acceptodds
Under review as a conference paper at ICLR 2027

Selection-Consistent Completion for Incomplete Multi-View Unsupervised Feature Selection

Abstract

Incomplete multi-view unsupervised feature selection aims to identify a compact subset of original features when some views are missing. Existing joint methods combine missing-view completion with feature selection, but completion may still rely on features that are ultimately excluded from the selected subset, creating a mismatch between the selected-feature representation and the information used for completion. We introduce selection–completion consistency, which requires missing-view completion to rely only on currently available selected features while allowing richer observed information to provide auxiliary supervision during training. Based on this principle, we propose Selection-Consistent Multi-View Feature Selection (SC-MVFS), which integrates exact-budget feature selection, probabilistic multi-view posterior fusion, and joint learning with a full-input teacher and a selected-input student. A conditional squared-risk decomposition further characterizes the predictive-information difference between selected and fully observed features. Experiments on eight real-world datasets and controlled synthetic data demonstrate competitive clustering performance. Matched controlled experiments further isolate the effect of using consistent feature inputs during training and final completion and identify the conditions under which this consistency becomes more important.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
incomplete multi-view learning, unsupervised feature selection, selection-completion consistency, missing-view completion

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_FWF2OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates incomplete multi-view unsupervised feature selection and identifies a potential inconsistency between feature selection and missing-view completion, where the completion process may rely on features that are ultimately discarded. To address this issue, the authors propose SC-MVFS, a framework that combines exact-budget feature selection with a teacher–student learning mechanism. The student performs missing-view completion using only the selected observed features, while the teacher leverages all available observed features to provide auxiliary supervision. The paper presents theoretical analysis and extensive experiments on real-world and synthetic datasets, demonstrating the effectiveness of the proposed approach and its competitive clustering performance.

Critical strengths

(1)Novelty and Research Significance. The paper identifies a potential inconsistency between feature selection and missing-view completion in existing incomplete multi-view feature selection methods and introduces the principle of Selection–Completion Consistency. This perspective offers a novel and meaningful insight into the problem and provides a new direction for jointly learning feature selection and missing-view completion. (2) Well-Motivated and Coherent Methodology. The proposed framework effectively integrates feature selection, probabilistic representation learning, and missing-view completion through an exact feature budget constraint and a teacher–student learning mechanism. Furthermore, the conditional squared-risk analysis provides a theoretical interpretation of the proposed consistency principle. (3)Comprehensive Experimental Evaluation. The paper evaluates the proposed method against several representative baselines on eight real-world datasets and examines its robustness under varying missing-view ratios and feature budgets. In addition, targeted consistency experiments, ablation studies, and controlled synthetic experiments provide further empirical evidence supporting the effectiveness of the proposed framework and its underlying motivation.

Critical weaknesses

(1)Potential Fairness Concerns in Baseline Evaluation. The paper adopts a unified clustering evaluation protocol. However, for methods such as C2IMUFS and UIMUFSLR, which do not directly produce completed views in the original feature space, mean imputation is employed. This strategy may not fully leverage the structural information learned by these methods, potentially affecting the fairness of the comparisons. Furthermore, the current evaluation does not sufficiently disentangle the contributions of feature selection quality and missing-view completion to the final clustering performance. (2)Lack of a Simpler Selection-Consistent Baseline. The central contribution of the paper lies in the proposed Selection–Completion Consistency principle. However, the current experiments do not sufficiently distinguish the benefits of this principle from those introduced by the additional teacher–student learning mechanism. In particular, the absence of a selection-consistent student-only baseline, without the full-input teacher or distillation losses, makes it difficult to assess the necessity and individual contributions of the additional components in the proposed framework.

Questions for authors

(1)Fairness of Baseline Evaluation. For methods such as C2IMUFS and UIMUFSLR, which do not directly produce completed views in the original feature space, the paper adopts mean imputation. Could the authors further justify this evaluation strategy? Would it be possible to provide additional comparisons under the baselines' original evaluation protocols or with a shared imputation strategy to better assess the fairness of the reported results? (2)Independent Contribution of Selection–Completion Consistency. Could the authors include a selection-consistent student-only baseline that does not rely on a full-input teacher or knowledge distillation losses, while maintaining the same feature budget and comparable model capacity and training settings? Such a comparison would help disentangle the benefits of the proposed Selection–Completion Consistency principle from those introduced by the additional teacher–student learning mechanism. (3) Necessity of Key Loss Components. The current ablation studies primarily investigate posterior matching, masked reconstruction, and additional view masking. Could the authors provide separate ablations for the assignment matching loss and mixture regularization, together with sensitivity analyses of their corresponding loss weights? These experiments would help clarify the individual contributions and necessity of these components within the proposed framework.

1 comment

We thank the reviewer for the constructive assessment and specific experimental suggestions. We completed the shared-completion comparison, the student-only baseline, and separate assignment/mixture ablations with loss-weight sensitivity analyses. We address each question below.

Q1. Can differences in completion procedures explain the comparison?

We evaluated every selected subset using common observed means and an independent Ridge completer (alpha=1). Ridge uses selected source coordinates, excludes the entire target view and predicts selected target coordinates. We reuse the original Table 1 subsets and five seeds, preserving its masks, native feature counts and clustering settings; both common completers use observed-fitted min-max preprocessing.

SC-MVFS retains the lowest mean NMI rank under both common completers among ten methods: 3.250 with common means and 3.250 with common Ridge, across eight datasets at MR=FR=0.3. The native evaluation has mean NMI rank 1.500. This supports selected-subset quality beyond our own decoder.

Q2. A selection-consistent student-only baseline.

We implemented the requested Student with the same neural architecture, initialization seed, natural observation masks and exact K as its paired Full, removing full-input warm-up, teacher targets, posterior/assignment distillation and mixture. Both Full and Student satisfy the selected-input completion constraint. Full achieves higher mean NMI on six of the eight datasets.

DatasetFullStudent only
3Sources11.60 ± 3.609.70 ± 4.79
100Leaves69.63 ± 0.9868.42 ± 1.40
BBCSport24.09 ± 5.4922.83 ± 5.18
BDGP55.90 ± 5.9755.10 ± 4.71
Caltech101-757.66 ± 0.5655.97 ± 1.28
CiteSeer15.08 ± 3.8916.09 ± 4.37
Handwritten84.22 ± 2.0383.34 ± 1.58
OutdoorScene52.12 ± 0.9152.86 ± 1.30

Q3. Separate assignment/mixture ablations and weight sensitivities.

As specified in Appendix A, the weights remain fixed globally rather than being selected separately using evaluation-set results. After incorporating these two additional ablation variants, SC-MVFS still achieves the best average NMI rank among all six variants.

Assignment weight

Weight3Sources100LeavesBBCSportBDGPCaltech101-7CiteSeerHandwrittenOutdoorScene
011.54 ± 3.8869.09 ± 0.9419.70 ± 4.7757.06 ± 5.3157.68 ± 0.4317.32 ± 4.6582.79 ± 3.3451.76 ± 1.15
0.0111.54 ± 4.2769.46 ± 0.7119.64 ± 4.6056.23 ± 4.5557.65 ± 0.4616.68 ± 4.4982.92 ± 1.5151.63 ± 1.30
0.19.13 ± 2.0570.06 ± 0.9318.95 ± 4.4956.11 ± 4.9757.64 ± 0.3414.55 ± 3.1884.22 ± 1.8451.90 ± 1.35
0.5 (default)11.60 ± 3.6069.63 ± 0.9824.09 ± 5.4955.90 ± 5.9757.66 ± 0.5615.08 ± 3.8984.22 ± 2.0352.12 ± 0.91
113.72 ± 7.0969.77 ± 1.0122.65 ± 2.0155.57 ± 5.9057.07 ± 0.3313.29 ± 4.8584.12 ± 2.4951.83 ± 1.11
29.29 ± 2.0669.35 ± 0.5123.74 ± 7.3956.12 ± 5.1256.86 ± 0.5211.05 ± 5.3483.92 ± 2.2851.57 ± 0.94

Mixture weight

Weight3Sources100LeavesBBCSportBDGPCaltech101-7CiteSeerHandwrittenOutdoorScene
016.29 ± 6.5369.11 ± 1.1420.31 ± 2.6156.44 ± 4.4356.38 ± 1.4816.90 ± 3.2183.23 ± 2.7751.94 ± 0.70
0.00112.61 ± 2.7169.94 ± 1.4217.16 ± 3.3655.02 ± 5.0957.70 ± 0.2816.44 ± 4.1982.02 ± 2.7152.15 ± 0.72
0.0114.10 ± 4.8269.37 ± 1.3021.29 ± 4.6857.05 ± 4.9357.72 ± 0.5214.64 ± 4.6382.63 ± 1.1352.09 ± 0.63
0.1 (default)11.60 ± 3.6069.63 ± 0.9824.09 ± 5.4955.90 ± 5.9757.66 ± 0.5615.08 ± 3.8984.22 ± 2.0352.12 ± 0.91
0.512.11 ± 6.2869.10 ± 0.8212.53 ± 4.4755.67 ± 7.2557.22 ± 0.9410.12 ± 4.5385.69 ± 1.0550.77 ± 1.49
18.60 ± 2.9867.79 ± 0.638.62 ± 3.4551.05 ± 4.1157.37 ± 1.897.39 ± 4.5485.84 ± 0.9450.46 ± 1.08
RankVariantAverage Rank
1SC-MVFS2.500
2No assignment matching3.250
2No mixture regularization3.250
4No masked reconstruction3.750
5No posterior matching4.000
6No reduced-view augmentation4.250
Reviewer_KxTnOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Selection-Consistent Multi-View Feature Selection (SC-MVFS) for incomplete multi-view unsupervised feature selection. The framework introduces selection-completion consistency, requiring missing-view completion to depend only on selected features. A full-input teacher provides posterior and assignment supervision, while a selected-input student performs inference and reconstruction under an exact global feature budget. Experiments on eight datasets evaluate clustering performance, completion consistency, and several training mechanisms.

Critical strengths
  • The paper explores the consistency between feature selection and missing-view completion, providing a reasonable motivation for restricting the final completion pathway to selected features.

  • The teacher-student framework distinguishes training-time auxiliary information from the selected features used for final inference and completion.

  • The experiments cover eight datasets with different missing rates and feature budgets, together with ablation studies and controlled completion experiments.

Critical weaknesses
  • The comparison protocol uses different completion strategies across methods. SC-MVFS employs its learned completer, whereas C2IMUFS and UIMUFSLR rely on mean-imputation adapters. Consequently, the reported improvements may reflect differences in completion quality rather than feature-selection effectiveness, limiting the fairness of the comparison.

  • The controlled experiments provide limited evidence for the central selection-completion consistency mechanism. In Table 6, the reconstruction improvements are small on most datasets, and the selected-input variant improves NMI on only five of eight datasets. These results do not sufficiently explain the substantially larger gains reported in the main comparisons.

  • The exact-K selector uses discrete view-budget allocation and per-view top-K operations in the forward pass, while the backward pass relies on a global soft surrogate that ignores these discrete operations. The paper does not sufficiently examine whether this approximation produces stable and reliable feature allocations during optimization.

  • The development-set criterion used for hyperparameter selection is insufficiently specified. In particular, it is unclear whether ground-truth labels are involved in selecting configurations for this unsupervised task. The search ranges and selection procedures for the baselines also require clarification to establish comparable tuning conditions.

Questions for authors
  • Do the reported improvements persist when all selected feature subsets are evaluated using the same independently trained, budget-matched completion model?

  • Can the authors provide paired statistical comparisons of the selected-input and full-input variants to establish the significance of the reported improvements?

  • How does Proposition 1 theoretically support the optimization and clustering effectiveness of SC-MVFS beyond the fixed-subset information restriction?

  • How stable are the learned per-view feature budgets across random seeds, and how well does the soft surrogate approximate the actual hard selection behavior?

  • What development-set criterion and hyperparameter search ranges were used for SC-MVFS and the baselines? Were ground-truth labels used during configuration selection?

4 comments

We thank the reviewer for the detailed questions and suggestions for controlled evaluation. We completed shared-completer comparisons, paired reconstruction analysis and selector diagnostics, and clarify the theoretical assumptions and configuration-selection procedure below.

Q1. A common, independent completer.

We evaluated all subsets with common observed means and independently fitted Ridge (alpha=1). Ridge reads selected source coordinates, excludes the entire target view from its predictors and predicts selected target coordinates; it does not reuse our decoder. Lower mean NMI rank is better:

Evaluation pipelineSC-MVFS mean rankBest other method / rank
Original Table 1 (native)1.500UIMUFSLR / 3.875
Common observed-mean3.250UIMUFSLR / 4.500
Common Ridge3.250Mean+CAE / 4.750

The comparison uses the original Table 1 subsets from eight datasets, ten methods and five seeds (42–46), at MR=FR=0.3. Native results reproduce Table 1; common completion uses observed-fitted min-max preprocessing with the original clustering settings and per-method random seeds. Original feature-count rounding is retained: on OutdoorScene, six ranking baselines select 374 features while the other methods select 375; all other dataset/method combinations match ceil(0.3D). SC-MVFS retains the lowest mean NMI rank under both common completers.

Q2. Paired C/F statistics.

The ten pairs per dataset match the split, missingness, initialization, warm-up, batches and view hiding.

Cross-view reconstruction. C has lower mean cross-view MSE on all eight datasets. Relative reductions are computed as 100*(F-C)/F:

DatasetC cross-view MSEF cross-view MSERelative reduction95% paired CI (C-F MSE)
3Sources1.9996472.0033180.18%[-0.004280, -0.003062]
100Leaves0.9721341.0346886.05%[-0.077997, -0.047110]
BBCSport2.6199042.6418280.83%[-0.022919, -0.020929]
BDGP1.0336521.0600932.49%[-0.031054, -0.021828]
Caltech101-70.5549180.5554600.10%[-0.001589, 0.000506]
CiteSeer0.9237990.9264140.28%[-0.003343, -0.001886]
Handwritten0.4689580.4695050.12%[-0.001314, 0.000220]
OutdoorScene0.6673680.6745061.06%[-0.011825, -0.002452]

MSE is the per-view standardized decoder error averaged across views and ten seeds. The intervals are unadjusted two-sided paired 95% t intervals for C-F; negative values favor C. Six of eight intervals are entirely below zero.

Q3. What does the theory say beyond fixed-subset reconstruction risk?

Proposition 1 specifies the information cost at fixed S and observation pattern. To connect posterior matching to clustering, fix S, preprocessing, the trained decoder and centroids. Let the decoder be L-Lipschitz, the teacher covariance satisfy Sigma_T <= Lambda I, and M_i be the teacher representation's second-nearest minus nearest centroid distance. For any delta>0:

mean[cluster disagreement] <= mean[M_i <= 2*delta] + (2*Lambda*L^2/delta^2)*mean KL(q_student || q_teacher).

Gaussian KL bounds squared posterior-mean displacement by 2LambdaKL. The decoder transfers that bound to representations; the triangle inequality preserves the nearest centroid outside small-margin or large-displacement cases. The unit-prior PoE permits Lambda<=1. This explains a conditional transfer mechanism.

Q4. Stability of per-view budgets and fidelity of the soft surrogate.

We evaluated per-view feature budgets across ten initialization seeds while fixing the data split and missingness mask. The table reports the mean ± sample standard deviation of KvK_v, with views listed in their dataset order.

DatasetsPer-view budgets KvK_v: mean ± SD
3Sources1024.4 ± 5.3; 1027.6 ± 7.0; 1026.0 ± 5.8
100Leaves20.0 ± 0.0; 18.0 ± 0.0; 20.0 ± 0.0
BBCSport954.0 ± 5.3; 962.0 ± 5.3
BDGP470.0 ± 0.0; 79.0 ± 0.0
Caltech101-748.0 ± 0.0; 40.0 ± 0.0; 231.0 ± 3.1; 276.3 ± 1.4; 260.4 ± 1.8; 274.3 ± 1.9
CiteSeer1097.2 ± 17.9; 1007.8 ± 17.9
Handwritten42.3 ± 0.5; 34.5 ± 0.5; 36.7 ± 0.5; 6.0 ± 0.0; 43.0 ± 0.0; 32.5 ± 0.5
OutdoorScene112.4 ± 0.8; 110.0 ± 0.5; 104.6 ± 0.7; 48.0 ± 0.0

The largest per-view coefficient of variation is approximately 1.78%, indicating limited budget variability under initialization changes in this setting. Every run satisfies the exact global budget, ∑vKv=K\sum_v K_v=K. Some constant allocations reflect view-capacity constraints, so stable counts alone do not establish optimal allocation.

To examine the surrogate, we fixed the trained network and evaluation mask, replaced a selected feature pp with an unselected feature qq, and compared the sign of the actual hidden-target reconstruction-loss change with the surrogate prediction gq−gpg_q-g_p, where gg is the loss gradient with respect to selector logits. Direction agreement counts matching signs among swaps for which both quantities have magnitude greater than 10−1210^{-12}.

The dataset-level mean agreement rates range from 84.4% to 93.4%. This provides empirical evidence that the surrogate often captures the direction of local hard-selection changes.

Q5. Development-set criterion and hyperparameter search ranges.

We used a 20/80 development/evaluation split and selected configurations using 0.5(ACC+NMI)0.5(\mathrm{ACC}+\mathrm{NMI}) on the development subset under the central setting MR=FR=0.3\mathrm{MR}=\mathrm{FR}=0.3. Ground-truth development labels were therefore used for configuration selection; sample class labels do not enter the training loss. All baseline methods also used the same development split and hyperparameter-selection criterion to ensure consistency in the tuning protocol.

For SC-MVFS, as specified in Appendix A, the main objective weights were fixed globally, and only λ2,1∈{10−5,10−3}\lambda_{2,1}\in\{10^{-5},10^{-3}\} was selected on the development subset. As specified in Appendix B, we followed the hyperparameter search ranges reported in the respective original papers for the specialized baselines. The selected configurations were then held fixed for subsequent control experiments.

I thank the authors for their detailed responses and additional experiments, which address some of my concerns. However, the framework remains overly complex, and the contribution of selection–completion consistency to the reported clustering improvements is still not sufficiently clear. Overall, I appreciate the authors' efforts and have decided to raise my rating.

Thank you for reassessing the paper and pointing out the issues that need further explanation. The main-table results evaluate the complete system. The controls provide a more specific account of the reconstruction-input mechanism. Appendix G.2 holds the feature subset fixed and trains the completion models without the teacher, mixture model, matching losses, or learned selector. Selected-input training achieves lower mean completion MSE on all eight datasets. Table 6 retains the other training components and directly compares reconstruction using selected features with reconstruction using all available features. Appendix J further connects this intervention to feature allocation. At the highest shortcut strength, η=0.9, the selected-input variant allocates 53.8% of its budget to class-independent shortcut features, compared with 66.9% for the full-input control, while achieving an 11.51-percentage-point NMI advantage. These results come from paired comparisons on 20 independently generated synthetic datasets. The two variants share the same architecture, global budget, and optimization settings, and both use only their final selected features at inference. This provides controlled evidence linking the training-input intervention to changes in feature allocation and downstream clustering, while the magnitude of the benefit on real datasets remains dataset-dependent. Regarding complexity, the C/F control changes only the reconstruction inputs and adds no network, parameters, or objective terms. Student-only also satisfies the same selected-input requirement; the Full–Student comparison mainly evaluates the incremental benefit of the teacher and auxiliary objectives. In addition, SC-MVFS retains the lowest average NMI rank under both common observed-mean and independent Ridge completion, at 3.250 in each setting. This supports the usefulness of the selected subsets with other completers and complements the mechanism controls.

Reviewer_jbJfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SC-MVFS to address the mismatch between feature selection and missing-view completion. It uses a teacher-student framework to ensure completion relies only on selected features.

Critical strengths
  • The motivation is clear and addresses a reasonable issue in existing methods.

  • The experimental results are strong, with fairly comprehensive comparisons and controlled studies.

Critical weaknesses
  • The novelty is somewhat limited. Most components are standard, and the main idea is essentially enforcing train-test consistency.

  • The theoretical analysis is rather straightforward and provides limited insight into why the proposed method works better.

  • The improvements specifically attributable to selection-completion consistency are quite small on several datasets.

Questions for authors
  • Why is such a complicated framework necessary? Would simply training the completion model on selected features achieve similar results?

  • How much of the performance gain actually comes from the proposed consistency principle, rather than the teacher-student framework or other components?

  • The evaluation focuses on clustering performance, but how well does the method actually select informative and stable features?

3 comments

We thank the reviewer for the questions about framework complexity, performance attribution and feature quality. We added the student-only comparison and examined the reconstruction-source controls, common-completer results and feature-recovery and stability diagnostics to address these questions below.

Q1. Does the auxiliary framework improve over a student-only model?

We implemented the requested Student with the same neural architecture, initialization seed, natural observation masks and exact K as its paired Full, removing full-input warm-up, teacher targets, posterior/assignment distillation and mixture. Full achieves higher mean NMI on six of the eight datasets.

DatasetsFullStudent only
3Sources11.60 ± 3.609.70 ± 4.79
100Leaves69.63 ± 0.9868.42 ± 1.40
BBCSport24.09 ± 5.4922.83 ± 5.18
BDGP55.90 ± 5.9755.10 ± 4.71
Caltech101-757.66 ± 0.5655.97 ± 1.28
CiteSeer15.08 ± 3.8916.09 ± 4.37
Handwritten84.22 ± 2.0383.34 ± 1.58
OutdoorScene52.12 ± 0.9152.86 ± 1.30

Q2. How much can be attributed to selected-source consistency?

First, Appendix G.3 (Table 6) compares two variants differing only in reconstruction inputs, while keeping the teacher–student framework and all other components unchanged. Consistent training reduces completion RMSE on 7/8 datasets (0.81% average relative reduction) and improves ACC and NMI on 5/8 datasets each. Second, Appendix G.2 removes the teacher and other joint-training components, showing that consistent completion reduces MSE on all eight datasets (14.9% average relative reduction). Third, Appendix J (Table 11) shows that, under matched synthetic settings, the NMI advantage attributable to consistency increases from 3.56 to 11.51 percentage points as the selection–completion mismatch becomes more pronounced. Meanwhile, the component ablations in Appendix H (Table 7) confirm the complementary contributions of posterior matching, masked reconstruction, and view augmentation. These results demonstrate that consistency provides an independent benefit in completion, while the overall clustering performance reflects its interaction with the teacher–student framework and other components.

Q3. Are the selected features informative and stable?

The independent-completer test re-evaluates the original Table 1 feature subsets over its five seeds: SC-MVFS has mean NMI rank 3.250 with common observed means and 3.250 with independently fitted Ridge, and retains the lowest mean NMI rank under both common completers. Both common pipelines use observed-fitted min-max preprocessing and the original clustering settings and per-method random seeds. Ridge reads only selected source coordinates and excludes the entire target view. This assesses subset quality using an independent completion model. In an additional synthetic feature-recovery experiment, relevant-feature precision is 1.000 at all three SNRs, versus approximately 0.446 for Random-K. Initialization Jaccard ranges from 0.241 to 0.578, indicating that selection stability varies across datasets.

Thank you for the additional experiments and clarifications. However, my main concerns remain.

The improvements over the student-only baseline are generally small relative to the reported standard deviations, and the full model does not consistently outperform it across datasets. Similarly, the gain directly attributable to selection-completion consistency is modest (0.81% RMSE reduction), making it difficult to justify the added complexity. The feature-quality experiments are helpful, but the reported Jaccard scores also suggest limited selection stability on some datasets.

Overall, while the additional results are useful, they do not substantially change my assessment of the methodological novelty or the necessity of the proposed framework. I will maintain my original rating.

Thank you for the follow-up. The student completion branch in Full and the Student-only model both use selected features, so their comparison mainly evaluates the benefit of auxiliary training. The C/F control examines the effect of training completion with features that will be discarded before inference. This change in reconstruction inputs requires no additional network, parameters, or loss terms. Table 6 reports a 0.81% average relative RMSE reduction. Appendix G.2 examines the same input-mismatch issue in a simpler setting: the feature subset is fixed, and only the completion encoders and decoders are trained. Selected-input training achieves lower mean MSE on all eight datasets. Relative to Full-to-Selected completion, the reductions average 14.9% across datasets. These experiments use different settings and error metrics, examining input mismatch during joint learning and with a fixed subset, respectively. Appendix J shows a setting where input mismatch has a larger effect. The base SNR is fixed at 2.0, with paired comparisons on 20 independently generated synthetic datasets at each shortcut strength. As η increases from 0 to 0.9, the reported mean NMI advantage grows from 3.56 to 11.51 percentage points. This helps explain when training with subsequently discarded features can harm feature selection and clustering. The feature-quality and stability experiments also address different questions. SC-MVFS achieves the best average NMI rank among ten methods under both common mean and independent Ridge completion, at 3.250 in each setting. This shows that the selected subsets remain useful with other completers. However, the Jaccard results also indicate limited repeatability of exact feature identities on some datasets. Our contribution is to jointly model feature selection and completion using only selected inputs under a global feature budget, and to use fixed-subset and controlled synthetic experiments to examine when training–completion input mismatch affects reconstruction, feature allocation, and clustering.

Reviewer_Z1ZdOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies incomplete multi-view unsupervised feature selection, focusing on the information mismatch between missing-view completion and the final selected feature subset. To address this issue, the authors propose Selection-Consistent Multi-View Feature Selection (SC-MVFS). The method uses an exact global feature budget to constrain the number of selected features, employs a Product-of-Experts (PoE) model to fuse posterior distributions from available views, and adopts an alternating optimization scheme with a full-input teacher and a selected-input student. The student performs posterior inference and missing-view completion using only the selected features. The authors also provide a conditional squared-risk decomposition to analyze the difference in optimal prediction risk between using selected features and all observed features. Experiments on eight real-world datasets demonstrate competitive clustering performance, while ablation studies, controlled experiments, and synthetic data experiments are used to examine the role of selection–completion consistency.

Critical strengths
  1. The paper identifies the issue of selection–completion consistency. Existing joint methods may use features that are eventually discarded to train the completion module, leading to a mismatch between the final selected features and the information used for completion. SC-MVFS explicitly addresses this issue by restricting the student model to selected features, ensuring that the final completion process does not depend on discarded features.
  2. The experiments cover multiple datasets and different missing-view rates. In addition to standard ablation studies, the paper includes fixed-subset completion experiments, joint reconstruction-source controls, and synthetic experiments to examine the consistency between training inputs and final completion inputs.
Critical weaknesses
  1. The reconstruction objective and joint control experiments do not sufficiently clarify the source of the improvements in missing-view completion. The reconstruction loss includes both same-view and cross-view reconstruction, making it difficult to determine whether the performance gains mainly come from missing-view completion or autoencoding with sparse inputs. This weakens the evidence supporting the effectiveness of the proposed mechanism.
  2. The hyperparameter selection procedure in the unsupervised setting is not sufficiently explained. It is unclear whether ground-truth labels are used for model selection and whether the same hyperparameter tuning protocol is applied consistently across all baseline methods.
Questions for authors
  1. What exactly changes in the computation graph between variants C and F in the joint reconstruction-source control in Appendix G.3? The paper states that F changes the reconstruction branch to use all observed features from the same available views instead of only the selected features. However, it remains unclear whether this change affects the encoder input, introduces a separate reconstruction branch, or only modifies the decoder-side pathway.
  2. Does the current reconstruction objective adequately reflect cross-view missing-view completion? The reconstruction loss is computed over all naturally observed views, including source views that remain available as inputs. Therefore, the model may perform both same-view and cross-view reconstruction. Could the authors provide a control experiment in which the reconstruction loss is computed only on artificially hidden target views, and report the reconstruction errors for same-view and cross-view reconstruction separately?
  3. Were ground-truth class labels used for hyperparameter selection on the development set? Appendix A states that 20% of the data is used for configuration selection, while the remaining 80% is used for transductive evaluation. Could the authors clarify which metrics were used for hyperparameter selection and confirm whether all baselines were tuned under a fair and consistent protocol?
2 comments

Thank you for the assessment and the specific requests for clarification. We completed the requested reconstruction controls and clarified the selection protocol below.

Q1. What exactly changes between C and F?

The change is at the encoder input of the auxiliary reconstruction pass:

C: Q(selected_features, reduced_view_mask) -> decoder

F: Q(all_observed_features, reduced_view_mask) -> decoder

F reuses the same encoder and decoder parameters, with the same reduced-view mask; it introduces no separate network. The posterior and assignment branches retain the selected-input student. Teacher refresh, natural-pattern reconstruction, mixture and loss weights follow the same rules. Final inference uses each model's selected features.

Q2. Hidden-target-only reconstruction and separate errors.

We completed the requested auxiliary control: its targets are only originally observed views hidden by additional masking. The table reports actual decoder predictions, separating naturally observed (same-view) and naturally missing (cross-view) blocks:

DatasetsFull sameFull crossHidden-aux sameHidden-aux cross
3Sources0.84061.99960.84941.9987
100Leaves0.37880.97210.44660.9466
BBCSport0.88472.61990.91252.6037
BDGP0.47301.03370.57260.9841
Caltech101-70.44200.55490.46250.5602
CiteSeer0.84550.92380.86480.9207
Handwritten0.38190.46900.40360.4757
OutdoorScene0.35320.66740.40450.6632

Values are per-view standardized MSE averaged over ten seeds. Compared with Full, Hidden-aux achieves lower cross-view MSE on six of the eight datasets, while Full consistently achieves lower same-view MSE. These results suggest a trade-off between same-view reconstruction and cross-view completion, while showing that hidden-target-only supervision can improve cross-view completion on several datasets.

Q3. Labels, selection criterion and baseline tuning.

Development labels were used to select configurations by 0.5*(ACC+NMI), using the 20% development split; the remaining 80% is used for transductive evaluation. The split is stratified and the number of clusters is known. Sample class labels do not enter the training objective. All baseline methods followed the same hyperparameter-selection protocol, using the same development split and selection criterion.

I apologize for the delay. Thank you for the detailed clarifications. All three of my questions are addressed and I decide to maintain my score.

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

Loading the map…

Discussion (0)

Sign in to comment.