acceptodds
Under review as a conference paper at ICLR 2027

Prior-Corrected Anchor Learning with Direct Discrete Partitioning for Multi-View Clustering

Abstract

Anchor-based multi-view clustering improves computational efficiency by representing the relationships among samples using a small number of anchors. However, existing methods typically employ shared semantic priors across different views. Since individual views may contain view-specific noise, feature distortions, and outliers, a shared semantic prior may fail to accurately characterize each view. Moreover, learning representations and performing clustering assignments in separate stages may result in representations that are not directly optimized toward the final clustering objective.To address these issues, we propose a Prior-Corrected Anchor Learning method, termed PCAL, which jointly learns view-specific anchor matrices, a shared anchor representation, and discrete clustering assignments within a unified framework. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection. Building upon this mechanism, we further introduce regularized view-specific correction matrices, enabling the shared prior to adapt to the feature discrepancies across different views. In addition, a discrete factorization term is incorporated to directly connect anchor representation learning with cluster label assignment, thereby avoiding additional spectral clustering post-processing.For the proposed model, we develop an alternating optimization algorithm, in which each variable subproblem can be efficiently solved either in closed form or through singular value decomposition. Experimental results demonstrate that, compared with other multi-view clustering methods, the proposed PCAL method reduces computational time while achieving competitive clustering performance in terms of metrics such as ACC and NMI.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, anchor 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_wc7yOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This article proposes a Prior-Corrected Anchor Learning method, termed PCAL, which jointly learns view-specific anchor matrices, a shared anchor representation, and discrete clustering assignments within a unified framework. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection. Building upon this mechanism, this work further introduces regularized view-specific correction matrices, enabling the shared prior to adapt to the feature discrepancies across different views. In addition, a discrete factorization term is incorporated to directly connect anchor representation learning with cluster label assignment, thereby avoiding additional spectral clustering post-processing.

Critical strengths
  1. This article proposes a Prior-Corrected Anchor Learning method and further introduces regularized view-specific correction matrices, enabling the shared prior to adapt to the feature discrepancies across different views.

  2. A discrete factorization term is incorporated to directly connect anchor representation learning with cluster label assignment, thereby avoiding additional spectral clustering post-processing.

Critical weaknesses
  1. The novelty of the proposed PCAL is not very clear in Figure 1 and the new parts in this figure can be highlighted. The collections among the part view-specific prior correction, shared anchor representation and discrete cluster assignment should be given in this description for Figure 1.

  2. The authors give the compact view-specific prior correction in 3.2 for prior-corrected anchor learning. PCAL uses a predefined category-anchor grouping matrix Y as a common reference. Constructing this matrix does not require ground-truth sample labels. The mentioned compact view-specific prior correction is not clearly presented for this part, especially how compact can be explained.

Questions for authors
  1. The shared anchor representation is the popular-used words for the existing multi-view clustering works for large-scale data. Therefore, using shared anchor representation is not very reasonable as the title of this part and the authors should use some representative words in summarizing this part.

  2. The authors do not give enough explanations or analysis for running time and scalability in 5.3 for the experiment. As is known, the efficiency is an important metric for multi-view clustering for large scale as well as effectiveness. Then further analysis for running time and scalability is expected to be given in this part.

  3. The authors can give the shortcomings of this work and further study can be presented in Conclusion part. The authors just give that future work will investigate which forms of view heterogeneity benefit most from prior correction and assess computational efficiency under more closely matched preprocessing and parameter-search protocols. How this mentioned future work relates with the proposed work in this paper?

5 comments

Q1: The novelty of the proposed PCAL is not very clear in Figure 1 and the new parts in this figure can be highlighted. The collections among the part view-specific prior correction, shared anchor representation and discrete cluster assignment should be given in this description for Figure 1.

A1: Good question! PCAL’s contribution is to learn a separate correction to the shared structural prior for each view. In PALCS, the fixed prior YY guides anchor learning, while the view-specific terms enter the graph semantics module. In PCAL, the correction E(v)E^{(v)} acts directly on YY, giving each view a corrected prior Y+E(v)Y+E^{(v)} for anchor learning. For Figure 1 on page 2 of the submitted manuscript, we propose retaining the three-column PCAL overview and adding a compact PALCS–PCAL comparison beneath it. In the left column, we propose highlighting the View-specific correction and Corrected Priors blocks and labeling the mapping path Guides anchor learning. The bottom comparison will show that the paths from YY and E(v)E^{(v)} lead to separate modules in PALCS and merge at Y+E(v)Y+E^{(v)} in PCAL. The three PCAL modules are connected through the view-specific anchors A(v)A^{(v)} and the shared representation CC. First, the mapped prior P(v)(Y+E(v))P^{(v)}(Y+E^{(v)}) provides an alignment target for the anchor update in Eq. 5. The anchors from all views then contribute to the update of CC in Eq. 8, together with the current partition through the term λGF\lambda GF. The updated CC informs the next prototype and assignment updates in Eqs.12–14. In this way, prior correction influences the partition through the anchors and CC, while the current partition feeds back into representation learning. We propose replacing the Figure 1 caption at lines 069–071 on page 2 of the submitted manuscript with the following text. Figure 1. Overview of PCAL. PALCS uses a fixed structural prior to guide anchor learning (Peng et al., 2025). PCAL learns a view-specific correction E(v)E^{(v)} to the shared prior YY and uses the mapped code P(v)(Y+E(v))P^{(v)}(Y+E^{(v)}) to guide anchor learning. The learned anchors reconstruct each view through the shared representation CC, and the factorization C≈GFC\approx GF couples CC to the discrete sample assignments FF.

Q2: The authors give the compact view-specific prior correction in 3.2 for prior-corrected anchor learning. PCAL uses a predefined category-anchor grouping matrix Y as a common reference. Constructing this matrix does not require ground-truth sample labels. The mentioned compact view-specific prior correction is not clearly presented for this part, especially how compact can be explained.

A2: Thanks! The predefined matrix YY in Eq.2 groups mm anchor positions into kk groups, with h=m/kh=m/k positions in each group. This grouping is constructed from the numbers of clusters and anchors. The sample assignments are learned through FF during optimization. The prior correction is compact because each view learns a single k×mk\times m correction matrix E(v)E^{(v)} shared across all its samples. Equation 10 computes this correction from the difference between the projected anchors (P(v))⊤A(v)(P^{(v)})^\top A^{(v)} and the shared prior YY. Both matrices have size k×mk\times m, so each entry of E(v)E^{(v)} adjusts the prior for one category–anchor pair. The VV corrections therefore require VkmVkm entries in total, independent of the sample count nn for fixed VV, kk, and$m$ . In the NUSWIDEOBJ scaling experiment, we fix V=5V=5 and k=m=31k=m=31as the sample count grows from 6,002 to 30,000. The shared representation CC grows from 31×6,00231\times 6{,}002 to$31\times 30{,}000$ , while the five correction matrices together contain 5×31×31=4,8055\times 31\times 31=4{,}805 entries at every sample size. This is the specific storage saving of the correction mechanism. We propose replacing the passage beginning with The adjustment concerns the anchor-grouping structure at lines 205–208 on page 4 of the submitted manuscript (Section 3.2) with the following text. Each view learns one k×mk\times m correction matrix E(v)E^{(v)}shared across its samples, with one entry per category–anchor pair. The VV corrections therefore require VkmVkm entries regardless of the sample count for fixed$V$ , kk, and mm . As samples are added, CC and FF each gain one column per sample, while the dimensions of the correction matrices remain fixed.

Q3: The shared anchor representation is the popular-used words for the existing multi-view clustering works for large-scale data. Therefore, using shared anchor representation is not very reasonable as the title of this part and the authors should use some representative words in summarizing this part.

A3: Good question! The current title Shared Anchor Representation names the variable CC, a concept widely used in anchor-based multi-view clustering. A more informative title would describe what Section 3.1 formulates. In Eq. 1, each view is reconstructed as X(v)≈A(v)CX^{(v)}\approx A^{(v)}C. The anchor matrix A(v)A^{(v)} accounts for the feature space of view vv, while the shared coefficient matrix CC gives each sample a common representation across views. We propose replacing the heading 3.1 Shared Anchor Representation on page 3, line 158 of the submitted manuscript with the following title. 3.1 Multi-View Reconstruction with Shared Anchor Coefficients

Q4: The authors do not give enough explanations or analysis for running time and scalability in 5.3 for the experiment. As is known, the efficiency is an important metric for multi-view clustering for large scale as well as effectiveness. Then further analysis for running time and scalability is expected to be given in this part.

A4: Thanks! PCAL scales with the number of samples nn by keeping its main linear systems in the mm-dimensional anchor space. As samples are added, the shared representation C∈Rm×nC\in\mathbb{R}^{m\times n} and the assignment matrix F∈{0,1}k×nF\in\{0,1\}^{k\times n} gain one column per sample. The coefficient matrices in Eqs.5 and 8 remain m×mm\times m, and the representation update reuses a single factorization for all nn right-hand sides. The VV view-specific correction matrices require$Vkm$ entries, independent of nn for fixed VV,kk , and mm. This structure yields two scaling properties. First, Eq.15 gives a per-iteration cost that grows linearly with nn when the anchor count, cluster count, feature dimensions, and number of views are fixed. Total optimization time also depends on the iteration count. Second, the principal storage bound in Eq. 16 is$\mathcal{O}(mn+Dm+Vkm+m^2)$ , where DD is the sum of the feature dimensions across views. This bound excludes input data and temporary buffers. For example, in the NUSWIDEOBJ experiment, increasing the sample count from 6,002 to 30,000 increases the number of entries in CC from 186,062 to 930,000, while both main coefficient matrices remain$31\times 31$ . To test this behavior in practice, we varied the sample count on cluster-stratified subsets of NUSWIDEOBJ while keeping the model parameters fixed. Each subset size was run three times with seeds 11–13. The solver timings include initialization and optimization and end before final label-index extraction.

Number of samplesPCAL solver time (s)
6,0028.95±0.238.95 \pm 0.23
12,00112.51±0.3712.51 \pm 0.37
17,99915.16±0.2415.16 \pm 0.24
23,99817.79±0.4917.79 \pm 0.49
30,00020.51±0.0920.51 \pm 0.09

All five sample sizes were processed with the same model settings. The experiment measures elapsed time through 30,000 samples, while Eq. 16 describes how the sizes of the stored matrices grow. In the separate cross-method comparison that times methods through final label output, PCAL takes 20.408s on the full dataset. This is faster than SMVSC (50.468s), FPMVS-CAG (98.619s), and OMSC (105.523s), which achieve higher ACC. FastMICE is faster (16.383s) and achieves higher NMI, while PCAL achieves higher ACC 0.1664 vs. 0.14910.1664\ \text{vs.}\ 0.1491. We propose replacing lines 416–431 on page 8 of the submitted manuscript, from On CCV, it achieves through the end of Section 5.3, with the following text. On CCV, PCAL achieves the highest mean ACC and NMI in 0.631s. On NUSWIDEOBJ, it produces labels for 30,000 samples in 20.408s on average. As the sample count increases, CC and FF gain one column per sample, while the main coefficient matrices remain m×mm\times m and the prior corrections retain VkmVkm entries. With the model dimensions fixed, both per-iteration computation and principal storage grow linearly with nn (Eqs.15–16). In the fixed-parameter NUSWIDEOBJ subset experiment, mean solver time including initialization rises from 8.95s at 6,002 samples to 20.51s at 30,000 samples. In the cross-method comparison on the full dataset, PCAL is faster than SMVSC, FPMVS-CAG, and OMSC, which achieve higher ACC. FastMICE is faster and achieves higher NMI, while PCAL achieves higher ACC.

Q5: The authors can give the shortcomings of this work and further study can be presented in Conclusion part. The authors just give that future work will investigate which forms of view heterogeneity benefit most from prior correction and assess computational efficiency under more closely matched preprocessing and parameter-search protocols. How this mentioned future work relates with the proposed work in this paper?

A5: Good question! PCAL learns one correction matrix E(v)E^{(v)} per view, shared across all samples in that view. This design keeps the correction compact, but limits it to adjusting the anchor grouping at the view level. A systematic feature distortion can shift the anchors across a view, whereas corruption in a small subset of samples can affect the estimated anchors without reflecting a shift that applies to the whole view. These two cases may therefore respond differently to the same correction mechanism. In Table 2, enabling correction on Caltech101-20 improves mean NMI from 0.6050 to 0.6092 while reducing mean ACC from 0.5811 to 0.5679. The ablation establishes that the benefit depends on the metric, but leaves open which types of view heterogeneity make correction useful. We chose to study systematic feature distortion and sample-level corruption separately because they test the assumption underlying PCAL’s design that a single correction shared within a view can capture the relevant deviation. Comparing Full PCAL with DiscreteOnly at fixed settings and across distortion levels would help identify when that assumption holds. A second limitation is that compact correction matrices can still increase solver time. On NUSWIDEOBJ, enabling correction raises the mean runtime from 6.87s to 20.49s, while mean ACC and NMI are unchanged at the reported precision 0.1664 and 0.12620.1664\ \text{and}\ 0.1262. This motivates measuring the additional solver cost of enabling correction and interpreting it together with the change in clustering quality. For comparisons across methods, the current timings also exclude parameter search, and preprocessing differs across implementations. For example, OPMC standardizes its input before timing. We therefore chose matched preprocessing and tuning budgets as a future direction so that differences in timing can be interpreted under comparable conditions. Reporting preprocessing, tuning, and solver time separately would show where the computational cost arises. We propose replacing the passage beginning with However, the benefits of prior correction at lines 482-485 on page 9 of the submitted manuscript, through the end of the Conclusion, with the following text. Each view uses one correction shared across its samples, so its benefit may depend on the source of view heterogeneity. We will test systematic feature distortion and sample-level corruption separately. Prior correction can also increase solver time. We will measure this overhead and evaluate practical efficiency under matched preprocessing and tuning budgets.

Reviewer_V5zAOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a Prior-Corrected Anchor Learning method, termed PCAL, which jointly learns view-specific anchor matrices, a shared anchor representation, and discrete clustering assignments within a unified framework. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection. Building upon this mechanism, the authors further introduce regularized view-specific correction matrices, enabling the shared prior to adapt to the feature discrepancies across different views.

Critical strengths
  1. The paper proposes a Prior-Corrected Anchor Learning method, which jointly learns view-specific anchor matrices, a shared anchor representation, and discrete clustering assignments within a unified framework.
  2. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection.
Critical weaknesses
  1. The authors are expected to compare the proposed PCAL with the most related work in the overview of PCAL for Figure 1. Then the novelty of the proposed PCAL can be more easily achieved by the readers.
  2. How to understand the claim that “View-specific corrections adapt a shared structural prior to guide” in Figure 1, the authors are expected to give more detailed explanations for this claim.
Questions for authors
  1. The authors introduce a view-specific prior-correction mechanism that learns regularized corrections in the category-anchor space. How this view-specific prior-correction mechanism differs from mechanisms in the existing works?
  2. The authors are expected to give more in-depth analysis of the clustering performance in 5.2 for the experiment, which will make the analysis for the clustering results more comprehensive.
  3. The authors construct four variants around prior correction and joint discrete partitioning in ablation study for the experiment. Then the reason why construct such four variants can be given in this part.
5 comments

Q1: The authors are expected to compare the proposed PCAL with the most related work in the overview of PCAL for Figure 1. Then the novelty of the proposed PCAL can be more easily achieved by the readers.

A1: Good question! For Figure 1 on page 2 of the submitted manuscript, we propose retaining the existing three-column PCAL overview and adding a compact side-by-side comparison beneath it. The left panel will show PALCS, where the fixed prior YY guides anchor learning and the view-specific term E(v)E^{(v)} enters the graph semantics module. The right panel will show PCAL, where YY and E(v)E^{(v)} combine to form the corrected prior Y+E(v)Y+E^{(v)} , which is mapped into the view’s feature space to guide anchor learning. In the proposed comparison, both panels use the same color for E(v)E^{(v)} . The arrows show the key distinction between the methods. PALCS uses the view-specific term in the graph semantics module, while PCAL uses it to adjust the structural prior that guides anchor learning.

Q2: How to understand the claim that “View-specific corrections adapt a shared structural prior to guide” in Figure 1, the authors are expected to give more detailed explanations for this claim.

A2: Thanks! The shared prior YY specifies how anchor positions are grouped. For view vv, PCAL learns a regularized correction E(v)E^{(v)} in the category–anchor space and uses P(v)P^{(v)} to map the corrected code Y+E(v)Y+E^{(v)} into that view’s feature space. The alignment term α∥A(v)−P(v)(Y+E(v))∥F2\alpha\lVert A^{(v)}-P^{(v)}(Y+E^{(v)})\rVert_F^2 in Eq.3 encourages the anchors A(v)A^{(v)} to stay close to the mapped code, while the reconstruction term encourages them to fit the observed data. The corrected prior thus provides a soft alignment target for anchor learning. In the upper part of the proposed Figure 1, we highlight the View-specific correction and Corrected Priors blocks and label the mapping path “Guides anchor learning.” We propose replacing the Figure 1 caption at lines 069–071 on page 2 of the submitted manuscript with the following text. Figure 1. Overview of PCAL. PALCS uses a fixed structural prior to guide anchor learning (Peng et al., 2025). PCAL learns a view-specific correction E(v)E^{(v)} to the shared prior YY and uses the mapped code P(v)(Y+E(v))P^{(v)}(Y+E^{(v)}) to guide anchor learning. The learned anchors reconstruct each view through the shared representation CC, and the factorization C≈GFC\approx GF couples CC to the discrete sample assignments FF.

Q3: The authors introduce a view-specific prior-correction mechanism that learns regularized corrections in the category-anchor space. How this view-specific prior-correction mechanism differs from mechanisms in the existing works?

A3: Good question! The key distinction lies in what the view-specific term modifies. PALCS uses a fixed structural prior to guide anchor learning and models view-specific noise and outliers in its graph semantics module. ALPC learns an anchor-cluster indicator shared across views and uses it to link the cluster structures of anchors and samples. PCAL uses a predefined category anchor grouping YY as a reference and learns a separate regularized correction E(v)E^{(v)} for each view. The corrected code Y+E(v)Y+E^{(v)} directly changes the alignment target used to learn that view’s anchors. The comparison at the bottom of the revised Figure 1 shows where the view-specific terms act in PALCS and PCAL. Equation 10 of our manuscript shows how PCAL adjusts the prior. The correction is obtained by scaling the discrepancy (P(v))⊤A(v)−Y(P^{(v)})^\top A^{(v)}-Y by the shrinkage factor αα+β\frac{\alpha}{\alpha+\beta}. For fixed A(v)A^{(v)} and P(v)P^{(v)}, the regularization weight β\beta controls the magnitude of the correction. Each correction matrix has size k×mk\times m, where kk and mm are the numbers of clusters and anchors, respectively. We propose replacing the passage from PALCS guides anchor learning through remains fixed across views at lines 075–077 on page 2 of the submitted manuscript with the following text. PALCS uses a fixed structural prior to guide anchor learning and models view-specific variation in its graph semantics module (Peng et al., 2025). PCAL learns a regularized correction E(v)E^{(v)} to the shared prior YY for each view (see Figure 1).

Q4: The authors are expected to give more in-depth analysis of the clustering performance in 5.2 for the experiment, which will make the analysis for the clustering results more comprehensive.

A4: Thanks! On CCV, PCAL achieves the highest mean ACC and NMI among the compared methods, at 0.2458 and 0.1965, respectively. The runner-up, OMSC, achieves 0.2402 and 0.1855. PCAL ranks second on both metrics on MSRC and BBCSport, although the gaps differ considerably. On MSRC, PCAL’s mean ACC of 0.7771 is close to FastMICE’s 0.7819. On BBCSport, PCAL achieves mean ACC/NMI of 0.6438/0.4755, compared with 0.9704/0.9022 for FastMICE. PCAL’s BBCSport ACC also has a standard deviation of 0.1128 across ten runs, indicating substantial variation across runs. On Caltech101-20, PCAL’s mean ACC of 0.5679 is 9.56 percentage points below the best result of 0.6635 from FPMVS-CAG. The NMI gap is smaller, with PCAL at 0.6092 and OMSC at 0.6420. On NUS-WIDE and NUSWIDEOBJ, PCAL achieves mean ACC/NMI of 0.1948/0.1723 and 0.1664/0.1262, respectively, below the best scores on both metrics. These comparisons distinguish settings where PCAL performs close to the best method from those where substantial gaps remain. For ease of comparison, we report PCAL’s ACC and NMI from Table 7 of the submitted manuscript below. The values are means ± standard deviations across ten runs.

DatasetACCNMI
MSRC0.7771±0.03130.7771 \pm 0.03130.6713±0.00940.6713 \pm 0.0094
BBCSport0.6438±0.11280.6438 \pm 0.11280.4755±0.09990.4755 \pm 0.0999
Caltech101-200.5679±0.04100.5679 \pm 0.04100.6092±0.01780.6092 \pm 0.0178
CCV0.2458±0.00630.2458 \pm 0.00630.1965±0.00250.1965 \pm 0.0025
NUS-WIDE0.1948±0.00800.1948 \pm 0.00800.1723±0.00380.1723 \pm 0.0038
NUSWIDEOBJ0.1664±0.00400.1664 \pm 0.00400.1262±0.00260.1262 \pm 0.0026

We propose replacing the performance discussion in Section 5.2 on page 7, lines 370–373 of the submitted manuscript with the following text. PCAL leads on both ACC and NMI on CCV and ranks second on both metrics on MSRC and BBCSport. Its MSRC accuracy is close to FastMICE’s 0.7771 vs. 0.78190.7771\ \text{vs.}\ 0.7819, while its BBCSport accuracy varies substantially across runs SD=0.1128\mathrm{SD}=0.1128. On Caltech101-20, the gap to the best score is larger for ACC than for NMI. On both NUS datasets, PCAL trails the best scores on both metrics.

Q5: The authors construct four variants around prior correction and joint discrete partitioning in ablation study for the experiment. Then the reason why construct such four variants can be given in this part.

A5: Good question! The four variants form a 2×22\times 2 ablation design in which prior correction and joint discrete partitioning are enabled or disabled independently. This allows us to assess each component while holding the other fixed. Comparing Base with CorrectionOnly shows the effect of prior correction when joint partitioning is disabled. Comparing DiscreteOnly with Full PCAL shows its effect when joint partitioning is enabled. Similarly, comparing Base with DiscreteOnly and CorrectionOnly with Full PCAL shows the effect of joint partitioning with prior correction disabled and enabled, respectively. For example, on MSRC, adding prior correction increases mean NMI from 0.6470 to 0.6639 when joint partitioning is disabled, but decreases it from 0.6802 to 0.6713 when joint partitioning is enabled. These paired comparisons suggest that the effect of prior correction can depend on whether joint partitioning is enabled. All variants use the full model’s parameter settings and the same random seeds, so the comparisons show how enabling or disabling each component affects performance under fixed settings. We propose replacing lines 444-450 on page 9 of the submitted manuscript, the opening paragraph of Section 5.4, with the following text. We use a 2×22\times 2 ablation design to evaluate each component with the other component enabled and disabled. Base disables both components, CorrectionOnly enables prior correction alone, DiscreteOnly enables joint discrete partitioning alone, and Full PCAL enables both. All variants retain reconstruction and prior alignment and use the full model’s parameter settings and the same random seeds. When prior correction is disabled, we set .E(v)=0E^{(v)}=0 When joint partitioning is disabled, we remove the discrete factorization term and apply kk-means to the learned representation to obtain cluster labels.

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

This paper proposes a Prior-Corrected Anchor Learning method, which jointly learns view-specific anchor matrices, a shared anchor representation, and discrete clustering assignments within a unified framework. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection. Experimental results demonstrate that, compared with other multi-view clustering methods, the proposed PCAL method substantially reduces computational time while achieving competitive clustering performance in terms of metrics.

Critical strengths
  1. PCAL reconstructs each view through the shared anchor representation and aligns the anchor matrix of each view with the shared semantic prior via orthogonal projection.
  2. Experimental results demonstrate that, compared with other multi-view clustering methods, the proposed PCAL method substantially reduces computational time while achieving competitive clustering performance in terms of metrics.
Critical weaknesses
  1. For the motivation part, the authors state that the existing works motivate an efficient anchor model that retains clustering feedback in a compact representation and supports joint optimization through small matrix operations. However, the further exploration behind this motivation is not very clear.
  2. The authors mention that the proposed model coordinates data reconstruction, prior correction, and discrete cluster assignment. Then how this model coordinates these three parts can be explained for the corresponding part in the work.
Questions for authors
  1. For the motivation part, the authors state that the existing works motivate an efficient anchor model that retains clustering feedback in a compact representation and supports joint optimization through small matrix operations. However, the further exploration behind this motivation is not very clear.
  2. The authors mention that the proposed model coordinates data reconstruction, prior correction, and discrete cluster assignment. Then how this model coordinates these three parts can be explained.
  3. The organized manner for the three parts in related work is not very clear and the authors should explain the relationship among these three parts. For instance, why matrix-factorization-based multi-view representation learning is presented in related work and what is the relationship between it and the proposed method.
  4. The authors should check the whole paper to avoid the possible existing typo issues, i.e., there should be a blank between “correction” and “and” for the last line in page 2. There is the corresponding command for trace “T” in latex and the authors can replace T with this standard symbol in latex.
4 comments

Q1: For the motivation part, the authors state that the existing works motivate an efficient anchor model that retains clustering feedback in a compact representation and supports joint optimization through small matrix operations. However, the further exploration behind this motivation is not very clear.

A1: Thanks! The motivation for compact clustering feedback and joint optimization with small matrices is to let the cluster assignments guide representation learning while preserving the computational advantages of anchor-based modeling. A low reconstruction error indicates that the representation retains information useful for explaining the observed features. Samples assigned to the same cluster, however, may still be far apart in the representation space. When labels are assigned after the representation has been fixed, the resulting partition cannot influence which sources of variation the representation retains.

PCAL addresses this gap by including the term λ∥C−GF∥F2\lambda\lVert C-GF\rVert_F^2 in Eq.3. For a sample currently assigned to cluster rjr_j, this term penalizes the distance between its anchor representation cjc_j and the corresponding prototype grjg_{r_j}. Equation 8 makes the feedback explicit. Its right-hand side contains both the reconstruction contribution ∑v(A(v))⊤X(v)\sum_v (A^{(v)})^\top X^{(v)} and the clustering contribution λGF\lambda GF. The current assignments thus influence the next representation update, and the updated representation informs the next assignments. PALCS includes a cluster partitioning module. In PCAL, the current partition guides the update of the shared anchor representation through the λGF\lambda GF term in Eq.8. PCAL also learns a regularized k×mk\times m correction matrix for each view to adapt the shared structural prior. All samples in a view share the same correction matrix. Here, optimization with small matrices refers to the size of the matrices that the algorithm factorizes. The anchor and representation updates in Eqs.5 and 8 both use m×mm\times m coefficient matrices, where m≪nm\ll n; The anchor update reuses the same factorization across views, and the prototype update computes a thin SVD of the m×km\times k matrix CF⊤CF^\top . The representation update still solves a system with nn right-hand sides, and the assignment step compares each sample with the prototypes. The algorithm therefore performs the required work for each sample without forming or factorizing an n×nn\times n sample affinity matrix. With the other dimensions and the iteration count fixed, the optimization cost in Eq.15 grows linearly with nn, excluding initialization. Table 2 shows how this feedback affects clustering quality. Under the same parameter settings on CCV, adding joint discrete partitioning to Base gives the DiscreteOnly variant and increases mean ACC/NMI from 0.0824/0.02060.0824/0.0206 to 0.2458/0.19650.2458/0.1965 . On the two NUS datasets, the same change increases ACC and decreases NMI. The effect of incorporating the partition during representation learning therefore depends on the dataset and evaluation metric. We propose replacing lines 083–088 on page 2 of the submitted manuscript, beginning with Coordination between representation learning and cluster assignment, with the following text. Low reconstruction error can leave samples poorly organized for clustering. Building on Fang et al. (2024) and Peng et al. (2025), PCAL couples the shared anchor representation to the current discrete partition so that cluster assignments can guide subsequent updates. Each view adapts the shared prior through a regularized k×mk\times m correction. The anchor and representation updates use m×mm\times m coefficient matrices, with m≪nm\ll n, and avoid constructing an n×nn\times n sample affinity matrix (see Section 4.2).

Q2: The authors mention that the proposed model coordinates data reconstruction, prior correction, and discrete cluster assignment. Then how this model coordinates these three parts can be explained for the corresponding part in the work.

A2: Good question! The view-specific anchors A(v)A^{(v)} and shared representation CC connect the three terms. The anchor update in Eq.5 uses both the data-fit term X(v)C⊤X^{(v)}C^\top and the corrected prior term αP(v)(Y+E(v))\alpha P^{(v)}(Y+E^{(v)}). Equation 10 updates the correction from the discrepancy (P(v))⊤A(v)−Y(P^{(v)})^\top A^{(v)}-Y, scaled by αα+β\frac{\alpha}{\alpha+\beta}. These updates let the data influence the correction through the learned anchors and let the corrected prior guide subsequent anchor updates. Equation 8 then updates CC using both the anchors from all views and the current prototype targets GFGF. The updated CC is used to update the prototypes and assignments in Eqs.12–14; the resulting GFGF feeds back into the next representation update. Because the anchor update also depends on CC, this feedback can influence subsequent anchor learning. The prior YY specifies a grouping of anchor positions, while FF is learned from the sample representations.

Q3: The organized manner for the three parts in related work is not very clear and the authors should explain the relationship among these three parts. For instance, why matrix-factorization-based multi-view representation learning is presented in related work and what is the relationship between it and the proposed method.

A3: Thanks! Section 2.1 provides the starting point for PCAL’s reconstruction model. The factorization X(v)≈A(v)CX^{(v)}\approx A^{(v)}C in Eq.1 uses a separate basis matrix A(v)A^{(v)} for each view because the views occupy different feature spaces. All views share the coefficient matrix CC, whose jj-th column represents the same sample across views. This factorization explains how PCAL integrates information from multiple views. The reconstruction objective alone leaves open how the anchors should be organized and how the learned representation should support clustering. Section2.2 motivates the use of a shared structural reference that each view can adapt through a regularized correction. Section2.3 explains how structured factorization allows discrete assignments to guide representation learning. We propose replacing lines 120–126 on page 3 of the submitted manuscript, the opening paragraph of Section2.1, with the following text. Multi-view matrix factorization combines view-specific bases with shared or coordinated latent representations (Liu et al., 2013). Related methods learn hierarchical representations (Zhao et al., 2017), derive cluster indicators through orthogonal factorization (Yang et al., 2021), or couple graph learning with tensor modeling (Wu et al., 2020). This literature motivates PCAL’s reconstruction model X(v)≈A(v)CX^{(v)}\approx A^{(v)}C. Sections 2.2 and 2.3 motivate complementing reconstruction with view-specific prior correction and structured discrete clustering, respectively.

Q4: The authors should check the whole paper to avoid the possible existing typo issues, i.e., there should be a blank between “correction” and “and” for the last line in page 2. There is the corresponding command for trace “T” in latex and the authors can replace T with this standard symbol in latex.

A4: Good question! On page 2, line 107 of the submitted manuscript, we propose replacing correction,and with correction, and. We also propose using the LaTeX operator \operatorname{Tr} consistently in the following expressions. Page 5, line 269, Eq. (6): Tr⁡((P(v))⊤A(v)(B(v))⊤)\operatorname{Tr}\bigl((P^{(v)})^\top A^{(v)}(B^{(v)})^\top\bigr). Page 6, line 298, Eq. (11):Tr⁡(G⊤CF⊤)\operatorname{Tr}(G^\top CF^\top). Page 12, line 633, Appendix A.2, Eq. (21): the expansion is ∥A(v)−P(v)B(v)∥F2=∥A(v)∥F2+∥B(v)∥F2−2Tr⁡((P(v))⊤A(v)(B(v))⊤)\Vert A^{(v)}-P^{(v)}B^{(v)}\Vert_{F}^{2}=\Vert A^{(v)}\Vert_{F}^{2}+\Vert B^{(v)}\Vert_{F}^{2}-2\operatorname{Tr}\left((P^{(v)})^{\top}A^{(v)}(B^{(v)})^{\top}\right) Page 12, line 637, Appendix A.2, Eq. (22):

\operatorname{Tr}\bigl((P^{(v)})^\top A^{(v)}(B^{(v)})^\top\bigr).$$ We will also check the manuscript for similar spacing and notation inconsistencies.
Area_Chair_1P6zOpenReview ↗

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.

I thank the authors for their detailed and constructive responses. The rebuttal has satisfactorily addressed my main concerns, particularly regarding the motivation of the proposed framework and the interactions among data reconstruction, prior correction, and discrete cluster assignment. I am therefore happy to increase my rating to 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.