborhood contrast ensure intra-class sub-distribution matching without cross-view pairings. Finally, a decoupled feature fusion strategy is introduced to mitigate manifold contamination caused by superimposing heterogeneous misaligned features. Extensive experiments on ten benchmark datasets demonstrate DeCA's superiority over various PVC baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
cross-view correspondences, geometric topologies, semantic confusion, manifold collapse

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zW2VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DeCA for partially view-aligned Clustering (PVC), aiming to address the neglect of intra-view geometric topology in existing methods and the potential false-negative repulsion caused by aligning unpaired samples. To preserve intra-view structural information, DeCA employs a dual-decoder Graph Autoencoder. For aligned samples, DeCA employs cross-view contrastive learning to establish semantic anchors. For unaligned samples, it combines intra-view neighborhood contrastive learning with pseudo-label-guided conditional MMD to achieve cross-view distribution alignment without explicit sample correspondences. Finally, a decoupled feature fusion strategy is introduced to prevent the direct fusion of mismatched representations. The reported experimental results on ten benchmark datasets show competitive clustering performance against eight existing PVC baselines.

Critical strengths
  1. The consideration of intra-view structural information is a reasonable design choice. DeCA employs a dual-decoder GAE to reconstruct both feature attributes and graph topology, incorporating intra-view structural information into representation learning.

  2. The experimental evaluation covers a relatively broad range of settings. The paper includes ten multi-view datasets with varying scales and eight representative PVC baselines, providing a comprehensive performance comparison.

Critical weaknesses
  1. The paper exhibits serious issues in scientific rigor and academic presentation. The first sentence of the Introduction incorrectly claims that unsupervised representation learning heavily relies on multi-view clustering, misrepresenting the relationship between these research areas. Additionally, Sections 2.2 and 2.3 have identical titles and substantially overlapping content. These issues raise concerns about the rigor of the paper.

  2. The experimental results contain multiple repetitions, raising serious concerns about their reliability. In Table 4, the "w/ MLP" and "w/o Decoupled Fusion" variants report identical results across all six entries. Furthermore, in Tables 2 and 3, three baselines (BIN, DWInfoNCE, and RMCNC) report exactly the same ACC, NMI, and ARI values on Fashion under different unaligned rates. These repetitions across different experimental settings raise serious concerns about the reliability of the reported experimental evidence.

Questions for authors
  1. Why do the "w/ MLP" and "w/o Decoupled Fusion" variants in Table 4 report identical results across all six entries? Furthermore, why do BIN, DWInfoNCE, and RMCNC report exactly the same ACC, NMI, and ARI values on Fashion under different unaligned rates in Tables 2 and 3? Explain these repetitions and verify the reported results.

  2. What is the scientific basis for the claim that unsupervised representation learning heavily relies on multi-view clustering in the opening sentence of the Introduction? Why do Sections 2.2 and 2.3 have identical titles and substantially overlapping content?

  3. In Table 4, simultaneously removing (L_{\mathrm{cscl}}) and (L_{\mathrm{lpca}}) yields higher ACC than removing either loss individually on both datasets. In particular, on NoisyMNIST-tune, all three metrics (ACC, NMI, and ARI) are higher when both losses are removed. Please explain the reason for this phenomenon.

  4. Section 3.4 states that the template view is selected based on clustering silhouette quality, whereas Section 4.3 specifies that the first view is used as the template view. Please explain this inconsistency and analyze how different template-view selection strategies affect clustering performance.

  5. How are the global cluster centers used for pseudo-label generation initialized and updated during joint training? The authors should explain how inaccurate pseudo-labels and empty pseudo-classes are handled.

Reviewer_Y1c8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

DeCA addresses Partially View-aligned Clustering (PVC) through a decoupled framework. It uses a GAE to preserve intra-view topologies, applies cross-view contrast on aligned samples for global anchoring, and utilizes conditional MMD for unaligned samples to match distributions without explicit pairings, concluding with a decoupled feature fusion strategy.

Critical strengths
  1. Decoupling the alignment of paired and unpaired samples to prevent "false-negative repulsion" in PVC scenarios is a logically sound and inspiring divide-and-conquer strategy.

  2. Using a GAE preserves intra-view topologies, effectively mitigating the single-view manifold disruption typical of existing deep PVC methods.

Critical weaknesses
  1. The method lacks core innovation, acting as a mere amalgamation of existing techniques (GAE, MMD). Additionally, it introduces numerous hyperparameters (e.g., k,λa,τsk, \lambda_a, \tau_s), making systematic tuning in label-free unsupervised scenarios highly impractical.

  2. Assigning pseudo-labels via aligned cluster centers for conditional MMD risks severe error accumulation if low alignment ratios produce noisy labels.

  3. Constructing full-sample KNN graphs imposes a severe O(VN2)\mathcal{O}(VN^2) quadratic time and space complexity. This prohibits scalability to large datasets like Cifar10.

  4. The claim of "highly competitive efficiency" contradicts the experimental results in Figure 5, which shows DeCA is noticeably slower than baselines like FUMC and DWInfoNCE on multiple datasets, undermining the efficiency claims.

Questions for authors
  1. Since conditional MMD pseudo-labels depend on aligned cluster centers, how severely does DeCA degrade at extreme low-alignment ratios (e.g., 5-10%)?

  2. Since Figure 5 shows DeCA is significantly slower than FUMC and DWInfoNCE, what is the objective evidence supporting the appendix's claim of "highly competitive efficiency"?

  3. With the GAE's O(VN2)\mathcal{O}(VN^2) spatial complexity, what is the model's practical sample limit on a standard single GPU?

  4. Given the critical role of λa\lambda_a and λl\lambda_l in the objective function, how should users systematically tune these hyperparameters in real-world unsupervised scenarios lacking validation labels?

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

This paper proposes DeCA for partially view-aligned clustering. DeCA uses a decoupled fusion strategy to retain the view representation for unaligned samples.

Critical strengths
  1. This paper addresses a practical problem in multi-view clustering with incomplete cross-view correspondence.
  2. The division between aligned and unaligned samples is conceptually appropriate.
Critical weaknesses
  1. The novelty is relatively weak. GAE, contrastive learning, and MMD has been used extensively in existing works. DeCA seems to be a straightforward combination of current modules.
  2. Existing graph-based multi-view clustering methods has widely adopted graph neural networks to capture the intra-viewgeometric topology, which weakens the first contribution listed in the introduction section.
  3. The method description and implementation details are inconsistent regarding template-view selection. Section 3.4 describes adaptive selection based on clustering silhouette, whereas the implementation Section 4.3 states that the first view is fixed as the template.
  4. Results are reported without standard deviations, which is necessary because training, sample shuffling, and K-means are stochastic.
  5. Lack of sensitivity analysis for key hyperparameters.
Questions for authors

Please see the weakness list.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.