acceptodds
Under review as a conference paper at ICLR 2027

Scalable Multi-View Clustering via Latent Tensor Low-Rank Representation

Abstract

Existing TLRR(Tensor low-rank representation) methods typically use the observed tensor itself as the dictionary to learn robust representations. However, when the observed data are limited and contaminated by noise, such a self-representation strategy is unable to adequately capture the underlying tensor subspace structure, resulting in degraded performance. To overcome this limitation, we propose Latent Tensor Low-Rank Representation (LT-LRR), a general tensor representation framework that augments the observed tensor dictionary with an unobserved latent component. We show that the resulting latent tensor self-representation admits a closed-form structural characterization under the t-product framework, which naturally yields low-rank constraints on both the representation tensor and the latent correction tensor. Building upon this formulation, we further develop ETRL-SMVC, a scalable multi-view clustering method that applies LT-LRR to anchor graph learning by replacing computationally expensive sample-level self-representation with compact dictionary-based tensor factorization over anchor graphs. Extensive experiments on benchmark datasets demonstrate that the proposed method consistently outperforms state-of-the-art competitors while maintaining favorable scalability.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view clustering, tensor low-rank representation, anchor graph, self-representation learning

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_mh6SOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To address the limitations of conventional tensor self-representation and factorization in capturing latent subspaces when observations are insufficient or noisy, this paper proposes Latent Tensor Low-Rank Representation (LT-LRR) and theoretically derives a low-rank correction formulation for latent information. Building on this framework, the authors develop ETRL-SMVC, which combines compact anchor-label tensor factorization with latent correction and Schatten-p regularization to enhance cross-view representation learning and reduce computational costs. Comparative experiments and ablation studies on four benchmark datasets validate the effectiveness of the proposed method

Critical strengths
  1. The work is well motivated, introducing latent representations to address insufficient sampling and extending the matrix formulation to the tensor setting.
  2. The method achieves competitive clustering results on benchmark datasets, with ablation studies demonstrating the effectiveness of its components.
Critical weaknesses
  1. The transition from anchor graph self-representation to tensor factorization is not sufficiently clear and requires further explanation.
  2. Scalability is primarily supported by theoretical complexity analysis, without runtime comparisons across varying sample sizes.
Questions for authors
  1. How are the hyperparameters selected for each dataset? Could the authors specify the search ranges and selection criteria?
  2. What advantages does the Schatten-p norm offer over the standard tensor nuclear norm (TNN)?
  3. How are anchors selected for different views? Is cross-view anchor correspondence required when stacking the anchor graphs into a tensor?
1 comment

Response to W1:

Thank you for your valuable comment. The main motivation for replacing self-representation learning with dictionary learning is to improve scalability while preserving the low-rank representation principle of LT-LRR. Specifically, directly applying LT-LRR to the anchor-graph tensor S∈Rm×n×V\mathcal{S}\in\mathbb{R}^{m\times n\times V} still requires learning a self-representation coefficient tensor Z\mathcal{Z} of size n×n×Vn\times n\times V, which incurs substantial computational costs for large-scale datasets.

To avoid explicitly learning Z\mathcal{Z}, we approximate the self-representation term using a compact tensor factorization, S∗Z≈G∗H⊤\mathcal{S}\ast\mathcal{Z}\approx\mathcal{G}\ast\mathcal{H}^{\top}, where G\mathcal{G} and H\mathcal{H} denote the anchor-label and sample-label tensors, respectively. This leads to the proposed formulation: S≈G∗H⊤+L∗S.\mathcal{S}\approx\mathcal{G}\ast\mathcal{H}^{\top}+\mathcal{L}\ast\mathcal{S}. Here, the tensor factorization term captures the underlying clustering structure, while the latent correction term preserves the model's ability to compensate for insufficient or noisy observations.

We will further clarify this transition in Section 3.2 of the revised manuscript and explicitly explain the mathematical relationship between the original LT-LRR formulation and the proposed compact tensor factorization model.

Response to W2:

Thank you for your valuable comment. To further evaluate the scalability of our method, we conducted additional experiments comparing runtime and peak memory consumption with several representative anchor graph-based clustering methods, including LMVSC, FPMVS-CAG, and AGFS-OMVC, FMCNOF. The experiments were conducted on two relatively large datasets, Reuters and NUS, containing 18,758 and 30,000 samples, respectively. The results are summarized below.

DatasetMethodRuntime (s)Peak Memory (MB)
ReutersLMVSC73.969981.27
FMCNOF15.7841099.88
FPMVS-CAG2571.31221260.85
AGFS-OMVC1348.27331946.89
Ours226.0601699.90
NUSLMVSC148.1591067.55
FMCNOF44.2501474.45
FPMVS-CAG241.3511286.76
AGFS-OMVC334.7621551.22
Ours183.3901175.40

These results demonstrate the practical feasibility of our method on large-scale datasets and provide additional empirical evidence to complement the theoretical complexity analysis. We will incorporate these experimental results and clarify the scalability analysis in the revised manuscript.

Response to Q1: Thank you for your valuable comment. We conduct a grid search to select the hyperparameters for each dataset. Specifically, for small-scale datasets, the anchor rate is searched from 0.1 to 1.0, while for large-scale datasets, we consider smaller anchor rates, such as 0.001, 0.002, and so on. The parameter (p) is searched from 0.1 to 1.0 with a step size of 0.1. The regularization parameters (\lambda_1), (\lambda_2), and (\lambda_3) are selected from ({0.001, 0.01, 0.1, 1, 10, 100, 1000}). We will clarify these hyperparameter search ranges and the selection procedure in the revised manuscript to improve reproducibility.

Compared with the standard tensor nuclear norm (TNN), the tensor Schatten-pp norm (0<p<10<p<1) provides a more flexible approximation to tensor rank.

TNN applies a uniform linear penalty to singular values, potentially suppressing important structures. In contrast, Schatten-pp regularization penalizes larger singular values relatively less while promoting the suppression of smaller ones, thereby better preserving dominant low-rank structures.

In our framework, it regularizes both the representation tensors and the latent correction tensor to capture cross-view structural relationships and exploit shared and complementary information. TNN is a special case when p=1p=1, highlighting the flexibility of the proposed regularization.

Response to Q3:

Thank you for your valuable comment. We adopt a variance-based decorrelation strategy for anchor selection. Specifically, we first concatenate the feature matrices from all views and calculate the feature variance of each sample. We then iteratively select the sample with the highest variance as an anchor and apply a decorrelation mechanism to reduce redundancy among the selected anchors.

Yes, cross-view anchor correspondence is required when stacking the anchor graphs into a tensor. Since the anchors are selected using the concatenated multi-view features, the same sample indices are used as anchors across all views, naturally ensuring their correspondence. Handling non-corresponding anchors across different views is an interesting and challenging problem, which we plan to investigate in our future work.

Reviewer_mLYzOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on the tensor low-rank representation problem in multi-view clustering. The authors point out that traditional TLRR methods usually directly use the observed tensor as the dictionary. However, when the number of samples is insufficient or the data contain noise, the observed dictionary may fail to adequately characterize the underlying subspace structure. To address this issue, this paper proposes Latent Tensor Low-rank Representation. The work introduces a latent tensor dictionary strategy, which supplements the observed tensor with unobserved latent components. Through rigorous mathematical derivations, a closed-form solution to the problem is obtained. Based on this theoretical foundation, the authors further apply this idea to anchor graph-based multi-view clustering and propose ETRL-SMVC. The method replaces large-scale sample-level self-representation with a compact factorization of anchor-label tensors and sample-label tensors, and imposes Schatten-p norm regularization on the related tensors. Experiments on multiple multi-view datasets demonstrate good clustering performance.

Critical strengths
  1. By introducing latent tensor components to enhance the observed dictionary and deriving a closed-form solution for tensor self-representation, this work addresses the limitations of traditional low-rank representation in robustness and scalability, offering a valuable theoretical contribution to multi-view clustering.
  2. The paper presents a relatively complete framework, including theoretical derivation, optimization algorithms, and complexity analysis. Experiments on four multi-view datasets demonstrate competitive clustering performance, while ablation studies and visualizations validate the effectiveness of key components.
Critical weaknesses
  1. The algorithm design in Section 3.2 is generally reasonable, but some key details are missing. The optimization problem appears to be complex, involving alternating optimization of multiple variables. Such nonconvex optimization problems are generally difficult to solve and often require careful initialization of each variable. However, the relevant descriptions in the paper are insufficient.
  2. The scalability claim lacks runtime and memory comparisons, while the anchor selection and graph construction procedures are not clearly described, limiting reproducibility.
Questions for authors
  1. Why might standard TLRR be restrictive in practical scenarios? Could the authors explain specifically why traditional TLRR may fail in such cases, while the proposed method can effectively address these limitations?
  2. Could the authors provide a detailed description of how each variable in optimization problem (17) is initialized?
3 comments

Response to W1/Q1:

Thank you for pointing this out. For each frequency slice, the sample-label variable HH is initialized as an n×cn\times c rectangular identity matrix, while the anchor-label variable GG is initialized by multiplying the transpose of the initial anchor-graph slice by HH. The auxiliary variables QQ, JJ, FF, and MM, together with the four Lagrange multipliers, are initialized to zero. The latent correction variable LL is not initialized separately but is directly updated in the first iteration. The four penalty parameters μ\mu, ρ\rho, σ\sigma, and β\beta are all initialized to 10−410^{-4}, with a growth factor η=1.2\eta=1.2. We will provide these initialization details in the revised manuscript to improve reproducibility.

Response to W2:

Thank you for your valuable comment.

1. Runtime and memory comparison.

To further evaluate scalability, we conducted additional experiments comparing runtime and peak memory consumption with several representative anchor graph-based clustering methods, including LMVSC, FMCNOF, FPMVS-CAG, and AGFS-OMVC, on Reuters (18,758 samples) and NUS (30,000 samples). The results are summarized below.

DatasetMethodRuntime (s)Peak Memory (MB)
ReutersLMVSC73.969981.27
FMCNOF15.7841099.88
FPMVS-CAG2571.31221260.85
AGFS-OMVC1348.27331946.89
Ours226.0601699.90
NUSLMVSC148.1591067.55
FMCNOF44.2501474.45
FPMVS-CAG241.3511286.76
AGFS-OMVC334.7621551.22
Ours183.3901175.40

Our method achieves lower runtime and memory consumption than FPMVS-CAG and AGFS-OMVC on Reuters. On NUS, it also outperforms FPMVS-CAG and AGFS-OMVC in runtime. These results demonstrate the practical feasibility of our method on large-scale datasets, complementing the theoretical complexity analysis.

2. Anchor selection and graph construction.

We adopt a variance-based decorrelation strategy for anchor selection. Specifically, we concatenate the feature matrices from all views, calculate each sample's feature variance, and iteratively select the highest-variance samples while reducing redundancy through decorrelation.

For each view, we construct a sample-anchor affinity graph using its corresponding feature representation and the selected anchors. The resulting anchor graphs are stacked into a third-order tensor. Since the same sample indices are selected as anchors across all views, cross-view anchor correspondence is naturally ensured.

We will incorporate the additional efficiency comparisons and clarify the anchor selection and graph construction procedures in the revised manuscript to improve reproducibility.

Response to Q1:

Thank you for your valuable question.

Traditional TLRR directly uses the observed tensor as the representation dictionary, which assumes that the observed data sufficiently capture the underlying subspace structure. However, when the observations are insufficient or contaminated by noise, the observed dictionary may fail to adequately represent the underlying subspaces, leading to inaccurate self-representation coefficients and degraded clustering performance.

To address this limitation, LT-LRR introduces an additional latent tensor dictionary to compensate for the insufficient representation ability of the observed data. The latent representation can be reformulated as a low-rank correction term, yielding:

\mathcal{B}=\mathcal{B}\ast\mathcal{Z}+\mathcal{L}\ast\mathcal{B}+\mathcal{E}.\$$ Here, the latent correction term captures structural information that cannot be adequately represented by the observed dictionary alone, while the error term accounts for reconstruction residuals. Furthermore, our ETRL-SMVC replaces the computationally expensive sample-level self-representation with compact tensor factorization, improving scalability.

Thank the authors for their response. My concerns have been properly addressed, and I will keep my score.

We sincerely thank the reviewer for your positive feedback and for acknowledging our clarifications. We greatly appreciate your time and valuable suggestions.

Reviewer_uQUQOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a scalable multi-view clustering framework based on Latent Tensor Low-Rank Representation (LT-LRR). Unlike conventional tensor low-rank representation methods that directly use the observed tensor as the dictionary, LT-LRR introduces an additional latent tensor component to compensate for incomplete or noisy observations. The authors provide a theoretical characterization showing that the latent tensor representation can be reformulated as a low-rank correction term. Based on LT-LRR, the authors further develop an efficient anchor graph-based clustering method. Instead of performing computationally expensive sample-level tensor self-representation, the proposed method learns compact anchor-label and sample-label tensors with tensor Schatten-p regularization and latent correction.

Critical strengths
  1. The paper proposes a latent tensor low-rank representation strategy that extends conventional tensor representation methods by introducing additional latent information. This provides a meaningful approach to improving representation robustness when the observed tensor is insufficient or noisy.
  2. The proposed method effectively combines tensor representation learning with anchor graph-based multi-view clustering, reducing computational complexity while exploiting cross-view structural information. The experimental results demonstrate competitive performance on several benchmark datasets.
Critical weaknesses
  1. Although introducing latent components is reasonable, the relationship between the proposed method and existing latent low-rank representation approaches is not sufficiently discussed. The authors should better explain the unique advantages brought by the tensor formulation.
  2. Although the ablation experiments demonstrate the effectiveness of different low-rank regularization terms, the specific contribution of the latent correction mechanism itself is not fully isolated. Further analysis would help clarify whether the improvements primarily arise from the latent representation mechanism or its associated low-rank regularization.
Questions for authors
  1. Could the authors further clarify the main theoretical and methodological differences from existing latent low-rank representation methods, particularly the additional benefits introduced by tensor modeling?
  2. Although the ablation experiments demonstrate the effectiveness of different regularization terms, the specific contribution of the latent correction mechanism remains insufficiently isolated. Could the authors provide a more direct comparison to verify its effectiveness in improving clustering performance?
  3. One of the main motivations of LT-LRR is to address the limitations of noisy or insufficient observations. Could the authors provide more systematic quantitative experiments under different noise levels or incomplete observation conditions to support this claim?
1 comment

Answer to W1/Q1:

Thank the reviewer for the valuable suggestion. The main contribution of this work lies in establishing a tensor-based latent low-rank representation framework that unifies latent information compensation and cross-view structural modeling. Specifically, the theoretical and methodological advantages of our approach are reflected in the following three aspects.

First, at the theoretical level, we introduce latent information modeling into the t-product-based tensor representation framework and derive a low-rank formulation of latent tensor representation based on t-SVD. This formulation not only provides a theoretical foundation for compensating for latent structural information that is insufficiently captured by the observed data, but also enables the latent correction to establish connections among different views through the t-product, thereby enhancing the representation capability of the observed data by exploiting cross-view structural information.

Second, at the structural modeling level, we employ the tensor Schatten-p norm to impose joint low-rank constraints on the latent correction tensor and representation tensors. Tensor low-rank modeling enables the learning of underlying low-rank structures from linear combinations of information across different views, thereby jointly exploiting shared and complementary information among multiple views and facilitating the recovery of more reliable clustering structures.

Finally, at the methodological level, we further develop ETRL-SMVC by integrating the latent tensor correction mechanism with anchor-graph tensor factorization. By jointly learning the sample-label tensor and anchor-label tensor, the proposed method avoids explicitly constructing a large-scale sample-level self-representation tensor, thereby reducing computational and storage costs while preserving the capability to model cross-view structural relationships.

Answer to W2/Q2:

We thank the reviewer for the valuable suggestion. We have conducted an additional ablation experiment in which the latent correction variable LL is fixed to zero at every iteration, while keeping the data preprocessing, anchor graphs, and all other update procedures unchanged. The results are presented in the following table.

DatasetFull model ACC / NMIL=0 ACC / NMI
MSRC1.000 / 1.0000.976 / 0.953
NUS0.566 / 0.7420.127 / 0.100
AWA0.539 / 0.7700.070/ 0.074
Reuters0.818/ 0.7390.333/ 0.089

These results demonstrate the effectiveness of the latent correction component, particularly on larger datasets.

Answer to W3:

We thank the reviewer for the valuable suggestion. We add view-wise scale-normalized Gaussian noise to the original features of the MSRC dataset:

\qquad N_{ij}^{(v)}{\sim}\mathcal{N}(0,1).$$ where \$\delta\in\{0,0.05,0.1,0.2\}\$. The ACC results at different noise levels are presented in the following table. | Noise level delta | Full model ACC | L=0 ACC | lambda3=0 ACC | | --- | --- | --- | --- | | 0 | 1.000| 0.976 | 0.504 | | 0.05 | 1.000 | 0.928 | 0.609 | | 0.1 | 1.000 | 0.947 | 0.366 | | 0.2 | 0.995 | 0.904 | 0.452 | These results suggest that learning a regularized latent correction helps preserve clustering structure when the anchor graphs are rebuilt from noisy observations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.