acceptodds
Under review as a conference paper at ICLR 2027

Topology-Aware Contrastive Anchors for Incomplete multi-view clustering

Abstract

Incomplete multi-view clustering remains challenging because missing observations disrupt cross-view correspondence and distort the underlying structural relationships among samples. We propose Topology-Aware Contrastive Anchors (TACA), a reconstruction-free framework that learns a unified representation directly from incomplete multi-view data through anchor-based contrastive modeling. Instead of explicitly recovering missing features, TACA constructs view-specific anchor representations and aligns their semantic information across views, thereby reducing the dependence on unreliable feature imputation. To further preserve structural information beyond a single anchor resolution, we introduce a topology-aware cross-scale consistency mechanism that coordinates multi-scale anchor graphs by encouraging agreement among their latent relational structures. Moreover, incomplete observations may generate semantically ambiguous negative pairs and weaken contrastive supervision. TACA addresses this issue through a negative sample decoupling strategy that suppresses unreliable negatives according to their semantic similarity. These components are jointly optimized to obtain discriminative and structurally consistent representations. Extensive experiments on eight benchmark datasets under different missing rates demonstrate that TACA consistently achieves competitive clustering performance against recent incomplete multi-view clustering methods while maintaining linear complexity with respect to the number of samples.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Contrastive Learning, Topology Consistency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_B5TLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

TACA addresses incomplete multi-view clustering (IMVC) via anchor-based contrastive learning, bypassing explicit imputation. It employs Cross-Scale Consistency (CSC) for multi-scale graph alignment and a hard-threshold Negative Decoupling (ND) strategy to filter false negatives, claiming linear computational scaling across five evaluated datasets.

Critical strengths
  1. By bypassing explicit imputation and directly aligning available observations at the anchor level, the method prevents noise accumulation from unreliable imputations at high missing rates.

  2. The CSC mechanism utilize multi-scale anchor graphs (e.g., m,2m,4mm, 2m, 4m) to enhance structural supervision beyond standard single-resolution approaches.

Critical weaknesses
  1. TACA is merely a mechanical amalgamation of mature, conventional techniques (multi-scale graphs, contrastive alignment, hard-threshold filtering η\eta) patched together via interpolation weights (γ\gamma) and regularizations.

  2. Relying on "shared observed landmarks" to define cross-view anchors is a fundamental flaw. Under severe missing rates (e.g., ρ=0.7\rho=0.7), finding enough shared landmarks is highly impractical, leaving disconnected views unaligned and severely restricting the method's real-world applicability.

  3. The theoretical claims (Theorems 3.1–3.3) appear hollow. The authors explicitly concede in the appendix that their proofs regarding spectral stability and gradient bounds neither guarantee improved clustering accuracy nor certify false-negative removal, severely undermining the genuine theoretical backing for the method's effectiveness.

  4. The claimed linear scalability misleadingly relies on a fixed anchor budget MM. Table 15 shows severe non-linear computational growth as anchors increase, meaning for large datasets, the model must either sacrifice representational capacity by fixing MM, or entirely invalidate its own efficiency claims by increasing it.

Questions for authors
  1. Does TACA's anchor mechanism collapse at extreme missing rates (e.g., ρ≥0.7\rho \ge 0.7) due to disconnected view graphs?

  2. Relying on a fixed hard threshold (η\eta) for negative decoupling is highly risky. Early in training, immature features may cause this hard truncation to permanently discard genuine "hard negatives," trapping the model in a suboptimal local minimum.

  3. Since the appendix concedes Theorems 3.2 and 3.3 guarantee neither improved clustering nor accurate false-negative filtering, what is their core value? Can these theoretical bounds be meaningfully linked to actual clustering recoverability?

  4. Since computational costs surge with the anchor count mm (Table 15), and a small, fixed mm cannot capture complex structures in massive datasets, how does TACA genuinely justify its claimed "high scalability" in practice?

Reviewer_8wjkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a reconstruction-free framework for incomplete multi-view clustering—Topology-Aware Contrastive Anchors (TACA). This method aims to address two issues in existing anchor-based clustering methods: structural inconsistency across different anchor scales, and unreliable negative sample supervision caused by semantically similar but mismatched anchors. Specifically, this method constructs multiscale anchor representations using available observational data, achieves semantic alignment through cross-view contrastive learning, and introduces cross-scale consistency constraints (CSC) and a negative sample decoupling strategy (ND) to enhance structural consistency and mitigate the exclusion of erroneous negative samples, respectively, while utilizing variance regularization to suppress representation collapse. Furthermore, the authors provide theoretical analyses regarding representation non-collapse, spectral stability, and gradient perturbations, and validate the method’s effectiveness on five benchmark datasets.

Critical strengths
  1. This paper addresses two key challenges in anchor-based incomplete multi-view clustering: cross-scale structural inconsistency and potentially unreliable negative supervision. By considering the impact of missing observations on cross-view semantic learning, the study presents a reasonable and practically meaningful motivation.

  2. TACA directly utilizes existing observable data for representation learning, without the need to generate features from missing views. This helps avoid, to some extent, introducing potentially unreliable reconstructed representations into the training process.

  3. Instead of directly aligning anchor representations with different dimensions, the paper enforces cross-scale structural consistency by comparing sample affinity matrices constructed at different anchor scales. Furthermore, an equivalent factorized formulation avoids explicitly constructing dense sample affinity matrices, thereby reducing memory overhead while effectively coordinating structural relationships across multiple anchor scales.

Critical weaknesses
  1. This paper assumes that non-matching anchors with high similarity are more likely to be false negative samples; however, high similarity does not necessarily imply that they belong to the same class—they may also correspond to hard negatives that possess discriminative value. Directly filtering out these samples may weaken the model’s discriminative ability. Although Theorem 3.3 proves the boundedness of gradient perturbations, it does not guarantee the semantic accuracy of negative sample filtering.

  2. This paper enhances structural consistency by minimizing the differences between affinity matrices of samples at different scales; however, structural consistency does not necessarily mean that the true clustering relationships are accurately reflected. Furthermore, overly strong consistency constraints may weaken the complementarity between different scales.

Questions for authors
  1. CSC objective aims to encourage similarity among the affinity matrices generated at different anchor scales. However, different scales may capture complementary structural information. How do the authors strike a balance between cross-scale consistency and scale-specific structural diversity? Could excessively strong CSC regularization lead to oversmoothing of the affinity matrix or reduce the discriminative power of the clustering representation?

  2. This paper uses a simple averaging strategy to fuse sample-anchor assignment results from different views; however, the clustering quality and reliability of these views may vary significantly. Could the authors explain why an equal-weighting fusion strategy that does not account for differences in view quality is justified? Furthermore, when some views contain severe noise or have weak semantic discrimination capabilities, can this method still maintain stable clustering performance?

  3. In Figure 2, using CSC alone results in a slight decrease in clustering ACC, while using the negative sample decoupling module alone yields only a very limited improvement. However, using the two together leads to a significant performance boost. Can the authors explain the reason behind this strong synergy? Specifically, does CSC improve the reliability of negative sample filtering, or does the negative sample decoupling module facilitate more accurate cross-scale structural alignment?

  4. The negative sample decoupling module filters out negative sample anchors with high similarity based on a fixed similarity threshold. How do the authors distinguish between false negative samples and difficult negative samples that belong to different classes but have high similarity?

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

This paper proposes TACA, an anchor-based framework for incomplete multi-view clustering without explicit feature reconstruction. The method combines cross-view anchor contrastive learning with cross-scale structural consistency and introduces a negative sample decoupling strategy to reduce unreliable contrastive supervision. Experiments on five benchmark datasets demonstrate competitive clustering performance and computational efficiency.

Critical strengths
  1. The reconstruction-free framework provides an efficient way to learn clustering representations from incomplete multi-view observations.
  2. The combination of cross-scale structural consistency and negative sample decoupling is reasonable, addressing both structural disagreement and unreliable contrastive supervision.
  3. The experiments cover different missing rates, ablation studies, and efficiency comparisons, providing useful evidence for the proposed design.
Critical weaknesses
  1. The methodological novelty is limited. The main components, including anchor-based contrastive learning, cross-scale structural consistency, and similarity-based negative sample decoupling, are closely related to existing techniques. Although their combination is reasonable, the paper does not sufficiently establish what new methodological insight it provides beyond previous work.
  2. The negative sample decoupling strategy relies on similarity estimates that may be unreliable during early training. The effect of inaccurate similarity estimates is not sufficiently discussed.
  3. The benefits of cross-scale consistency may depend on the selected anchor resolutions. More analysis would help explain when different scales provide complementary information.
Questions for authors
  1. What is the main distinction between the proposed cross-scale consistency objective and existing multi-scale structural alignment methods?
  2. How does negative sample decoupling behave when the learned representations are still unreliable?
  3. How sensitive is TACA to the choice and number of anchor scales?
1 comment

Without the authors' reply, I maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.