acceptodds
Under review as a conference paper at ICLR 2027

Direction-Preserving Anchor Learning for Multi-View Clustering

Abstract

Anchor-based multi-view clustering commonly represents samples using orthogonal anchors and simplex-constrained assignment weights, and performs reconstruction with squared Euclidean distance. We identify an overlooked problem in this common formulation: under orthogonal anchors, the norm of the assignment vector directly determines the norm of the reconstructed sample. Since squared Euclidean loss depends on both direction and magnitude, the norm of the assignment weights also participates in minimizing the reconstruction error. As a result, the assignment weights, which are intended to describe clustering structure, are additionally influenced by reconstruction magnitude, interfering with the learning of the clustering representation. This additional influence may also hinder the clustering representation from adapting well to different numbers of anchors, making the model more sensitive to the anchor number. To address these problems, we propose Direction-Preserving Anchor Learning (DPAL), which normalizes each anchor combination before computing the reconstruction error. This preserves the representation direction determined by the assignment weights while making the reconstruction error independent of their norm. DPAL requires no additional regularization trade-off parameter and admits efficient closed-form alternating updates. Experiments on ten multi-view benchmark datasets show that DPAL substantially improves over conventional Euclidean reconstruction models, remains competitive with recent anchor-based methods, and is more stable with respect to the number of anchors.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, Anchor learning, Representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vTETOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Direction-Preserving Anchor Learning (DPAL) for anchor-based multi-view clustering. It revisits the commonly used anchor reconstruction formulation and identifies that, under orthogonal anchors and simplex-constrained assignments, the norm of assignment weights determines the norm of reconstructed samples. Since Euclidean reconstruction depends on both direction and magnitude, the authors argue that assignment learning is affected by reconstruction scale in addition to clustering structure.

To address this issue, the paper proposes to normalize the anchor combination before reconstruction, preserving only the directional information of the representation. The resulting objective admits closed-form alternating optimization based on orthogonal Procrustes updates and normalized correlation-based assignment updates. Experiments on multiple multi-view clustering benchmarks demonstrate improvements over matched Euclidean reconstruction baselines and reduced sensitivity to the number of anchors.

Critical strengths
  1. The paper identifies a non-obvious coupling between assignment norm and Euclidean reconstruction when anchors are orthogonal. The decomposition of reconstruction error into scale and angular components provides a useful interpretation of why the proposed objective behaves differently from standard Euclidean reconstruction.

  2. The proposed method does not introduce additional regularization coefficients or complicated optimization procedures. The alternating optimization admits closed-form updates, making the method computationally attractive.

Critical weaknesses

1. The motivation relies on a restrictive formulation and its general relevance is unclear.

The main observation depends on two strong assumptions: orthogonal anchors and simplex-constrained assignments. However, many existing anchor-based multi-view clustering methods employ more flexible anchor learning formulations without strict orthogonality constraints or simplex assignment constraints. In these cases, the claimed assignment-norm issue may not exist in the same form.

The paper would be stronger if it could demonstrate that this phenomenon is prevalent in modern anchor-based MVC methods rather than mainly occurring in the specific formulation analyzed.

2. The paper does not sufficiently justify why assignment norm should be removed.

The paper treats assignment norm as an undesirable reconstruction-related factor. However, under simplex-constrained assignments, the norm of an assignment vector also reflects assignment concentration: ∥∣zi∥∣22=∑jzij2\||z_i\||_2^2=\sum_j z_{ij}^2.

For example, consider two assignments: z1=[1,0,0]z_1=[1,0,0] and z2=[1/3,1/3,1/3]z_2=[1/3,1/3,1/3]. The former indicates that the sample is confidently associated with one anchor, while the latter indicates ambiguity among multiple anchors. Their norms are substantially different: ∥∣z1∥∣2=1,∥∣z2∥∣2=1/3\||z_1\||_2=1, \||z_2\||_2=\sqrt{1/3}. Therefore, the norm component may encode useful clustering information such as assignment confidence or anchor concentration, rather than merely introducing an undesirable bias.

By removing the assignment norm from the reconstruction objective, DPAL may discard this information. The paper should provide stronger justification that directional information alone is sufficient or demonstrate cases where preserving assignment concentration is harmful for clustering.

Questions for authors
  1. How does DPAL perform when the orthogonality constraint on anchors is relaxed? Is the observed advantage preserved for more general anchor learning formulations?

  2. Since DPAL normalizes AziAz_i before reconstruction, it effectively changes the reconstruction geometry from Euclidean similarity to angular similarity. Could the authors compare with existing cosine/spherical clustering approaches to clarify whether the improvement comes from the proposed anchor learning mechanism or from adopting angular reconstruction?

  3. Many practical pipelines apply L2 normalization to input features before anchor learning. How significant is the assignment norm effect after feature normalization, and does DPAL still provide substantial improvement under commonly used preprocessing protocols?

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

The paper presents a simple and potentially useful modification to conventional anchor-based multi-view clustering. The geometric motivation and efficient optimization are appealing, and the controlled experiments provide encouraging evidence. However, the limited distinction from existing direction-based objectives, together with concerns about the evaluation protocol and the interpretation of the theoretical analysis, prevents me from being fully convinced of its contribution at the ICLR level. Addressing these concerns would substantially strengthen the paper.

Critical strengths
  1. The paper provides an interesting geometric interpretation of Euclidean reconstruction in anchor-based multi-view clustering. The analysis offers a reasonable motivation for reconsidering the conventional reconstruction objective.
  2. The proposed method is simple and computationally efficient. The closed-form updates for both anchors and assignment matrices are appealing, particularly for large-scale clustering applications.
  3. The experimental evaluation covers ten benchmark datasets. The controlled comparisons between Euclidean reconstruction and DPAL under different input normalization protocols are useful for understanding the contribution of the proposed objective.
Critical weaknesses
  1. The central operation of DPAL is to normalize the reconstructed anchor combinations, making the objective closely related to cosine-similarity-based or spherical reconstruction formulations. Although the paper provides a specific geometric motivation, the distinction from existing direction-based representation learning methods remains insufficiently discussed. A more explicit comparison with related formulations would help clarify the methodological contribution.
  2. Propositions 1 and 2 establish the relationship between assignment norms and Euclidean reconstruction, but the connection between this relationship and clustering degradation is primarily supported by empirical observations. Moreover, under the simplex constraint, the assignment norm and direction are inherently coupled. It would be helpful to clarify the precise sense in which DPAL removes the influence of the assignment norm and whether stronger theoretical insights into clustering stability can be obtained.
  3. Section 4.1 states that 30 k-means runs are performed for each learned representation, with the run achieving the highest ACC retained. Since ACC requires ground-truth labels, this selection procedure introduces label information into the evaluation process. The authors should clarify the protocol and consider reporting results using an unsupervised selection criterion, such as the k-means objective, to ensure a more appropriate evaluation.
  4. Although DPAL achieves substantial improvements over the matched Euclidean baselines, its advantages over recent competitive methods are less consistent. For example, the improvements on BBCSport are marginal, while several existing methods achieve better results on COIL20 and Wiki. More detailed comparisons, including statistical significance analysis and comparisons under consistent anchor numbers and preprocessing settings, would strengthen the empirical conclusions.
  5. The analysis of anchor utilization deserves further investigation. The authors use assignment norms and effective anchor numbers to explain the improved stability of DPAL. However, broader anchor utilization does not necessarily imply more discriminative clustering representations. It would be interesting to investigate whether the improved utilization is associated with better inter-cluster separation and intra-cluster compactness, rather than merely more balanced assignment weights.
Questions for authors
  1. Could the authors clarify the differences between DPAL and existing cosine-similarity-based or spherical reconstruction objectives?
  2. How is the anchor number selected for the main comparisons? Does the selection procedure require ground-truth labels?
  3. Would the conclusions remain consistent if k-means solutions were selected according to their clustering objective instead of ACC?
  4. Have the authors considered evaluating the normalized representation (Q) directly, rather than only the simplex-constrained representation (Z), to further examine the effects of normalization on downstream clustering?
Reviewer_nR4AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Direction-Preserving Anchor Learning (DPAL) for multi-view clustering. The authors identify a coupling between assignment weights and reconstruction magnitude in conventional Euclidean anchor reconstruction, which may interfere with clustering representation learning. To address this issue, DPAL normalizes anchor combinations before reconstruction to remove the explicit magnitude-fitting requirement. The method admits efficient closed-form alternating updates without additional regularization parameters. Experiments on ten benchmarks demonstrate improved clustering performance over matched Euclidean baselines and reduced sensitivity to anchor numbers.

Critical strengths

1.The paper identifies a meaningful limitation in conventional Euclidean anchor reconstruction, offering an interesting perspective beyond improving anchor structures or introducing additional constraints. 2.The proposed method is simple and well motivated. It directly addresses the identified issue without additional regularization parameters and allows efficient closed form updates. 3.The experiments are reasonably comprehensive. Controlled comparisons isolate the effect of reconstruction normalization, while analyses of anchor number sensitivity and anchor utilization further support the main claims.

Critical weaknesses

1.The experiments demonstrate improved clustering performance, but do not fully establish whether the gains primarily come from removing magnitude fitting. Controlled synthetic experiments separating directional and magnitude information would provide stronger evidence for the proposed mechanism. 2.The proposed objective is closely related to angular similarity maximization. The authors should clarify the distinction from existing spherical or cosine-based clustering methods to better establish the methodological novelty. 3.The paper selects the best result among 30 k-means runs using ground truth ACC. This introduces label information into unsupervised evaluation and may overestimate performance. Results using a selection criterion independent of ground truth labels would be more convincing. 4.The paper provides limited analysis of sensitivity to anchor initialization and noisy views. Additional robustness experiments would help assess the stability of the proposed method under different initialization and noise conditions. 5.The applicability of directional reconstruction is not fully discussed. Although DPAL removes explicit magnitude fitting, it remains unclear whether this modification is always beneficial, especially when feature magnitudes contain useful clustering information. Further experiments would help clarify the conditions under which DPAL is preferable to Euclidean reconstruction.

Questions for authors

See weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.