acceptodds
Withdrawn from ICLR 2027

Harnessing Causality-Driven Meta-Learning for Imputation in Incomplete Multi-view Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In recent years, incomplete multi-view clustering algorithms (IMVC) have attracted extensive attention. In IMVC, the quality of missing-view imputation is a key factor determining clustering quality. However, existing IMVC approaches predominantly exploit statistical correlations and cross-view contrastive information for view completion. They fail to account for two intrinsic data characteristics: 1) the inherent causal relationships that define the logical connections between samples; 2) the superfluous features that lack discriminative power for downstream clustering tasks. This joint deficiency ultimately renders clustering predictions heavily dependent on extrinsic observed data patterns, resulting in suboptimal clustering performance. To address these issues, we propose a novel IMVC method based on Causal Meta-Learning Imputation (CMLI), which embeds causal knowledge into the imputation module orchestrated by meta-learning. Specifically, CMLI adopts meta-learning to distill critical causal information from features, thereby improving the discrimination of task-irrelevant information in the imputation process. Leveraging the Information Bottleneck (IB) principle, CMLI further optimizes the trade-off between compressing local features and preserving global features by simultaneously minimizing one mutual information term while maximizing another. Finally, the consistency learning between the label features effectively exploits cross-view clustering consistency and complementarity, resulting in a unified framework that couples clustering assignment with multi-view representation learning. Comprehensive experimental results validate that our proposed method surpasses state-of-the-art approaches.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
deep clustering, meta-learning, incomplete multi-view clustering, causal

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

The paper proposes CMLI, a framework for incomplete multi-view clustering (IMVC) that replaces conventional statistical imputation with a causal meta-learning imputation mechanism. The pipeline consists of: (i) view-specific autoencoders with an Information Bottleneck objective for compression and preservation; (ii) dual embeddings that construct view-specific causal graphs, which are refined via meta-learning into convolution kernels for imputing missing views; (iii) contrastive alignment between recovered and observed views; and (iv) an end-to-end cluster consistency learning module combining category-level contrastive loss and a KL-based high-confidence objective. Experiments on six datasets under multiple missing rates show consistent gains over eight IMVC baselines, with ablations and sensitivity analyses supporting the design.

Critical strengths
  1. Framing missing-view imputation as a causal discovery problem rather than a purely statistical correlation problem is a meaningful departure from existing DIMVC work. The critique that statistical imputation is vulnerable to confounders and view-private attributes is reasonable and well articulated.
  2. The overall framework is reasonably coherent. Combining representation compression, imputation, cross-view alignment, and end-to-end clustering is a meaningful design choice.
  3. The empirical evaluation is relatively broad, including six datasets, three missing rates, and multiple recent baselines. The proposed method achieves strong results on the main datasets, and the ablation study indicates that the imputation, information-bottleneck, and clustering-consistency components all contribute to performance.
Critical weaknesses
  1. The causal-graph construction is motivated by prior causal-discovery work using dual cause/effect-oriented embeddings. However, that mechanism was originally developed and empirically validated in a time-series Granger-causality setting, whereas CMLI constructs a graph over latent feature dimensions for clustering. The paper should more clearly explain why the causal interpretation transfers to this non-temporal feature setting, state the precise identifiability assumptions, and empirically distinguish the proposed causal graph from a generic learned dependency graph.
  2. The empirical results are reported as single values without standard deviations, confidence intervals, or the number of random runs. Since clustering outcomes and missing-pattern generation can be sensitive to random seeds, it is difficult to assess whether some improvements are statistically reliable.
  3. There are spelling errors and other non-standard writing issues present in the work.
Questions for authors
  1. The paper should more clearly explain why the causal interpretation transfers to this non-temporal feature setting, state the precise identifiability assumptions, and empirically distinguish the proposed causal graph from a generic learned dependency graph.
  2. The empirical results are reported as single values without standard deviations, confidence intervals, or the number of random runs. Since clustering outcomes and missing-pattern generation can be sensitive to random seeds, it is difficult to assess whether some improvements are statistically reliable. 3.What is the biggest difference from other current works? It's not very clear at the moment.
1 comment
AuthorsOpenReview ↗

Clarified Scope, Reliability, and Novelty

We thank the reviewer for the constructive feedback. We agree that the original causal framing was not sufficiently supported. We have therefore narrowed the paper's positioning and renamed our method GCDI, a predictive graph-conditioned dynamic-imputation framework rather than a causal discovery or causal-identification method.

Causal interpretation and dependency graph (W1/Q1). We agree that the previous formulation did not justify transferring a causal interpretation from time-series causal discovery to non-temporal latent feature dimensions. The revision therefore does not make this transfer. In revised Eq. (10), A(v)A^{(v)} is explicitly defined as an asymmetric predictive feature-dependency matrix, not a causal graph. It is not constrained to be acyclic and is only used to condition feature-specific dynamic kernels. Accordingly, we no longer assert causal identifiability assumptions, causal effects, or confounder adjustment.

Graph-conditioning ablation at missing rate 0.3 (ACC/NMI/ARI, %).

VariantSynthetic3dFashionLandUse-21
Fixed kernel73.83/41.68/35.7092.50/85.80/84.7518.24/18.05/2.87
Identity graph86.51/61.15/64.0690.67/84.37/81.9721.57/22.46/6.03
Symmetric graph88.50/64.17/69.4192.57/85.89/84.8926.14/30.56/12.58
Shuffled graph88.83/67.34/70.4685.41/82.66/77.1923.62/30.05/10.13
Full90.33/68.64/73.6793.51/87.71/86.8026.62/31.08/14.15

These results distinguish the proposed predictive graph conditioning from fixed, identity, symmetric, and randomly permuted dependency structures. They support predictive utility rather than a causal interpretation.

Statistical reliability and repeated runs (W2/Q2). We apologize that the original presentation did not clearly state the evaluation protocol. All clustering results are averages over 10 independent random seeds. Within each seed, all methods use the same missing-view mask, which controls for variation caused by incomplete-observation construction. We have clarified this protocol in Sec. 4.2. We agree that mean values alone do not establish statistical significance and now describe the comparisons as descriptive rather than statistically conclusive.

Writing and presentation (W3). We appreciate this observation and have carefully proofread the manuscript. We corrected spelling, grammar, notation, and non-standard phrasing throughout, and revised the terminology to consistently distinguish predictive dependency learning from causal inference.

Difference from existing work (Q3). The main difference of GCDI is not a causal claim. Compared with IMVC methods based on fixed predictors, global similarity structures, prototype completion, or generative recovery, GCDI learns asymmetric feature dependencies and uses them to generate feature-specific dynamic kernels for missing-view recovery. It further combines this adaptive imputation mechanism with information-bottleneck regularization, cross-view alignment, and cluster-consistency learning in one end-to-end objective. Section 2 now clarifies this positioning and distinguishes GCDI from both existing imputation-based IMVC methods and causally formulated multi-view methods.

Reviewer_vJcZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CMLI, an incomplete multi-view clustering (IMVC) framework that imputes missing views by combining causal modeling with meta-learning. It integrates view-specific autoencoders with an information bottleneck objective, a causal meta-learning imputation module that constructs view-specific causal graphs to recover missing views, a feature MLP for representation learning, and a label MLP with cluster consistency learning into a unified clustering objective.

Critical strengths

The paper's main strength is its unified end-to-end formulation, which jointly optimizes compression, imputation, feature extraction, and clustering under a single objective rather than relying on a separate representation-learning stage followed by k-means clustering. This design is consequential because it reduces the mismatch between representation learning and the final clustering objective.

Critical weaknesses
  1. The paper's central causal claims appear to be largely terminological rather than methodological. The causal graph in Equation (10) is constructed by embedding the same mean-imputed features twice, computing their outer product, and applying tanh-based normalization. This construction does not explicitly model directed causal relationships, enforce acyclicity, or account for confounders or exogenous variables. Thus, it remains unclear how the graph captures causal relationships rather than statistical associations.
  2. Although Section 3.3 refers to the method as "causal meta-learning," the meta-convolution kernel in Equation (11) is generated by a two-layer fully connected network applied to the causal graph. The method lacks task-level training and inner-loop adaptation, making it more akin to a hypernetwork than to a conventional meta-learning framework. The authors should provide a more precise justification for the use of the term "meta-learning."
  3. The performance gains over the baselines are often modest, approximately 1–2%, yet no statistical significance tests are reported. Without such tests, it is difficult to determine whether these improvements are statistically significant or merely attributable to random variation.
  4. The paper does not adequately discuss existing research on causal multi-view clustering, nor does it compare the proposed method with relevant causal multi-view clustering methods. Given that causal modeling constitutes a central contribution of the paper, this omission makes it difficult to assess the novelty of the proposed approach.
  5. The current ablation study removes only the entire imputation module, without directly examining the contribution of the proposed causal graph. In particular, the authors do not compare the proposed graph with alternative structures, such as correlation-based graphs. Consequently, the reported results do not establish whether the observed performance gains arise specifically from the claimed causal modeling mechanism.
Questions for authors
  1. The proposed causal graph appears to capture pairwise associations rather than causal relationships. What theoretical basis justifies interpreting the graph constructed in Equation (10) as a causal graph?
  2. Page 2 asserts that statistical methods are "incapable of confounder discrimination" and that the proposed approach addresses this issue. Yet the method includes no confounder variable, backdoor adjustment, intervention, or counterfactual reasoning. Where in Equations (9)–(16) is confounding actually addressed?
  3. What are the tasks, task-level objectives, and inner- or outer-loop adaptation steps that constitute meta-learning in the current formulation?
  4. Since missing-view recovery is a central component of the proposed method, can the authors quantify the discrepancy between the causally imputed views and the corresponding ground-truth views on samples where complete views are available? Additionally, how does the imputation error affect downstream clustering performance?
  5. How does the proposed method compare with the baselines in terms of training time and memory usage, and how well does it scale as dataset size increases?
1 comment
AuthorsOpenReview ↗

Clarifications on Scope and Additional Empirical Evidence

We thank the reviewer for identifying the central issue in the original positioning. We agree that the previous formulation did not justify causal or conventional meta-learning claims. The revision narrows the positioning of our method (now termed as GCDI) to a predictive graph-conditioned dynamic-imputation framework; we no longer claim causal effects, confounder handling, interventions, counterfactuals, causal identification, or task-level meta-learning.

Causal interpretation and confounding (W1/Q1-Q2). We agree that the previous construction did not establish directed causal relations, acyclicity, confounder control, or causal identification. In the revision, Eq. (10) is explicitly defined as an asymmetric predictive feature-dependency matrix, not a causal graph. It only conditions the parameter generator for feature-specific dynamic kernels. We removed the claim regarding confounder discrimination and do not claim adjustment, intervention, or counterfactual reasoning.

Use of “meta-learning” (W2/Q3). The original method has neither task-level training nor inner-loop adaptation. We therefore removed “causal meta-learning” and related terminology. The two-layer module in Eq. (11) is now described as a lightweight parameter generator conditioned on the learned dependency matrix.

Performance variation (W3). We apologize for not clearly stating that all clustering results are averages over 10 independent runs. Within each run, all methods use the same missing-view mask. We have clarified this in Sec. 4.2. Since means alone do not establish statistical significance, the revision describes these comparisons as descriptive rather than statistically conclusive.

Relation to causal multi-view literature (W4). Section 2.2 now distinguishes causally formulated multi-view methods, which require stronger structural assumptions and inferential targets, from our predictive missing-view recovery objective. We therefore no longer claim novelty as a causal method or treat causal methods as directly interchangeable baselines.

Graph-mechanism ablation (W5). Table 3 directly compares alternative graph-conditioning and kernel-generation mechanisms. Values are ACC/NMI/ARI (%).

VariantSynthetic3dFashionLandUse-21
Fixed kernel73.83/41.68/35.7092.50/85.80/84.7518.24/18.05/2.87
Identity graph86.51/61.15/64.0690.67/84.37/81.9721.57/22.46/6.03
Symmetric graph88.50/64.17/69.4192.57/85.89/84.8926.14/30.56/12.58
Shuffled graph88.83/67.34/70.4685.41/82.66/77.1923.62/30.05/10.13
Full90.33/68.64/73.6793.51/87.71/86.8026.62/31.08/14.15

These results support the predictive utility of learned asymmetric dependencies and dynamic kernel generation, not a causal interpretation.

Recovery fidelity (Q4). Appendix H compares GCDI with mean imputation. On Fashion, relative error decreases from 0.947 to 0.768 and cosine similarity increases from 0.324 to 0.796; on LandUse-21, they improve from 0.965 to 0.939 and from 0.283 to 0.674, respectively. Table 2 further shows that removing imputation and alignment degrades clustering performance.

Ablation of loss components at missing rate 0.3 (ACC/NMI/ARI, %).

LI\mathcal{L}_ILIB\mathcal{L}_{IB}LC\mathcal{L}_CSynthetic3dFashionLandUse-21
×✓✓57.17 / 25.19 / 26.6849.73 / 46.90 / 20.3317.38 / 18.49 / 1.97
✓×✓88.00 / 63.58 / 68.4782.35 / 82.69 / 75.9624.29 / 29.54 / 13.52
✓✓×88.33 / 64.98 / 69.3348.40 / 62.11 / 26.9612.81 / 16.28 / 2.60
✓✓✓90.33 / 68.64 / 73.6793.51 / 87.71 / 86.8026.62 / 31.08 / 14.15

Efficiency and scalability (Q5). Appendix G reports training time, inference time, and peak memory on six datasets. GCDI incurs higher offline training cost but achieves lower inference latency and comparable inference memory than DCG. These are empirical results across datasets of varying sizes; we do not claim a formal asymptotic scalability analysis.

Reviewer_UUfzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Aiming at the problem of incomplete multi-view clustering, this paper claims to use meta-learning to incorporate causal structure into imputation and complete clustering using information bottleneck and cluster consistency. The causal graph is implemented as the normalized affinity of two embeddings of features, and the CGCG is mapped to the weights and biases of the convolutional kernel by two fully connected layers. The experiments cover six datasets and missing rates of 0.1/0.3/0.5.

Critical strengths

The motivation is clear, and the writing is clear and easy to read. Some necessary experiments are conducted to verify the effectiveness of the proposed method.

Critical weaknesses

This paper claims to use meta-learning to incorporate causal structure into imputation and complete clustering using information bottleneck and cluster consistency. However, the causal graph is implemented as the normalized affinity of two embeddings of features, and the CGCG is mapped to the weights and biases of the convolutional kernel by two fully connected layers. There is no causal identification proof, no task distribution and inner-outer optimization. Therefore, the technical contribution of this paper may not be related to causal-driven meta-learning.

Questions for authors
  1. Causal claims require identifiable structural learning objectives, but the causal and meta-learning motivations are not formalized into a formal problem, and it is not clear what the causal variables to be identified are, what the task distribution to be adapted is, and why ordinary imputation is not enough. "Assume that ZvZ^v satisfy causal identifiability" is followed by no conditions to make the assumption true.
  2. The CUB dataset shows that the ACC and ARI are better when the missing rate increases from 0.1 to 0.3, while other methods do not show this phenomenon. Although the author mentions this phenomenon in the paper, it is not clear whether the result is the average of multiple experiments. It is recommended to supplement the average and standard deviation of multiple experiments.
  3. It is recommended to supplement the complete ablation experiments, including the ablation experiments results under other missing rates, and the complete dataset.
  4. The parameter sensitivity analysis generally requires not only the parameter selection range, but also the parameter selection basis. The parameter settings in this paper may lead to data leakage, and how to set parameters on new datasets is not clear. The figure in the appendix shows that the performance of CMLI fluctuates greatly under different parameters, so the conclusion in the main text is not accurate.
  5. The visualization analysis can be compared with other methods, it is recommended to supplement the visualization comparison with other methods.
  6. The relative percentage like 3.27%, 4.91%, 5.47% should be changed to absolute difference.
  7. The author did not perform efficiency analysis, it is recommended to supplement the efficiency analysis, including the comparison of training time and inference time.
  8. It is recommended to supplement the DIMVC method without imputation for comparison.
2 comments
AuthorsOpenReview ↗

Clarifications and Additional Analyses

We thank the reviewer for the constructive feedback. We agree that the previous causal and meta-learning framing was stronger than what the formulation supports. We have therefore narrowed the paper’s positioning and renamed our method GCDI, a predictive graph-conditioned dynamic-imputation framework rather than a method for causal discovery, causal identification, confounder adjustment, or conventional meta-learning.

Causal interpretation and formulation (Q1). We agree that the previous version did not define causal variables, causal estimands, a task distribution, or inner-/outer-loop adaptation. These claims have been removed. In revised Eq. (10), A(v)A^{(v)} is an asymmetric predictive feature-dependency matrix generated by separate source and target projections. It only conditions the feature-specific kernels in Eq. (11), is not constrained to be acyclic, and is not interpreted as a causal graph.

Graph-mechanism ablation at missing rate 0.3 (ACC/NMI/ARI, %).

VariantSynthetic3dFashionLandUse-21
Fixed kernel73.83/41.68/35.7092.50/85.80/84.7518.24/18.05/2.87
Identity graph86.51/61.15/64.0690.67/84.37/81.9721.57/22.46/6.03
Symmetric graph88.50/64.17/69.4192.57/85.89/84.8926.14/30.56/12.58
Shuffled graph88.83/67.34/70.4685.41/82.66/77.1923.62/30.05/10.13
Full90.33/68.64/73.6793.51/87.71/86.8026.62/31.08/14.15

The results support the predictive utility of asymmetric graph conditioning and dynamic kernels, rather than a causal interpretation.

CUB non-monotonicity and repeated runs (Q2). All Table 1 results are means over 10 independent seeds; within each seed, all methods use the same missing-view mask. We clarified that the CUB increase from missing rate 0.1 to 0.3 is not evidence that more missingness improves clustering, but may reflect missing-pattern and optimization variability. Since only means are reported, the revised paper describes these comparisons as descriptive rather than statistically conclusive.

Ablation completeness (Q3). We appreciate the suggestion to evaluate all ablated variants over every missing rate and the complete-data setting. The full model is already evaluated on all six datasets and three missing rates (0.1/0.3/0.5), whereas the ablation is designed as a controlled component analysis at the representative intermediate rate of 0.3: one objective is removed while all other settings are fixed. This design isolates each component's contribution without conflating it with changes in missingness.

The full objective is best in all nine dataset-metric entries below. For example, removing LI\mathcal{L}_I reduces ACC from 90.33 to 57.17 on Synthetic3d and from 93.51 to 49.73 on Fashion. We agree that a complete 6×36\times3 ablation grid would provide additional breadth; accordingly, we restrict our conclusions to the evaluated settings rather than claiming exhaustive verification.

LI\mathcal{L}_ILIB\mathcal{L}_{IB}LC\mathcal{L}_CSynthetic3dFashionLandUse-21
×✓✓57.17/25.19/26.6849.73/46.90/20.3317.38/18.49/1.97
✓×✓88.00/63.58/68.4782.35/82.69/75.9624.29/29.54/13.52
✓✓×88.33/64.98/69.3348.40/62.11/26.9612.81/16.28/2.60
✓✓✓90.33/68.64/73.6793.51/87.71/86.8026.62/31.08/14.15

Parameter selection and potential leakage (Q4). λ1\lambda_1 and λ2\lambda_2 are fixed across seeds and missing rates rather than retuned for each condition. The revised sensitivity analysis states that GCDI is stable within a moderate range, not insensitive to hyperparameters. We will further make the selection protocol explicit.

Visualization comparison (Q5). Fig. 3 now directly compares GCDI with DCG and DCP on Fashion at missing rate 0.3. Appendix F provides the same raw-vs-GCDI protocol on additional datasets.

Absolute improvements (Q6). We corrected the wording: improvements such as 1.42, 2.33, and 2.81 are absolute percentage-point differences, not relative percentages.

Efficiency analysis (Q7). Appendix G compares GCDI and DCG on six datasets. GCDI has higher offline training cost on five datasets but lower inference latency on all six (69.87 ms vs. 163.88 ms on average), while maintaining comparable inference memory.

Imputation-free comparison (Q8). Thank you for this suggestion. We would like to clarify that our compared methods already include APADC, which is explicitly formulated in its original paper as an imputation-free deep incomplete multi-view clustering method. Rather than reconstructing missing views, APADC projects available observations into a common feature space and performs distribution alignment for clustering. It therefore provides the requested imputation-free comparator. We have revised Section 4.2 to highlight this distinction.

Thanks for your rebuttal. Since using meta-learning to incorporate causal structures is no longer the main contribution of this paper, I argue that the motivation of this work lacks novelty. Could you provide further evidence, rather than speculation, to support your current claim that "dependencies differ across views and feature dimensions"? Furthermore, among the existing results, which ones directly demonstrate that GCDI improves model performance by resolving such discrepancies in dependency relationships?

Comments on the paper

1 comment

Dear Reviewers and Authors,

As a quick reminder, the author-reviewer discussion period will close on November 18.

@Reviewers: Please make sure to read the authors' rebuttals, acknowledge their responses, and engage in further discussion if any points remain unclear. While final decisions are not made at this stage, we ask that you carefully re-evaluate your ratings in light of the new clarifications and experimental results:

Increase your score if the authors have satisfactorily resolved your concerns and strengthened the paper.

Decrease your score if new, unaddressed issues have emerged during the discussion.

Maintain your score if critical concerns remain unresolved.

Please note that you are under no obligation to change your rating, provided your final score accurately reflects the current state of the manuscript. To ensure a fair evaluation, please ground your assessment in concrete evidence rather than personal preference, and consider whether the work offers meaningful scientific value to the ICLR community.

Thank you very much for your time, thoughtful engagement, and contributions to the review process.

Best regards,

Area Chair

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

Related papers

Discussion (0)

Sign in to comment.