acceptodds
Under review as a conference paper at ICLR 2027

Recoupling Multi-Tensorized Boost Consistency Graph Learning for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) has grasped growing attention recently, which aims to address unsupervised clustering problems caused by missing views. While several method has been proposed, the current approaches still suffer from key shortcomings: Affinity graph learning methods based on Euclidean distances can be sensitive to feature magnitudes, leading to inadequately reflect informative relative response patterns. Moreover, existing methods recover missing view-specific information without considering cross-view diversity, resulting in suboptimal imputation. To overcome these challenges, we introduce a novel framework termed recoupling multi-tensorized boost consistency graph learning (RMTC-GL) for IMVC. Our approach first constructs correlation graphs using non-parametric Spearman correlation on complete samples, then recovers missing entries by preserving and leveraging all observed correlations via a masking strategy. To maintain view diversity, we incorporate a tensor-based diversity constraint that explicitly enforces heterogeneity among the recovered graphs across views. Additionally, we design a recomposed multi-tensor enhanced consensus learning module that jointly exploits cross-view complementarity for direct consensus graph recovery and captures high-order consistency between complete and incomplete correlation structures. Extensive experiments demonstrate that RMTC-GL achieves superior clustering accuracy compared to state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Graph learning, Multi-Tensorized Consistency

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_sibfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RMTC-GL for incomplete multi-view clustering. The method first constructs a Spearman correlation graph based on the feature rankings of observed samples, fitting observed associations and recovering missing ones through masking constraints. It then encourages diversity by imposing pairwise overlapping penalties on cross-view graph structures, and combines the recovered view-specific graphs with a shared consensus graph into a tensor, applying joint regularization via tensor nuclear norm. The authors solve the problem using alternating optimization and compare their method with nine others across six datasets and three missing rates, additionally providing ablation studies, parameter sensitivity analysis, and convergence analysis.

Critical strengths
  1. This paper addresses a promising research question in multi-view learning: when recovering missing information, one should not only leverage consistency across views but also avoid making all recovered views completely identical. This research motivation is well justified, and graph modeling provides a clear framework for discussing this issue.
  2. The experiments cover a broad range. Performance improvements on multiple datasets, along with ablation results for different components, indicate that the proposed framework deserves further investigation.
Critical weaknesses

1.Diversity constraints directly penalize the overlap of view graph structures, but it does not specify how to avoid weakening the consistency that is beneficial for clustering.
2.The additional contribution of tensor reorganization, compared with conventional low-rank regularization and consensus graph learning, is not clear enough. 3.In Table 2, the OIMC-IAGD on BBC4 and GUITAR on COIL show duplicate values, which need to be checked by the authors.
4.Some graph update formulas are inconsistent with the objective function, and it needs to be clarified which update method was actually implemented.

Questions for authors

1.Which differences does the diversity term aim to preserve in the view graph structures? How does it avoid weakening connections that correctly reflect the same clustering structure across different views?
2.What information can the proposed 'view-consensus graph' tensor construction capture that simple graph alignment or conventional low-rank tensor models cannot? Can this be verified experimentally on its own?
3.Please explain the repeated values in Table 2 and indicate whether, if there is an error in result compilation, correcting it would change the comparison conclusions between methods. Also, please clarify whether the graph updates in the actual implementation are consistent with the objective function in the paper.

2 comments
Area_Chair_MqDgOpenReview ↗

Please leave some comments for the authors!

Thank you for your thoughtful reviews. When you have a chance, please read the authors’ responses to your comments and update the discussion with your assessment. Please note whether your concerns have been addressed and briefly mention any remaining questions.

Reviewer_sibfOpenReview ↗

Raise my rating

My concerns have basically been addressed, and I decide to raise my rating to 4.

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

This paper addresses incomplete multi-view clustering (IMVC) by proposing RMTC-GL, a framework that integrates Spearman correlation-based graph construction, masked incomplete graph recovery, tensorial cross-view diversity embedding, and multi-tensor enhanced consensus graph learning into a unified model. The method constructs view-specific correlation graphs from complete samples, recovers missing entries while preserving observed structures, enforces cross-view diversity via a tensor-based constraint, and captures high-order consistency between recovered and consensus graphs through a recomposed multi-tensor nuclear norm. Experiments on six benchmark datasets under varying missing rates demonstrate competitive performance against nine state-of-the-art methods.

Critical strengths
  1. The motivation is well articulated: the paper identifies a genuine gap in IMVC, namely, the tension between recovering incomplete view-specific structures and preserving cross-view diversity while exploiting high-order consistency, and proposes a coherent pipeline that addresses each aspect.

  2. The empirical evaluation is reasonably thorough, covering six datasets, three missing rates, nine baselines, Nemenyi statistical tests, convergence analysis, parameter sensitivity, and ablation studies. The consistent margins over strong baselines (e.g., ASCR, CAL) across datasets are encouraging.

Critical weaknesses
  1. The theoretical justification for the tensorial diversity constraint (TGD) in Eq. (4) is underdeveloped. The equivalence between the element-wise product formulation and the trace formulation is stated without proof, and the claim that minimizing this term "penalizes high similarity" lacks formal grounding. Moreover, the interaction between the diversity term (weighted by γ) and the consensus alignment term (weighted by β) is only empirically tuned; no theoretical guidance is provided on how these two competing objectives should be balanced.

  2. The convergence analysis is purely empirical. The paper reports that objective values decrease monotonically in Fig. 3, but provides no convergence guarantee for the nonconvex multi-block alternating optimization scheme. Given that the subproblem for J^(v) involves tensor nuclear norm minimization and the Z^(v) update involves simplex projection, a discussion of whether the algorithm converges to a stationary point (or under what conditions) would strengthen the soundness.

  3. The ablation study (Table 3) is informative but incomplete. RMTC-E replaces Spearman correlation with Euclidean distance, yet the paper does not report the computational overhead of Spearman correlation construction relative to the Euclidean alternative. Since the complexity analysis includes O(Vn_v d_v log d_v + Vn_v² d_v) for correlation construction, a runtime comparison would help justify the choice beyond clustering accuracy.

Questions for authors
  1. The paper claims that the multi-tensor formulation avoids the need for SVDs of n × n matrices in the Fourier domain by stacking Z^(v) and S into a two-slice tensor H^(v). However, the t-SVD of a two-slice tensor still involves Fourier-domain operations. Could the authors clarify the precise computational savings relative to conventional full-graph tensor formulations, ideally with a concrete complexity comparison?

  2. The choice of the Spearman correlation for graph construction is motivated by its non-parametric nature and robustness to nonlinear relationships. How does the method handle ties in the Spearman rank correlation, and does the discretization of ranks affect the quality of the recovered graphs, particularly for datasets with many categorical or ordinal features (e.g., BBC4)?

  3. The optimization uses the inexact ALM with penalty parameters ρ and λ updated by a factor θ. The paper does not report the sensitivity of the method to θ or to the initial values of ρ and λ. Given that the convergence speed and final solution quality can depend on these hyperparameters, could the authors provide guidance or sensitivity analysis?

5 comments
AuthorsOpenReview ↗

Response to W1

Response to W1:

Thank you for pointing out that Eq. (4) requires a proof and a more precise interpretation. We will explicitly restrict the identity to the nonnegative recovered graph. For any two such graphs Z(v)\mathbf{Z}^{(v)} and Z(w)\mathbf{Z}^{(w)}, \begin{equation} \left|\mathbf{Z}^{(v)} \odot \mathbf{Z}^{(w)}\right|_{1}= \sum_{i,j}\mathbf{Z}_{ij}^{(v)}\mathbf{Z}_{ij}^{(w)} =\sum_i(\mathbf{Z}_{ij}^{(v)}\mathbf{Z}_{ij}^{(w)\top})_{ii}= \operatorname{Tr}\Big(\mathbf{Z}^{(v)}{\mathbf{Z}^{(w)}}^{\top}\Big) \end{equation} where ∥⋅∥1\|\cdot\|_1 denotes the entrywise ℓ1\ell_1 norm. For a fixed view vv, the derivative of the TGD penalty with respect to Zij(v)\mathbf{Z}_{ij}^{(v)} is \sum_\{w\ne v} \mathbf{Z}_{ij}^{(w)}. Thus, assigning additional weight to an edge incurs a larger marginal penalty when that edge already carries more weight in other views.

For the interaction between β\beta and γ\gamma, we analyze the sub-objective consisting of these two terms.

Theorem 1 (Consensus-diversity decomposition)

Let Z(v)∈D\mathbf Z^{(v)}\in\mathcal D and define $F({\mathbf Z^{(v)}},\mathbf S) = \beta \sum_{v=1}^{V} |\mathbf Z^{(v)}-\mathbf S |^2_F+\gamma\sum_{w>v} \langle\mathbf Z^{(v)},\mathbf Z^{(w)}\rangle$ For fixed {Z(v)}\{\mathbf Z^{(v)}\} and β>0\beta>0, the minimizer over S∈D\mathbf S\in\mathcal D is S⋆=1V∑v=1VZ(v)\mathbf S^\star = \frac1V\sum_{v=1}^{V}\mathbf Z^{(v)}. Furthermore, \begin{equation} F({\mathbf Z^{(v)}},\mathbf S^\star) =\left(\frac{\beta}{V}-\frac{\gamma}{2}\right) \sum_{w>v} |\mathbf Z^{(v)}-\mathbf Z^{(w)}|^2_F+ \frac{\gamma(V-1)}{2} \sum_{v=1}^{V}|\mathbf Z^{(v)}|^2_F. \end{equation} proof. The arithmetic mean of row-stochastic nonnegative matrices belongs to D\mathcal D. Differentiation with respect to S\mathbf S therefore gives the stated minimizer. Using the identities \begin{equation} \sum_v|\mathbf Z^{(v)}-\mathbf S^\star|^2_F= \frac1V\sum_{w>v} |\mathbf Z^{(v)}-\mathbf Z^{(w)}|^2_F \end{equation} and \begin{equation} \begin{aligned} \sum_{w>v}\langle\mathbf Z^{(v)},\mathbf Z^{(w)}\rangle= \frac{V-1}{2}\sum_v|\mathbf Z^{(v)}|^2_F -\frac12\sum_{w>v} |\mathbf Z^{(v)}-\mathbf Z^{(w)}|^2_F \end{aligned} \end{equation} establishes the result.

Theorem 1 has been added to the revised manuscript, which can be found on Page 4, Section 3.3. It reveals that the combined regularization can be decomposed into a pairwise disagreement term and a graph-norm term. This characterizes how β\beta and γ\gamma jointly regulate cross-view consistency and structural diversity. Thus, RMTC-GL seeks to retain view-specific complementary information without abandoning the shared clustering structure.

AuthorsOpenReview ↗

Response to W2, W3:

Response to W2:

Thank you for raising this point. We agree that the iteration curves provide only empirical evidence of numerical stabilization on the tested datasets and do not constitute a formal convergence guarantee. In the revised manuscript, we have added the discussion of convergence, which can be found at Page 7, Section 4. The optimization problem in Eq.~(5) is generally non-convex due to the bilinear diversity term coupling the view-specific graphs. In the proposed alternating optimization framework, the S\mathbf{S}-subproblem admits a closed-form solution, the Z(v)\mathbf{Z}^{(v)}-subproblem is solved by weighted simplex projection, and the E\mathbf{E}- and J(v)\mathcal{J}^{(v)}-subproblems are solved by soft thresholding and tensor singular value thresholding, respectively. With the multipliers and penalty parameters fixed, exact block updates do not increase the augmented Lagrangian during the primal update sweep. Thus, each block can be efficiently optimized or decreased when the other variables are fixed. Since the objective value is lower bounded, the alternating procedure is expected to reach a stable solution in practice.

Response to W3:

Thank you for this suggestion. Table 3 compares clustering accuracy, and we will add a runtime comparison to quantify the cost of using Spearman correlation. For each view vv, Spearman graph construction adds an O(nvdvlog⁡dv)O(n_v d_v\log d_v) feature-ranking step, while both Spearman and Euclidean graph construction require O(nv2dv)O(n_v^2d_v) pairwise computation. These bounds describe how the cost scales, while measured runtimes will show the practical overhead.

Table 1 Graph construction time and ACC gain of Spearman over Euclidean affinity construction at different missing rates.

DatasetSpearman (ms)Euclidean (ms)ACC Gain (pp)
10%30%50%10%30%50%10%30%50%
ALOI24.8821.7418.3422.9719.6915.04+4.15+13.55+13.94
Leaves50.4739.2828.7647.0039.1028.19+7.45+7.75+4.21
COIL713.97617.94529.06265.29223.28172.92+2.08+1.71+1.55

We benchmarked the construction of the pairwise Spearman correlation and Euclidean distance matrices using identical normalized inputs and missing-view masks. Table 1 reports graph construction time alongside the ACC gain of RMTC over RMTC-E.

On ALOI and Leaves, Spearman adds only 0.18–3.47 ms to graph construction across the tested missing rates, with ACC gains of 4.15–13.94 percentage points. On COIL, Spearman takes 2.69–3.06 times as long to construct the graph and yields smaller ACC gains of 1.55–2.08 percentage points. Thus, Spearman improves ACC in all nine settings, but the runtime–accuracy trade-off varies by dataset. The added cost is small on ALOI and Leaves, while COIL shows more construction cost for a modest accuracy gain.

AuthorsOpenReview ↗

Response to Q1, Q2

Response to Q1:

Thank you for pointing this out. We have revised Eq. (19) to explicitly include the mode permutation before tensor singular value thresholding. For each view, we stack Z(v)Z^{(v)} and SS into an n×n×2n\times n\times 2 tensor and permute it to n×2×nn\times 2\times n before applying the FFT along the third mode. The FFT is still required, and the permutation changes the size of each Fourier-domain frontal slice from n×nn\times n to n×2n\times 2.

Specifically, define W(v)=H(v)+T(v)ρ{\mathcal{W}^{(v)}} = {{\mathcal{H}^{(v)}}} + \frac{{{\mathcal{T}^{(v)}}}}{\rho}. Let $\widetilde{{\mathcal W}}^{(v)} = P\left({\mathcal W}^{(v)}\right) \in \mathbb R^{n\times2\times n}$, where P(⋅)P(\cdot) denotes the mode permutation Rn×n×2→Rn×2×n\mathbb R^{n\times n\times2} \rightarrow \mathbb R^{n\times2\times n}. This arrangement models the high-order relationships across samples while performing SVD on compact n×2n\times2 frontal slices. Therefore, the optimal solution is: \begin{equation} {\mathcal J}^{(v)} = P^{-1} \left[ \Gamma_{1/\rho} \left( \widetilde{{\mathcal W}}^{(v)} \right) \right]. \end{equation}

For a conventional n×n×Vn\times n\times V tensor, the forward and inverse FFTs cost O(Vn2log⁡V)O(Vn^2\log V), and the VV dense SVDs cost O(Vn3)O(Vn^3). In the revised formulation, the FFTs cost O(n2log⁡n)O(n^2\log n) per view, and the nn thin SVDs cost O(n2)O(n^2) per view because each n×2n\times 2 SVD costs O(n)O(n). Across all VV views, the tensor thresholding step in each iteration therefore costs O(Vn2log⁡n)O(Vn^2\log n), compared with O(Vn2log⁡V+Vn3)O(Vn^2\log V+Vn^3) for the conventional formulation.

Response to Q2: Thank you for raising this question. In each view, we rank the feature values within each observed sample. Equal values receive the average of the ranks they occupy. For example, two values tied at positions 2 and 3 both receive rank 2.5. We then compute the Spearman similarity between two samples as the Pearson correlation of their rank vectors.

We tested sensitivity to additional ties on BBC4, where 98.81% of the observed feature entries are zero across the four views. We compared the original features with three progressively coarser representations. In each view, we grouped the nonzero values by magnitude into at most 16 or 4 bins and assigned all values in each bin the same code, creating additional ties. In the single-bin setting, we set every nonzero entry to 1. With the original zeros preserved, this last setting produces binary features that indicate whether each feature is present. We summarize the extent of ties using the median nonzero rank redundancy across observed samples, defined per sample as one minus the ratio of distinct nonzero values to nonzero entries. To assess local graph quality, we report same-class precision at 10 (P@10), the average proportion of each sample’s ten highest-weight neighbors that share its ground-truth class. We compute this metric for both the initial Spearman graph and the recovered consensus graph using the same observed samples and candidate neighbors in each view.

Table 2 Sensitivity to additional rank ties on BBC4 at 30% missingness. The original zero entries were preserved in all settings. In the binary setting, all nonzero values were set to 1. P@10 is the percentage of the top 10 neighbors that share the query sample's class label.

Nonzero feature treatmentMedian nonzero rank redundancy (%)Initial graph P@10 (%)Consensus graph P@10 (%)ACC (%)
Original values93.7574.0098.1198.41
At most 16 levels94.9274.0098.1198.41
At most 4 levels96.4573.9698.1198.41
One level (binary)98.0473.9098.1198.41

As Table 2 shows, using the same 30% missingness split and model settings, the initial graph’s mean same-class precision@10 changed from 74.00% with the original features to 73.90% with binary features. Across all four settings, the recovered consensus graph achieved 99.11% precision@10 on the same observed subsets, with 98.41% ACC. On this BBC4 split, preserving which features were present was sufficient to maintain the measured quality of the recovered graph under this test. In addition, the learned consensus graph in Fig. 4(b) exhibits a clear block-diagonal pattern, showing that RMTC-GL recovers a consensus graph with clear cluster structure on BBC4 as well.

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

This paper proposes RMTC-GL for incomplete multi-view clustering. The method constructs correlation graphs using Spearman correlation and recovers missing graph entries while preserving observed relationships. Tensor-based diversity constraints and multi-tensor regularization are then introduced to capture complementary information and learn a consensus graph. Experiments on several benchmark datasets demonstrate strong clustering performance under different missing rates.

Critical strengths
  1. The paper considers both cross-view diversity and consistency during incomplete graph recovery, which is relevant to the IMVC problem.
  2. The proposed optimization framework integrates correlation graph construction, graph recovery, and tensor-based consensus learning in a unified formulation.
  3. The experimental results are strong on several datasets, and the ablation studies provide some evidence for the effectiveness of the proposed components.
Critical weaknesses
  1. The methodological novelty is limited. The main components, including Spearman correlation, graph completion, and tensor-based regularization, are established techniques. Their integration does not clearly demonstrate a substantial advance over existing IMVC methods.
  2. The use of Spearman correlation is not sufficiently justified. Its robustness to feature magnitudes does not imply that it produces more informative clustering graphs than other similarity measures.
  3. The improvements on some datasets are considerable, but the underlying reasons are not well explained. More controlled comparisons would help establish which components account for these gains.
Questions for authors
  1. What is the key technical difference between RMTC-GL and existing tensor-based graph recovery methods for IMVC?
  2. How does Spearman correlation compare with other similarity measures within the same framework?
  3. Can the authors explain why the method achieves particularly large improvements on certain datasets?
2 comments
Area_Chair_MqDgOpenReview ↗

Please leave some comments for the authors!

Thank you for your thoughtful reviews. When you have a chance, please read the authors’ responses to your comments and update the discussion with your assessment. Please note whether your concerns have been addressed and briefly mention any remaining questions.

Reviewer_FWKNOpenReview ↗

Thanks for the detailed responses

The authors have adequately addressed my concerns regarding the methodological novelty of RMTC-GL, the justification for employing Spearman correlation, and the reasons behind its substantial improvements on certain datasets. In particular, the clarification of the consensus-coupled multi-tensor recovery formulation and its distinction from conventional tensor-based graph recovery methods helps establish the technical contributions of the proposed approach. Overall, I appreciate the authors' detailed responses and consider my major concerns satisfactorily resolved. Therefore, I support acceptance of this paper and assign a rating of 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.