acceptodds
Under review as a conference paper at ICLR 2027

Multi-view Collaborative Label Learning with Balanced Regularization

Abstract

Anchor graph-based methods reduce computational complexity and improve scalability in multi-view clustering. However, existing approaches suffer from two major limitations: they treat anchor graphs merely as sample representations — leading to under- representation and thus losing fine-grained structural information — and they ignore the consistency between anchor distribution and sample distribution, which results in suboptimal and unstable performance. To address these issues, we propose a balanced multi-view collaborative label learning method based on anchor representation. Specifically, instead of using anchors to represent samples, we use samples to represent anchors, which enhances the representational power of anchors and effectively resolves the under-representation problem inherent in the former approach. In addition, we design an anchor balancing mechanism based on the Schatten p-norm, supported by theoretical analysis, to solve the distribution consistency problem between the anchor space and the sample space. Collaborative learning is achieved by alternately updating sample labels and anchor labels, while a low-rank tensor constraint is imposed on the sample labels to capture complementary information across views. Extensive experiments demonstrate the effectiveness and superiority of our method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, anchor graph, tensor Schatten p-norm

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ZuqZOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a balanced multi-view collaborative label learning method based on anchor representations. Unlike conventional anchor-based clustering approaches, the proposed framework uses sample information to represent anchors and establishes a collaborative relationship between sample labels and anchor assignments. A Schatten-p norm maximization regularizer is introduced to encourage balanced cluster assignments, supported by theoretical analysis. Furthermore, a tensor low-rank regularizer is employed to exploit complementary information across multiple views. Extensive experiments on seven benchmark datasets demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper introduces a different perspective on anchor-based clustering by using samples to represent anchors. The collaborative relationship between anchor assignments and sample labels provides a meaningful approach to multi-view label learning.
  2. The Schatten-p norm-based balancing strategy is supported by theoretical analysis, providing an interesting interpretation of spectral regularization for clustering. The proposed method achieves impressive performance on seven benchmark datasets, demonstrating its potential effectiveness.
Critical weaknesses
  1. Although the overall experimental results are strong, additional ablation experiments would help clarify the individual contributions of the anchor balancing regularization and tensor low-rank constraint.
  2. The relationship between the theoretical balancing formulation and the final multi-view objective could be explained more explicitly, particularly regarding the normalization of the anchor assignment matrix.
Questions for authors
  1. Could the authors further clarify how the balancing property established in Theorems 1 and 2 is maintained when the anchor assignment matrix is defined as Q^{(v)}=(G^{(v)})^\top H^{(v)} in the final objective?
  2. Could the authors provide an ablation study to distinguish the contributions of the Schatten-p norm balancing regularizer and the tensor low-rank constraint?
  3. Could the authors further explain why the proposed sample-to-anchor representation is more effective than conventional anchor-to-sample representations, beyond the empirical performance improvements?
2 comments
AuthorsOpenReview ↗

Clarifications on Anchor Matrix Normalization, Theoretical Inheritance, and Additional Ablations

We sincerely thank the reviewer for the careful review and valuable comments. We respond point-by-point as follows.

Critical Weakness 1 / Question 2: Supplementing ablation studies to distinguish the contributions of anchor balance regularization and tensor low-rank constraint

We thank the reviewer for the suggestion. We agree that supplementing ablation studies will help more clearly elucidate the respective contributions of these two parts. We will add the relevant ablation studies in the revised manuscript. We expect that these supplementary ablation experiments will clearly demonstrate that both the balance regularization term and the tensor low-rank constraint make substantial contributions to the final clustering performance, thereby verifying the effectiveness of these two designs.

Critical Weakness 2 / Question 1: The relationship between the balancing formulation and the final multi-view objective, and the normalization of the anchor assignment matrix

First, regarding the normalization of the anchor assignment matrix and the preservation of the theorem's conditions. In the final objective, the anchor assignment matrix is defined as Q^(v) = (G^(v))^T H^(v). In the model constraints, we have explicitly required the sample label matrix H^(v) >= 0 and H^(v)1 = 1. To ensure that Q^(v) also satisfies the row sum of 1, we applied a column normalization operation when constructing the anchor graph G^(v) (i.e., ensuring the weight sum of each anchor is 1, (G^(v))^T 1 = 1). Based on this, we can strictly derive: Q^(v)1 = (G^(v))^T H^(v)1 = (G^(v))^T 1 = 1. This proves that Q^(v) strictly satisfies the condition of row sum being 1, and the balancing properties established in Theorem 1 and Theorem 2 can be directly applied to the final multi-view objective function. We will add this crucial detail in the revised manuscript.

Second, regarding the relationship between the balancing formulation and the final multi-view objective. The first term of the final multi-view objective function utilizes the conclusion of Theorem 2 view-by-view at the single-view level. By maximizing the Schatten p-norm of the anchor assignment matrix, it avoids the degenerate distribution of anchor labels and encourages the anchor space to evenly cover the structure of each cluster. Subsequently, the second term in the multi-view objective function, namely the tensor low-rank constraint, acts on the sample label tensor. It coordinates the H^(v) of different views, ensuring that each view learns cross-view consistent sample labels while pursuing its own anchor-side balance. Ultimately, consensus labels are obtained through average fusion. Therefore, the balancing formulation functions in the single-view anchor space, while the tensor low-rank constraint functions in the cross-view sample space. These two complement each other in the final objective function, jointly guaranteeing the balance and consistency of the multi-view clustering results.

Question 3: Why is the sample-to-anchor representation more effective than the traditional anchor-to-sample representation?

Traditional anchor-to-sample methods treat the anchor graph as a compact representation of samples and learn sample labels via GQ ≈ H. Essentially, they compress n samples into m anchors, where m << n. This creates an information bottleneck, making it easy to lose fine-grained structural information of the sample space, and the relationship between anchor assignments and sample labels is not explicitly modeled. In contrast, this paper adopts sample-to-anchor, which means using samples to represent anchors. Anchor labels are aggregated from sample labels via Q = G^T H. In this way, each anchor is represented by a weighted combination of multiple samples, which can leverage abundant sample information to enhance the representational capability of the anchors and alleviate the under-representation problem. Meanwhile, anchor labels and sample labels calibrate each other during optimization. The anchor graph is no longer just a compressed representation of samples, but a bridge connecting the sample space and the anchor space. Combined with the anchor-side balance regularization, the anchor space can evenly cover the structure of each cluster and avoid being dominated by the majority class. Therefore, it is more effective when the classes are imbalanced or the differences between views are significant.

We once again thank the reviewer for their constructive comments.

Reviewer_ZuqZOpenReview ↗

Thanks for your reply

I intend to increase my score

Reviewer_8ETBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies scalable multi-view clustering based on anchor graphs. The authors argue that existing anchor-based methods may suffer from under-representation by using a small number of anchors to represent a much larger set of samples, and may also overlook the consistency between anchor and sample distributions. To address these issues, the paper proposes a balanced multi-view collaborative label learning framework that uses sample labels to represent anchors, introduces a Schatten-pp norm based regularization to encourage balanced anchor assignments, and jointly refines sample and anchor labels. In addition, a low-rank tensor constraint is imposed on the sample label matrices to capture complementary information across multiple views. Experiments on several benchmark datasets show that the proposed method achieves competitive or superior clustering performance compared with existing multi-view clustering approaches.

Critical strengths
  1. Clear and intuitive motivation. The paper identifies two concrete limitations of existing anchor-based multi-view clustering methods: the potential under-representation caused by using a small number of anchors to represent many samples, and the lack of explicit consistency between anchor and sample distributions. The proposed reversal from “anchors representing samples” to “samples representing anchors” provides a simple and memorable perspective.
  2. A conceptually interesting balance regularization. The use of Schatten-pp norm maximization to encourage balanced and discrete anchor assignments is an interesting idea, and the paper attempts to support this design with theoretical analysis rather than introducing it purely as a heuristic.
  3. Strong empirical performance across multiple benchmarks. The method is evaluated on seven multi-view clustering datasets and compared with a broad set of recent baselines. The proposed method achieves strong results on both small- and large-scale datasets, and the comparison between shared and view-specific anchors also suggests some robustness to different anchor construction strategies.
Critical weaknesses
  1. Theorem 1 and Theorem 2 are developed in the QQ-space under the explicit constraint Q1=1Q\mathbf{1}=\mathbf{1}. However, starting from Eq. (14), the optimization is performed over HH with only H1=1H\mathbf{1}=\mathbf{1}. Since Q=G⊤HQ=G^\top H, the latter does not in general imply Q1=1Q\mathbf{1}=\mathbf{1} without additional assumptions on GG. Therefore, it is unclear whether the theoretical analysis can be directly inherited by the actual optimization problem.
  2. Some details of the theoretical analysis remain questionable. For example, when p=2p=2, the Schatten-pp norm reduces to the Frobenius norm, whose maximization under the simplex constraint does not by itself prevent all assignments from collapsing into a single cluster. In addition, the statement that the anchor labels can "naturally" be obtained as Q=G⊤HQ=G^\top H requires further justification, since this is not the general least-squares solution of the preceding formulation min⁡Q∣GQ−H∣F2\min_Q \\|GQ-H\\|_F^2.
  3. The authors state that "the learned anchors should also be distributed across these cc clusters, and should be as evenly distributed as possible among the clusters." This assumption may be reasonable for roughly balanced datasets, but it does not generally hold for imbalanced data. For highly skewed cluster distributions, enforcing uniformly balanced anchors may instead be inconsistent with the underlying sample distribution.
  4. The paper evaluates the method on multiple datasets and against many baselines, but most experiments focus on overall performance, parameter sensitivity, and convergence. There is no controlled ablation isolating the contributions of the reversed anchor representation, the Schatten-pp balancing term, and the tensor low-rank regularization. Therefore, it remains unclear which components are responsible for the reported gains and whether the experiments directly support the proposed mechanisms.
Questions for authors
  1. Could the authors clarify whether the constructed anchor graph GG satisfies any normalization condition that guarantees Q1=1Q\mathbf{1}=\mathbf{1} when Q=G⊤HQ=G^\top H?
  2. Could the authors clarify the case of p=2p=2, and explain whether Q=G⊤HQ=G^\top H is intended as a definition of anchor labels or as a solution derived from the preceding least-squares formulation?
  3. How does the proposed balance regularization behave when the underlying clusters are highly imbalanced? Does the method assume approximately balanced cluster sizes?
1 comment
AuthorsOpenReview ↗

Clarifications on Theoretical Rigor, Imbalanced Data, and Ablation Studies

We sincerely thank the reviewer for the highly professional and detailed review. These suggestions are of great help in improving the rigor of our paper. We respond point-by-point to the critical weaknesses and questions raised as follows.

  1. On the theoretical inheritance of Q1=1 and the normalization of G (Corresponding to Critical Weakness 1 & Question 1)

We need to clarify that the designs of the theorem and the objective function are sound; the key lies in the preprocessing operation in the experiments. In our actual experiments, the anchor graph G for each view is pre-constructed and fixed, and we apply column normalization to it, such that the sum of each column of G is 1 (i.e., G^T 1 = 1). Meanwhile, the model imposes a row normalization constraint on the sample label matrix H (i.e., H1 = 1). Therefore, the anchor assignment matrix Q = G^T H strictly satisfies: Q1 = G^T (H1) = G^T 1 = 1. We will add the column normalization condition of G and the corresponding derivation in the revised manuscript.

  1. On the degeneration at p=2 and the definition of Q = G^T H (Corresponding to Critical Weakness 2 & Question 2)

We sincerely apologize that the parameter range "0 < p <= 2" in the original manuscript is a typo, and the correct range should be "0 < p < 2". Furthermore, in our actual experimental settings, we strictly constrain the parameters to be 0 < p <= 1, thereby avoiding the degeneration issue at the source. We will correct this typo in the revised manuscript. Regarding the definition of Q = G^T H, it is a label representation directly assigned to the anchors by us, thereby achieving collaborative updating between sample and anchor labels during the optimization.

  1. On the assumption of anchor distribution under highly imbalanced data (Corresponding to Critical Weakness 3 & Question 3)

We need to clarify that this paper does not treat the "absolute uniform distribution" of samples as a hard constraint. Since the real-world datasets used (e.g., Reuters, AWA) are inherently class-imbalanced, forcing equal cluster sizes on the sample labels would distort the true data distribution and erroneously assign minority classes to majority clusters. Meanwhile, if the anchors are not constrained during optimization, the anchor space is highly susceptible to being dominated by the majority class, thereby drowning out the structural information of the minority classes. Therefore, the essence of maximizing the Schatten p-norm is to force the anchors to uniformly cover each cluster, preventing degenerate solutions. Since H and G^T H are collaboratively updated during optimization, the balanced distribution of anchor labels imposes a regularization constraint on the learning of H, thereby preventing the sample labels from degenerating into a trivial solution of a single cluster. In this way, even if the final output sample distribution is imbalanced (consistent with the real data prior), the model can still correctly identify the true clustering structure. Thus, the "anchor balance" in this paper does not replace "sample balance", but rather robustly guides sample clustering under imbalanced data through regularization constraints in the representation space.

  1. On supplementing controlled ablation studies (Corresponding to Critical Weakness 4)

We thank the reviewer for the suggestion. We agree that supplementing controlled ablation studies will more clearly elucidate the contributions of each regularization module. We will add the following two sets of comparative experiments in the revised manuscript: (1) Removing the tensor low-rank constraint to verify the contribution of cross-view consistency; (2) Replacing the balance regularization term by substituting it with the Frobenius norm to verify the effectiveness of the balance term.

Regarding the comparison of "Anchor-to-Sample", we have already conducted extensive comparisons with multiple representative baseline methods in this field (e.g., LMVSC, FPMVS-CAG, FMCNOF, etc.) in Tables 1 and 2. As stated in Section 2 of this paper, these traditional methods essentially belong to the Anchor-to-Sample paradigm that utilizes anchor graphs as compact representations of samples. The significant advantages of our method (Sample-to-Anchor) on these datasets have empirically demonstrated its effectiveness in alleviating the information bottleneck of m << n (the under-representation problem).

We once again thank the reviewer for their valuable time and constructive comments.

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

This paper proposes a balanced multi-view collaborative label learning method based on anchor representation. Firstly, the method replaces using anchors to represent samples by using samples to represent anchors, and designs an anchor balancing mechanism based on the Schatten p-norm. Secondly, the feasibility of this balancing mechanism has been theoretically analyzed. Thirdly, an alternately updating method for collaborative learning the sample labels and anchor labels is proposed.

Critical strengths
  1. The paper designs an anchor balancing mechanism based on the Schatten p-norm and theoretically analyzes the feasibility of this mechanism.
  2. Unlike existing methods that use anchors to represent samples, the paper uses samples to represent anchors.
  3. the paper proposes an alternately updating method for collaborative learning the sample labels and anchor labels.
Critical weaknesses
  1. In the clustering results, the anchor is balanced, but the sample is not balanced. Our goal should be for the sample to be balanced, not for the anchor to be balanced. 2.For each view, the method can obtain a label indicator matrix. which one is taken as the finally clustering result?
  2. In 4.4, a curve of the objective function value should be drawn to determine the convergence performance of the algorithm.
Questions for authors
  1. In Eq(14), why do we define v H's instead of defining one H?
  2. In 4.2, why does the View-specific Anchors performs better the Shared Anchors?
  3. The sample clustering result is balanced?
  4. In 4.4, Why are these curves not monotonic?
1 comment
AuthorsOpenReview ↗

Clarifications on Anchor-Sample Balance, Label Fusion, and Convergence Behavior

We sincerely thank the reviewer for the careful review. We respond point-by-point as follows.

Regarding "the anchor is balanced, but the sample is not balanced. Our goal should be for the sample to be balanced, not for the anchor to be balanced."

In real-world data (such as class-imbalanced data), imposing a "hard equal-cluster constraint" directly on sample labels would distort the true data distribution. Therefore, the "anchor balance" in this paper does not replace "sample balance" as the ultimate goal, but rather serves as a regularization mechanism at the representation level. During the optimization process, if the anchors are not constrained, the anchor space is highly susceptible to being dominated by the majority class and collapsing. The essence of maximizing the sp-norm is to force the anchor space to evenly cover the structure of each cluster, thereby preventing representation degeneration (avoiding sample labels collapsing into a single cluster). Through the collaborative updating of anchor labels and sample labels, this balance in the representation space will robustly guide sample clustering, enabling the model to accurately identify the true clustering structure even under skewed data distributions.

Regarding "Which label indicator matrix is taken as the final clustering result?"

The final result is not selected from a single view. Sample label matrices from all views are averaged to obtain a consensus label matrix, followed by row-wise argmax to derive cluster labels. Since the tensor low-rank constraint promotes cross-view consistency, this average fusion is natural and consistent with our model design.

Regarding "a curve of the objective function value should be drawn..."

Section 4.4 uses reconstruction error as the convergence criterion, a standard ALM practice. ALM alternately updates variables and penalty parameters to align auxiliary and original variables, with reconstruction error directly reflecting constraint satisfaction. Figure 3 shows that reconstruction error continuously decreases and stabilizes while ACC improves and stabilizes, indicating convergence. Thus, minimizing reconstruction error is a reliable criterion, and Figure 3 sufficiently demonstrates the convergence behavior.

Questions For Authors

Q1: Why define v H's instead of one H?

Because anchor graphs and structural information differ across views. A single shared label matrix would force premature identical labels, erasing complementary information. Each view's label matrix serves as a view-specific intermediate variable, with cross-view consistency achieved via the tensor low-rank constraint and final average fusion. This preserves view differences while obtaining a consensus label. These matrices are intermediate variables in collaborative learning, not final independent results.

Q2: Why do View-specific Anchors perform better than Shared Anchors?

Shared anchors concatenate all views before uniform selection, making them susceptible to noisy views and hard to adapt to individual distributions. View-specific anchors independently select anchors per view, flexibly preserving local structures. Collaborative learning and the tensor constraint then align the labels. Thus, when data scale increases and view differences are significant, view-specific anchors are advantageous.

Q3: Is the sample clustering result balanced?

We do not treat "strictly uniform sample distribution" as an objective, as benchmark datasets like Reuters, NUS, and AWA are inherently class-imbalanced. Achieving leading results on these datasets demonstrates that our model recovers meaningful cluster structures under imbalanced conditions rather than pursuing uniform sample quantities.

Q4: Why are these curves not monotonic?

The non-monotonicity arises from the ALM update mechanism, not a failure to converge. ALM optimizes an augmented objective that changes with the penalty parameter at each step, rather than the original objective. After updating H and D, the multipliers and penalty parameter are updated, altering the augmented objective. Thus, neither the original objective nor reconstruction error is guaranteed to decrease monotonically. Additionally, ACC is an external metric not involved in optimization, so local fluctuations are normal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.