acceptodds
Under review as a conference paper at ICLR 2027

Quality-to-Relation Learning for Multi-View Clustering under Heterogeneous Noise

Abstract

Quality-aware multi-view clustering (QAMVC) has recently been explored to address heterogeneous observation reliability by estimating data quality and using it to regulate representation learning and multi-view fusion. However, existing QAMVC methods often provide coarse-grained quality estimates based on limited quality cues, and mainly use them to adjust feature contributions or fusion weights, leaving the relational influence of unreliable observations insufficiently regulated. Moreover, weak relations may be suppressed or discarded even when they receive partial support from other views. To address these issues, we propose Cross-Granularity Quality-to-Relation Learning (CGQRL), a unified framework that establishes a path from sample-view quality to relational supervision. CGQRL first estimates dynamic sample-view quality by jointly considering reconstruction fidelity, cross-view consistency, and cluster compatibility. The resulting quality scores are used to adaptively weight view-specific representations for consensus fusion and, more importantly, transferred to the relational level to characterize the reliability of sample neighborhood relations and cross-view assignment relations. Based on these quality-aware relations, we further develop a reliability-preserving structural alignment strategy that provides soft structural supervision: reliable relations receive stronger supervision, partially supported relations are down-weighted rather than discarded, and relations unsupported across views are treated as explicit negatives. By jointly modeling observation quality and relational influence, CGQRL suppresses unreliable structural information while preserving potentially useful weak relations. Extensive experiments on multiple multi-view clustering benchmarks demonstrate the effectiveness of CGQRL under heterogeneous noise.

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

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

This paper addresses the multi-view clustering problem in the presence of heterogeneous noise, and proposes a new method called CGQRL, which aims to enhance the modeling of the relationships between samples by leveraging the quality of the features. Specifically, given multi-view data, CGQRL first measures the quality of the samples based on the dynamic quality estimation module. Then, the quality of the samples is incorporated into the subsequent sample relationship learning and cross-view relationship analysis to enhance the clustering performance.

Critical strengths
  1. This paper ingeniously combines the adaptive construction of neighbor relationships with feature quality, enabling the structured information to be dynamically updated and capable of handling multi-view scenes with noise.

  2. The experiments in the paper include a wide range of datasets and noise contamination rates, and the measurement indicators include ACC, NMI, ARI, and RI, providing compelling support for the effectiveness of the proposed method.

Critical weaknesses
  1. Many details of the experiments are not explained clearly. For instance, in different experimental procedures, are the settings for pollution the same?

  2. There is a lack of explanation regarding the reasons for the different performance of the method on different datasets.

Questions for authors
  1. For each experiment, it was conducted five times. What were the settings for these five experiments? What were the differences in the pollution settings and the model settings among the five experiments?

  2. As shown in Table 2, the improvement in clustering effect of the proposed method varies across different datasets. Why is this so?

  3. When the given dataset contains mostly reliable samples with a relatively low proportion of noise, how does the proposed CGQRL method perform?

2 comments
AuthorsOpenReview ↗

Experimental Protocol and Robustness

We sincerely appreciate your thoughtful questions regarding the construction of the experimental runs, the dataset-dependent performance gains, and the behavior of CGQRL under low corruption. These questions provide an important opportunity to clarify the fairness of our evaluation protocol and the conditions under which the proposed method is most effective. We address each question below.

Question 1. How are the five experimental runs constructed? Are corruption realizations independently resampled across runs, or do the runs share the same corrupted datasets and differ only in model initialization?

A1: Thank you for requesting this clarification. The corruption protocol approximates a realistic multi-view setting in which observation quality varies across samples and views, rather than imposing a binary division between entirely clean and entirely corrupted multi-view samples. Some sample-view observations remain unaffected, whereas others undergo mild to severe degradation caused by illumination changes, motion artifacts, sensor instability, or transmission noise.

We model corruption prevalence and severity separately. Within each view, a proportion ρ\rho of sample-view observations is selected for perturbation. Each selected observation receives an intensity α∈{0.2,0.4,0.6,0.8,1.0}\alpha\in\{0.2,0.4,0.6,0.8,1.0\} and is generated as

\qquad \epsilon_i^{(v)}\sim\mathcal{N}(0,I).$$ A smaller $\alpha$ retains more of the original observation, whereas a larger $\alpha$ represents more severe degradation. The protocol therefore varies both the location and severity of corruption at the sample-view level. For each dataset and corruption ratio, we generate five corruption realizations indexed by $s\in\{0,1,2,3,4\}$. The corruption seed determines the selected observations, perturbation intensities, and Gaussian noise values. Each realization is generated once, stored, and reused by all methods. Each method is trained once on each stored realization. In run $s$, all methods receive the same corruption file indexed by $s$, while the corresponding run-specific training seed controls parameter initialization and mini-batch ordering. Therefore, the five runs vary in both corruption realization and training randomness, while comparisons across methods remain paired under identical corrupted inputs. **Question 2. Could the authors explain why the performance gains differ across datasets?** **A2:** Thank you for raising this question. The gains vary across datasets because the available views differ in their structural information and in how strongly corruption disrupts the sample relations used for clustering. The datasets have markedly different configurations: HW2sources contains 2,000 samples represented by two views, whereas MSRCV16view contains 210 samples represented by six views; Mfeat also has six views but contains 2,000 samples. Thus, the number of observations available to establish relations, as well as the information supplied by each view, differs across datasets. Their measured responses to corruption are also distinct. As the corruption ratio increases from 0.1 to 0.7, CGQRL's ACC decreases by 13.10 points on HW2sources, 5.14 points on MSRCV16view, and 3.32 points on Mfeat. These differences suggest that the same corruption ratio does not disrupt the clustering structure equally across datasets, which helps explain why the observed gains vary. **Question 3. How does CGQRL perform when the corruption ratio is low and most observations remain reliable?** **A3:** Thank you for raising this question. The main experiments include $\rho=0.1$, and CGQRL remains competitive or superior in this regime. At $\rho=0.1$, CGQRL ranks first on 14 of the 24 dataset-metric combinations. It leads all four metrics on HW2sources, MSRCV16view, and MNIST-4, and achieves the best ACC and F1 on Mfeat. The gains are particularly clear on MSRCV16view, where CGQRL improves over the strongest baseline by 5.71, 7.50, 8.72, and 6.02 percentage points in ACC, NMI, ARI, and F1, respectively. On MNIST-4, the corresponding gains are 8.83, 5.71, 8.79, and 11.94 points. On Hdigit and COIL20, CGQRL remains close to the strongest baseline: its four metrics are within 1.00 point on Hdigit and within 1.02 points on COIL20. Thus, the method does not sacrifice performance when corruption is sparse, even when a strong baseline is already near saturation. These results show that the quality-to-relation mechanism remains effective when perturbations are sparse. Most sample-view evidence is unchanged, and the model preserves consistent relations while limiting the influence of perturbed observations. Gains are naturally smaller when baseline relations are already reliable or performance is near saturation.

Thanks for the response. After reading the response and the comments from other reviewers, I think my score is appropriate, and I am currently inclined to keep it as it is.

Reviewer_Y3NWOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CGQRL, a multi-view clustering framework that uses observation quality to regulate structural relations and cross-view information transfer. The method dynamically updates a candidate graph based on view-specific representations and assigns reliability weights to sample-level relations. It also introduces a reliability-scaled cross-view transfer objective to control the contribution of information exchanged between views. The framework is evaluated on six datasets under different noise conditions, with additional ablations and convergence analyses to examine its behavior.

Critical strengths

The dynamic graph and reliability-aware objectives provide a reasonable way to adapt relational learning to changing representations and heterogeneous observation quality. The framework considers both local sample relations and information transfer across views.

Critical weaknesses
  1. The sensitivity of the proposed framework to graph construction and optimization settings is not fully characterized.

  2. The computational cost and scalability of the additional graph and relation-learning operations are not sufficiently discussed.

  3. How sensitive is CGQRL to the neighborhood size (k) and graph refresh frequency?

  4. A brief complexity analysis would help clarify the computational overhead introduced by the proposed components.

  5. Could the authors provide an ablation study that removes each quality cue individually?

Questions for authors
  1. The sensitivity of the proposed framework to graph construction and optimization settings is not fully characterized.

  2. The computational cost and scalability of the additional graph and relation-learning operations are not sufficiently discussed.

  3. How sensitive is CGQRL to the neighborhood size (k) and graph refresh frequency?

  4. A brief complexity analysis would help clarify the computational overhead introduced by the proposed components.

  5. Could the authors provide an ablation study that removes each quality cue individually?

2 comments
AuthorsOpenReview ↗

Sensitivity Analysis (1/2)

We sincerely appreciate your constructive comments and questions. Since Questions 1 and 3 both concern sensitivity to graph construction and optimization settings, we address them jointly below through additional experiments on the neighborhood size and graph refresh frequency.

Question 1. The sensitivity of the proposed framework to graph construction and optimization settings is not fully characterized.

Question 3. How sensitive is CGQRL to the neighborhood size (kk) and graph refresh frequency?

A1 & A3: Thank you for raising these questions. The existing analysis varies λRP\lambda_{\mathrm{RP}} and τp\tau_p over five seeds. We additionally vary neighborhood size k∈{5,10,15,20}k\in\{5,10,15,20\} and graph refresh interval T∈{5,10,15,20}T\in\{5,10,15,20\}, keeping all other settings and paired corruption realizations fixed. The defaults are k=10k=10 and T=10T=10. Each entry reports ACC/NMI/ARI/F1 (%) over five seeds.

Neighborhood-size sensitivity

ρ\rhokkHW2sourcesMSRCV16viewMfeat
0.1595.00/89.08/89.27/95.0092.00/85.75/82.49/91.8087.67/84.43/80.54/87.14
0.11094.65/88.47/88.54/94.6690.76/84.27/80.19/90.4593.66/87.48/86.47/93.70
0.11594.54/88.39/88.34/94.5588.76/82.03/76.96/88.2690.92/86.07/83.83/90.80
0.12094.28/87.94/87.81/94.2987.14/80.22/74.52/86.5590.91/85.96/83.73/90.67
0.3591.99/83.10/83.13/91.9889.71/83.12/78.78/89.1392.54/85.86/84.23/92.58
0.31090.78/81.45/80.78/90.7891.52/84.71/81.56/91.3593.51/87.15/86.21/93.53
0.31591.34/82.13/81.87/91.3486.76/79.39/74.09/86.1592.67/85.97/84.51/92.71
0.32091.12/81.81/81.44/91.1286.29/78.37/73.18/85.6392.64/85.89/84.48/92.67
0.5586.98/73.97/73.43/86.9786.48/79.00/73.55/85.8791.50/83.65/82.09/91.52
0.51086.10/72.83/71.86/86.1088.57/79.04/75.82/88.3492.82/85.42/84.80/92.83
0.51586.77/73.84/73.09/86.7685.52/77.00/71.34/84.9891.69/83.89/82.49/91.72
0.52086.40/73.35/72.41/86.3983.71/75.10/68.73/83.0491.71/83.92/82.53/91.74
0.7580.96/64.33/62.66/80.9684.76/75.27/70.24/83.8289.22/79.56/77.75/89.23
0.71081.55/65.37/63.62/81.5685.62/73.89/69.98/85.2090.34/81.13/79.87/90.37
0.71580.53/64.19/61.93/80.5884.19/75.24/69.26/83.4989.65/80.19/78.54/89.66
0.72080.09/63.58/61.22/80.1283.14/74.01/68.01/82.7189.58/80.13/78.46/89.59

Graph-refresh-interval sensitivity

ρ\rhoTTHW2sourcesMSRCV16viewMfeat
0.1594.70/88.58/88.67/94.7088.29/82.04/76.81/87.8188.29/84.88/81.56/87.87
0.11094.65/88.47/88.54/94.6690.76/84.27/80.19/90.4593.66/87.48/86.47/93.70
0.11594.74/88.70/88.76/94.7488.29/82.11/76.85/87.8290.46/85.08/82.31/90.17
0.12094.72/88.69/88.72/94.7288.10/81.90/76.55/87.5890.42/85.06/82.27/90.13
0.3591.54/82.59/82.26/91.5587.81/81.06/76.08/87.2392.58/85.83/84.32/92.61
0.31090.78/81.45/80.78/90.7891.52/84.71/81.56/91.3593.51/87.15/86.21/93.53
0.31591.61/82.56/82.40/91.6187.71/80.84/75.86/87.1592.62/85.94/84.39/92.66
0.32091.58/82.53/82.34/91.5887.71/80.84/75.86/87.1592.65/86.00/84.44/92.69
0.5586.60/73.72/72.78/86.5986.95/78.65/73.38/86.5291.73/83.99/82.54/91.76
0.51086.10/72.83/71.86/86.1088.57/79.04/75.82/88.3492.82/85.42/84.80/92.83
0.51586.89/74.14/73.31/86.8886.76/78.39/73.21/86.3691.77/84.02/82.64/91.80
0.52086.89/74.08/73.32/86.8886.76/78.36/73.20/86.3491.71/83.94/82.52/91.74
0.7580.45/63.87/61.81/80.4884.38/75.03/69.35/83.7089.46/80.01/78.18/89.47
0.71081.55/65.37/63.62/81.5685.62/73.89/69.98/85.2090.34/81.13/79.87/90.37
0.71580.78/64.26/62.33/80.8284.57/75.20/69.62/83.9189.67/80.22/78.57/89.68
0.72080.84/64.21/62.40/80.8784.48/75.08/69.43/83.8089.60/80.17/78.44/89.62

Overall, k=10k=10 is a practical default, particularly under severe corruption, although k=5k=5 can be preferable in a few low-corruption settings. The default T=10T=10 achieves the highest ACC on MSRCV16view and Mfeat at all four ratios and leads on HW2sources at ρ=0.7\rho=0.7. Noticeable sensitivity is limited to a small subset of dataset-setting combinations; we will discuss these cases explicitly in the revision.

AuthorsOpenReview ↗

Complexity and Ablation (2/2)

We sincerely appreciate your valuable comments and suggestions. Since Questions 2 and 4 both concern the computational overhead and scalability of the additional graph operations, we address them jointly through a complexity analysis. We then answer Question 5 with a leave-one-quality-cue-out ablation.

Question 2. The computational cost and scalability of the additional graph and relation-learning operations are not sufficiently discussed.

Question 4. A brief complexity analysis would help clarify the computational overhead introduced by the proposed components.

A2 & A4: Thank you for requesting this clarification. For NN samples, VV views, and representation dimension dd, computing view-wise pairwise similarities for graph construction requires O(VN2d)\mathcal{O}(VN^2d) operations per refresh. The graph is first constructed after warm-up and then refreshed every 10 joint-training epochs; this full-graph operation is not repeated at every mini-batch update. Relation weights are subsequently used in the training objective. This analysis describes the overhead of the current full-dataset graph implementation; we do not claim experimentally verified scalability beyond the evaluated datasets.

Question 5. Could the authors provide an ablation study that removes each quality cue individually?

A5: Thank you for raising this question. We perform three separate ablations that remove reconstruction fidelity, cross-view consistency, or cluster compatibility from quality estimation. Each removed cue is excluded during warm-up, subsequent updates, graph refresh, and inference; the remaining cues are renormalized, while the underlying training objectives and network architecture are retained. All variants use the same datasets, corruption files, and paired seeds as the full model.

Leave-one-quality-cue-out ablation (%, five runs; each entry is ACC/NMI/ARI/F1)

ρ\rhoVariantHW2sourcesMSRCV16viewMfeat
0.5No reconstruction84.28/71.33/70.20/85.2987.43/78.52/74.03/86.9491.25/83.01/81.60/91.27
0.5No consistency85.16/72.39/72.83/87.1587.62/79.05/74.50/87.1991.77/84.03/82.67/91.79
0.5No cluster compatibility85.92/73.13/72.33/86.9186.86/78.72/73.28/86.3791.63/83.89/82.36/91.66
0.7No reconstruction77.02/62.03/57.76/77.3282.38/70.38/67.19/82.7587.22/78.04/75.85/88.23
0.7No consistency79.61/63.03/60.51/79.6484.67/75.55/69.65/84.0988.06/79.99/79.35/90.08
0.7No cluster compatibility80.66/64.05/62.09/80.7184.29/74.36/69.07/83.6189.63/80.17/78.52/89.64

The results are mixed across datasets and ratios. On Mfeat, removing any cue reduces all four metrics at both ratios. On MSRCV16view, all three removals reduce ACC and F1 at ρ=0.5\rho=0.5, whereas at ρ=0.7\rho=0.7 removing reconstruction raises some metrics. HW2sources also varies by setting and metric. This is consistent with our design: reconstruction fidelity, cross-view agreement, and cluster compatibility characterize distinct facets of reliability, and their utility depends on the available views and corruption patterns. The study therefore clarifies how each perspective contributes across datasets and corruption levels.

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

This paper addresses multi-view clustering (MVC) under heterogeneous noise, where each sample-view observation may be corrupted to a different degree. Existing quality-aware MVC methods use quality estimates mainly as feature/fusion weights and discard weak relations. The paper proposes CGQRL, which estimates dynamic sample-view quality by combining reconstruction fidelity, cross-view consistency, and cluster compatibility. The studied problem is significant, and the proposed method indicates certain novelty.

Critical strengths
  • The studied problem in this paper is significant, which is rarely studied by existing works.

  • The proposed countermeasures are technically solid and novel.

  • Comprehensive evaluation is conducted.

Critical weaknesses
  • The core design of the work is not fully validated with the experiment. To be specific, how could the authors demonstrate the estimated scores are able to find the corrupted observations?

  • The effectiveness of the estimated scores should also be attached with some analyses.

Questions for authors

Refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Validation of Quality Scores

We sincerely appreciate your important suggestion to directly verify whether the estimated quality scores can distinguish corrupted observations from clean ones. This suggestion helps us strengthen the empirical validation of the core quality-estimation component and clarify the intended interpretation of the learned scores. We address this concern below.

Weakness. The core design of the work is not fully validated with the experiment. To be specific, how could the authors demonstrate the estimated scores are able to find the corrupted observations? The effectiveness of the estimated scores should also be attached with some analyses.

Response: Thank you for raising this concern. We added a direct evaluation of whether the estimated observation-quality scores distinguish corrupted from clean sample-view observations. The corruption mask is used only for evaluation and is not given to the model during training. For each dataset and corruption ratio, we evaluate five independently generated corruption realizations with paired training seeds.

AUROC measures whether a clean observation receives a higher quality score than a corrupted one; 0.5 corresponds to random ranking. We also report the Spearman correlation between quality and the clean-observation indicator, and the mean quality of clean and corrupted observations.

Datasetρ\rhoAUROCSpearmanClean QCorrupted Q
HW2sources0.10.726 +/- 0.0090.242 +/- 0.0090.6510.532
HW2sources0.30.741 +/- 0.0060.427 +/- 0.0100.6770.550
HW2sources0.50.761 +/- 0.0100.578 +/- 0.0140.7060.570
HW2sources0.70.758 +/- 0.0080.688 +/- 0.0120.7260.596
MNIST-40.10.834 +/- 0.0070.357 +/- 0.0070.7190.506
MNIST-40.30.812 +/- 0.0030.554 +/- 0.0040.7220.526
MNIST-40.50.783 +/- 0.0030.631 +/- 0.0050.7250.562
MNIST-40.70.753 +/- 0.0050.662 +/- 0.0040.7240.591
MSRCV16view0.10.716 +/- 0.0140.236 +/- 0.0130.6970.591
MSRCV16view0.30.698 +/- 0.0150.387 +/- 0.0220.7050.608
MSRCV16view0.50.696 +/- 0.0140.519 +/- 0.0240.7200.623
MSRCV16view0.70.709 +/- 0.0120.657 +/- 0.0150.7340.634
Mfeat0.10.805 +/- 0.0150.327 +/- 0.0150.6990.520
Mfeat0.30.813 +/- 0.0090.553 +/- 0.0120.7170.536
Mfeat0.50.818 +/- 0.0050.688 +/- 0.0050.7400.564
Mfeat0.70.798 +/- 0.0130.752 +/- 0.0140.7480.594
COIL200.10.778 +/- 0.0090.297 +/- 0.0090.7510.628
COIL200.30.791 +/- 0.0060.515 +/- 0.0090.7750.641
COIL200.50.807 +/- 0.0080.673 +/- 0.0100.8020.662
COIL200.70.815 +/- 0.0070.799 +/- 0.0080.8250.684
Hdigit0.10.762 +/- 0.0050.279 +/- 0.0050.6690.529
Hdigit0.30.783 +/- 0.0050.493 +/- 0.0070.6920.538
Hdigit0.50.796 +/- 0.0060.640 +/- 0.0090.7160.551
Hdigit0.70.792 +/- 0.0030.743 +/- 0.0040.7370.575

Across all 24 dataset-ratio conditions, mean AUROC ranges from 0.696 to 0.834. Clean observations receive a higher mean quality score than corrupted observations in every condition, and the quality-cleanliness Spearman correlation is consistently positive. These results show that the estimated scores contain meaningful, though imperfect, information for ranking observation reliability. Because corruption masks are used only for evaluation, this analysis directly supports the quality-estimation component. We interpret the scores as continuous reliability weights and do not claim that CGQRL is a standalone corruption detector.

Thanks for your careful response, and I think my previous rating reasonable and will maintain it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.