acceptodds
Under review as a conference paper at ICLR 2027

Acquisition-Oriented Incomplete Multi-View Clustering under Limited Budgets

Abstract

Incomplete multi-view clustering (IMVC) typically recovers missing view information from the observed data, overlooking practical scenarios where missing sample-view entries can instead be directly acquired at a cost. This raises a fundamental question: how should a limited acquisition budget be allocated to improve multi-view clustering? To address this problem, we propose Evidence-Adaptive Acquisition (EAA), a fully unsupervised framework for acquisition-aware incomplete multi-view clustering. Specifically, EAA first assesses whether the available observations in each view are sufficient to support reliable entry-level probabilistic modeling. With sufficient observations, EAA quantifies the acquisition value of each missing entry through its conditional information gain about the latent cluster assignment and selectively acquires the most informative entries. Under insufficient observations, EAA avoids unreliable fine-grained utility estimation and instead selects views by residuals and samples eligible missing entries uniformly within those views. The acquisition policy iteratively incorporates newly acquired entries to update the model and reassess the remaining candidates, enabling adaptive budget allocation under a strict acquisition budget. Extensive experiments on multiple datasets demonstrate that EAA effectively allocates limited acquisition budgets and improves clustering performance over existing acquisition and incomplete multi-view clustering methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering, active acquisition, budgeted acquisition, multi-view learning, unsupervised clustering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EhxXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes EAA (evidence-adaptive acquisition), a fully unsupervised framework for acquisition-aware incomplete multi-view clustering. It first evaluates whether the available observations in each view are sufficient to support reliable entry-level probabilistic modeling. Then it quantifies the acquisition value of each missing entry through its conditional information gain. The paper expect that EAA work well under a strict acquisition budget.

Critical strengths

It seems that several methods are compared by experiments.

Critical weaknesses

The manuscript has the following drawbacks. The method requires known ground-truth cluster number K. Actually, in the practical applications, K is almost always unknown. Moreover, for clustering algorithms, it is basic that K is unknown. If this is not obeyed and ignored, the algorithm is indeed unacceptable. Under the condition that K is known the experimental evaluation is unfair. 2. It also assumes that the acquired observations are noisefree. 3. Feature type must be given. In summary these drawbacks are serious and the corresponding assumptions are very improperly and even violates basic principles. In addition, it is claimed that the authors used generative AI tools to assist with code implementation, manuscript language editing and polishing, and table formatting. In my opinion, the reliance on the AI tools is too much.

Questions for authors

I hope that the authors comply by basic and important points. And my following concers are also provided for you: if limited budgets for clustering are significant and is a serious academic toptic, more informative analyses and demonstrations should be offered. Moreover, more convincing evidences to support the conclusions of the paper are expected.

3 comments
AuthorsOpenReview ↗

Assumption of a Known K

Dear Reviewer EhxX,

Thank you for your careful review and suggestions.

Q1. Number of clusters

Response. We thank the reviewer for raising this important concern. We acknowledge that the number of clusters is generally unknown in real-world applications, and estimating the cluster number is indeed an important issue in clustering.

However, we would like to clarify that the primary objective of this work is to investigate budget-constrained multi-view clustering, particularly how to effectively select informative view-instance pairs under a limited acquisition budget, rather than to address the additional problem of automatically estimating the number of clusters. Therefore, we treat the number of clusters KK as a given task parameter to focus on the central research problem of budget-aware view acquisition.

Furthermore, using the ground-truth cluster number as the input KK is a commonly adopted experimental protocol in clustering research. More importantly, all compared methods are evaluated under the same setting, ensuring a consistent basis for performance comparison. Although this setting does not fully reflect scenarios where the cluster number is completely unknown, it does not invalidate the evaluation of the proposed budget-constrained acquisition strategy.

We agree that extending the proposed framework to automatically estimate KK would further improve its applicability to real-world scenarios. However, this involves an additional research problem beyond the primary scope of this work. We will clarify the assumption and its implications in the revised manuscript and explicitly acknowledge unknown cluster-number estimation as an important direction for future research.

To examine sensitivity to a misspecified cluster count, we evaluated EAA on Mfeat, COIL20, Caltech101-20, and scene15 with KK set to approximately 0.75, 1.00, and 1.25 times the ground-truth class count, rounded to the nearest integer. The acquisition model and final clustering use the same working KK within each setting. The initial missing rate is 50%, and the acquisition budget is 30% of the initially missing entries, allocated over three rounds; all other experimental conditions are held fixed.

Table R3.1. Sensitivity to the working number of clusters

NMI (%), reported as mean ±\pm sample standard deviation over five fixed seeds. The ground-truth class counts are 10, 20, 20, and 15 for Mfeat, COIL20, Caltech101-20, and scene15, respectively.

\hline \text{Dataset}&K\\ \text{(low/base/high)}&\text{Lower }K&\text{Original }K&\text{Higher }K\\\\ \hline \text{Mfeat}&8/10/13&85.63\pm0.56&87.29\pm0.61&83.10\pm1.18\\\\ \text{COIL20}&15/20/25&80.01\pm0.98&83.43\pm0.85&79.40\pm0.55\\\\ \text{Caltech101-20}&15/20/25&58.37\pm1.52&59.12\pm0.97&59.71\pm0.76\\\\ \text{scene15}&11/15/19&42.49\pm1.04&42.08\pm0.43&42.08\pm0.79\\\\ \hline \end{array}$$ Across the tested cluster counts, EAA achieves NMI values of 83.10–87.29% on Mfeat, 79.40–83.43% on COIL20, 58.37–59.71% on Caltech101-20, and 42.08–42.49% on scene15. Caltech101-20 and scene15 show relatively small changes, while Mfeat and COIL20 have maximum decreases of 4.20 and 4.03 percentage points from their respective ground-truth-K settings. These results indicate a degree of tolerance to moderate changes in $K$ on these four datasets, with the extent varying by dataset.
AuthorsOpenReview ↗

Acquisition Noise, Feature Types, and AI Use

Q2. Noisy acquisitions

Response. The main experiments simulate acquisition by revealing the original feature values hidden by the observation mask. To examine additional measurement error, we added zero-mean Gaussian noise to newly acquired features. For each feature, the noise standard deviation is σ\sigma times the standard deviation of its initially observed values, with σ∈0,0.05,0.10,0.20\sigma\in\\{0,0.05,0.10,0.20\\}.

For a given seed, the same candidate entry has the same noisy feature values across policies. Once acquired, these noisy features are used in subsequent model updates, acquisition decisions, and final clustering. We use a 50% initial missing rate and a budget equal to 30% of the initially missing entries, allocated over three rounds. Each setting is evaluated over five seeds.

We compare EAA with Random, View-rate, and Uncertainty on Mfeat, COIL20, Caltech101-20, and scene15.

Table R3.2. Clustering performance with noisy acquisitions

NMI (%), reported as mean ±\pm sample standard deviation over five fixed seeds. Bold indicates the highest mean NMI in each row.

\hline \text{Dataset}&\sigma&\text{EAA}&\text{Random}&\text{View-rate}&\text{Uncertainty}\\\\ \hline \text{Mfeat}&0.00&\mathbf{87.29\pm0.61}&84.24\pm1.39&85.46\pm1.20&84.92\pm0.88\\\\ &0.05&\mathbf{86.70\pm1.11}&84.96\pm0.39&85.62\pm1.25&83.90\pm0.80\\\\ &0.10&\mathbf{86.83\pm1.08}&84.83\pm2.05&85.00\pm0.96&84.57\pm0.43\\\\ &0.20&\mathbf{86.63\pm1.38}&84.53\pm1.39&84.84\pm0.81&84.39\pm0.55\\\\ \text{COIL20}&0.00&\mathbf{83.43\pm0.85}&81.29\pm0.84&81.54\pm0.39&82.38\pm0.44\\\\ &0.05&\mathbf{83.64\pm1.17}&81.03\pm1.11&81.32\pm0.52&82.39\pm0.47\\\\ &0.10&\mathbf{83.40\pm0.78}&81.54\pm0.60&81.29\pm0.35&82.35\pm0.42\\\\ &0.20&\mathbf{83.67\pm1.33}&81.80\pm0.63&81.70\pm0.54&82.22\pm0.34\\\\ \text{Caltech101-20}&0.00&\mathbf{59.12\pm0.97}&56.70\pm0.56&56.42\pm1.44&55.90\pm0.52\\\\ &0.05&\mathbf{58.66\pm0.72}&56.11\pm1.38&56.33\pm1.53&55.54\pm1.06\\\\ &0.10&\mathbf{58.43\pm1.07}&55.96\pm0.68&55.77\pm1.84&56.05\pm0.72\\\\ &0.20&\mathbf{58.85\pm0.98}&56.71\pm0.95&56.03\pm1.61&56.39\pm0.96\\\\ \text{scene15}&0.00&\mathbf{42.08\pm0.43}&38.92\pm0.81&38.23\pm1.81&37.86\pm0.58\\\\ &0.05&\mathbf{41.62\pm0.70}&38.57\pm1.58&38.25\pm0.88&37.90\pm0.61\\\\ &0.10&\mathbf{41.69\pm0.61}&38.61\pm0.81&38.06\pm1.29&37.82\pm0.66\\\\ &0.20&\mathbf{41.37\pm1.01}&38.25\pm1.03&38.57\pm0.71&37.73\pm0.51\\\\ \hline \end{array}$$ EAA has the highest mean NMI among the four acquisition policies in all 16 dataset–noise settings. Across the tested noise levels, its NMI ranges from 86.63% to 87.29% on Mfeat, 83.40% to 83.67% on COIL20, 58.43% to 59.12% on Caltech101-20, and 41.37% to 42.08% on scene15. These results support the effectiveness of EAA under the tested levels of acquisition noise on these four datasets. Q3. Feature-type information Response. We thank the reviewer for raising this concern. In multi-view clustering, the feature representation of each view is part of the problem input. In our experiments, the compared methods operate on the same feature representations. EAA uses known properties of these representations, obtained from the feature-construction process, to specify the observation model for acquisition scoring. This model choice makes the representation assumptions explicit and does not require additional class annotations or access to evaluation scores. Our focus is how to allocate a limited acquisition budget across missing sample–view pairs given these feature representations. The observation models allow EAA to evaluate potential acquisitions, while the observed data determine their acquisition priorities. The required type information concerns how the features are represented and modeled; it supplies no supervision about sample class membership. We will clarify its source and role in the revised manuscript. Q4. AI assistance Response. We have disclosed our use of generative AI to assist with code implementation, language editing and polishing, and table formatting in the paper’s AI use statement. The authors are responsible for reviewing AI-assisted outputs and take responsibility for the methods, implementation, experimental results, and conclusions reported in the paper.

Thank you again for your time and careful review. We would be grateful for your feedback on whether our explanations have clarified the points you raised and if any concerns remain. We would be happy to provide further clarification.

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

This paper studies IMVC in a setting where some missing views can be acquired at a cost under a limited budget. It introduces EAA, which decides how to spend the budget based on how much observed data each view has. With enough observations, it ranks missing entries by information gain about the cluster assignment. With too few observations, it instead selects views by their residuals and samples entries within those views. The final clustering is obtained by spectral clustering on a fused graph. Experiments on 10 datasets show gains over several baselines.

Critical strengths
  1. The setting is new and practical. Treating missing views as something that can be acquired rather than only recovered is a reasonable question for incomplete multi-view clustering.
  2. The experimental protocol is fair. All methods use the same initial masks, the same budget, and the same final clustering step.
Critical weaknesses
  1. Table 2 lacks the two reference points needed to interpret the gains. There is no zero-budget row (the consensus-graph decoder on the initial mask) and no full-observation ceiling, although the masks are generated from complete data and both are cheap to obtain. Without them, the absolute benefit of a 30% budget and the remaining headroom cannot be read from the main table. Please add both as reference rows, excluded from the bold/underline ranking.
  2. The comparison does not include any IMVC method run with its own clustering pipeline. All IMVC competitors are converted into acquisition scores and decoded by the authors' pipeline, and Fig. 5 only tracks Random. The paper should include at least one strong IMVC method evaluated natively, both on the original mask and on the EAA-augmented mask, so that readers can judge acquisition against recovery and against the state of the art.
  3. EAA's provisional assignments come from the same kNN consensus graph and spectral embedding used for final evaluation (§3.2.1, §4.1), whereas adapted baselines such as CAL and ASCR select entries from their own representations and are then decoded by EAA's pipeline. This aligns selection and evaluation for EAA but not for the baselines, which may favor the proposed method. Two checks would help: report the adapted baselines with their native clustering pipelines, and repeat the main comparison with a different decoder, for example K-means on concatenated PCA features or a deep IMVC model, to verify that the gains persist.
Questions for authors

See the above weaknesses.

4 comments
AuthorsOpenReview ↗

Reference points and native IMVC evaluation

Dear Reviewer jNVo,

Thank you for your careful review and suggestions. We address the three points below.

Q1. Zero-budget and full-observation references.

Response. Thank you for pointing this out. We evaluated both reference settings using the same dataset-specific preprocessing and clustering configuration as in Table 2. We will add both settings as reference rows for all ten datasets and all three metrics—ACC, NMI, and ARI—and exclude these rows from the bold/underline ranking. Table R1 summarizes the NMI results.

Table R1. Zero-budget, EAA acquisition, and full-observation results.

Mean NMI (%).

\hline \textbf{Dataset}&\textbf{Zero budget}&\textbf{EAA, 30\\% budget}&\textbf{Full observation}\\\\ \hline \text{Yale}&46.93&63.98&65.81\\\\ \text{ORL}&66.17&83.79&88.66\\\\ \text{BBCSport}&39.84&90.37&92.68\\\\ \text{Reuters-1200}&26.19&27.22&17.80\\\\ \text{COIL20}&78.53&83.43&82.39\\\\ \text{Mfeat}&82.34&87.29&88.36\\\\ \text{Caltech101-20}&52.65&59.12&59.67\\\\ \text{scene15}&34.11&41.85&44.67\\\\ \text{FashionMNIST-30k}&60.59&67.00&69.28\\\\ \text{EMNIST-30k}&60.28&64.53&61.61\\\\ \hline \end{array}$$ The acquisition budget is measured relative to the initially missing sample-view entries. All acquisition comparisons below use a 50% initial missing rate; a 30% budget reveals 30% of those initially missing entries. **Q2.** *Native IMVC evaluation on the original and EAA-augmented masks.* **Response.** To examine how EAA acquisition affects existing IMVC methods, we evaluated HCLS-CGL and RMoGL on both the original and EAA-augmented masks using their native clustering pipelines. Each method receives two inputs: the original incomplete data and the data augmented by EAA acquisition. The named native method performs final clustering in both settings, with its parameters held fixed. EAA only selects the additional entries to acquire. **Table R2.** Native clustering on the original mask and the EAA-augmented mask. NMI (%), mean ± sample standard deviation. $$\begin{array}{lccc} \hline \textbf{Dataset}&\textbf{Native pipeline}&\textbf{Original mask 0\\%}&\textbf{EAA-augmented mask 30\\%}\\\\ \hline \text{ORL}&\text{HCLS-CGL}&76.57\pm1.10&84.28\pm1.00\\\\ \text{ORL}&\text{RMoGL}&76.91\pm0.72&88.76\pm0.95\\\\ \text{Mfeat}&\text{HCLS-CGL}&83.47\pm0.87&84.62\pm1.31\\\\ \text{Mfeat}&\text{RMoGL}&82.36\pm1.84&86.57\pm3.18\\\\ \text{BBCSport}&\text{HCLS-CGL}&39.30\pm4.28&82.68\pm8.46\\\\ \text{BBCSport}&\text{RMoGL}&27.41\pm2.87&44.80\pm5.89\\\\ \text{Reuters-1200}&\text{HCLS-CGL}&13.92\pm2.94&14.76\pm2.37\\\\ \text{Reuters-1200}&\text{RMoGL}&16.28\pm2.03&15.25\pm1.52\\\\ \hline \end{array}$$ Both native pipelines improve on ORL, Mfeat, and BBCSport. On Reuters-1200, HCLS-CGL improves slightly, while RMoGL decreases slightly. These results show how EAA-acquired observations affect existing IMVC methods when their native clustering pipelines are retained. The complete budget curves will be included in the supplementary analysis.
AuthorsOpenReview ↗

Acquisition policies and decoder choice

Q3. Native CAL/ASCR evaluation and an alternative common decoder.

Response. We followed both suggestions by evaluating CAL and ASCR with their native clustering pipelines and repeating the acquisition comparison with an alternative common decoder.

(a) Native CAL and ASCR evaluation.

We evaluated the CAL- and ASCR-derived acquisition policies using their respective native clustering pipelines. CAL uses its learned graph and native spectral clustering; ASCR uses its normalized Z representation and native K-means. We compare each method’s own adapted acquisition policy with EAA acquisition at the same 30% budget, keeping the native pipeline and its parameters fixed.

Table R3. Same-budget acquisition comparison using native CAL and ASCR clustering.

NMI (%), mean ± sample standard deviation.

\hline \textbf{Dataset}&\textbf{Native pipeline}&\textbf{Own acquisition 30\\%}&\textbf{EAA acquisition 30\\%}\\\\ \hline \text{ORL}&\text{CAL}&\mathbf{97.31\pm0.68}&97.01\pm0.37\\\\ \text{ORL}&\text{ASCR}&95.52\pm0.69&\mathbf{96.52\pm0.24}\\\\ \text{Mfeat}&\text{CAL}&90.29\pm0.91&\mathbf{90.64\pm0.75}\\\\ \text{Mfeat}&\text{ASCR}&\mathbf{96.39\pm0.43}&95.40\pm0.68\\\\ \text{BBCSport}&\text{CAL}&79.20\pm6.92&\mathbf{80.33\pm7.22}\\\\ \text{BBCSport}&\text{ASCR}&\mathbf{96.51\pm1.05}&95.90\pm1.41\\\\ \text{Reuters-1200}&\text{CAL}&\mathbf{18.40\pm5.19}&2.73\pm0.73\\\\ \text{Reuters-1200}&\text{ASCR}&43.20\pm3.72&\mathbf{61.39\pm2.45}\\\\ \text{Caltech101-20}&\text{CAL}&66.43\pm1.06&\mathbf{67.65\pm0.93}\\\\ \text{Caltech101-20}&\text{ASCR}&69.01\pm1.25&\mathbf{70.14\pm0.97}\\\\ \hline \end{array}$$ Bold indicates the higher mean NMI within each row. “Own acquisition” denotes the CAL- or ASCR-derived acquisition policy used in the paper. Both columns use the indicated method’s native final clustering routine. Within each repetition, we averaged ten native clustering runs. The table reports the mean and sample standard deviation across five paired repetitions. The relative performance of the acquisition policies varies with the dataset and native clustering pipeline. **(b) Alternative common decoder.** We kept each policy’s acquired mask fixed and replaced the common graph decoder with per-view PCA, feature concatenation, and K-means. Preprocessing and PCA use only observed rows, with at most 32 components per view. Missing views are represented by the observed mean, giving zero coordinates in the centered PCA space. All policies use the same preprocessing rules and final decoder. **Table R4.** Acquisition policies evaluated with the PCA-concatenation-K-means decoder at a 30% budget. NMI (%), mean ± sample standard deviation. $$\begin{array}{lcccc} \hline \textbf{Acquisition policy}&\textbf{ORL}&\textbf{Mfeat}&\textbf{BBCSport}&\textbf{Reuters-1200}\\\\ \hline \text{Random}&66.61\pm2.22&72.44\pm2.51&46.23\pm5.42&30.53\pm2.17\\\\ \text{View-rate}&65.78\pm2.15&74.41\pm3.99&37.51\pm10.82&28.87\pm4.99\\\\ \text{Uncertainty}&63.38\pm2.24&70.65\pm0.39&41.37\pm6.80&27.96\pm2.48\\\\ \text{CAMA}&69.16\pm1.70&78.68\pm0.68&49.99\pm9.92&\underline{31.82\pm4.26}\\\\ \text{CAL}&\underline{73.08\pm1.39}&73.91\pm1.71&44.07\pm10.02&22.33\pm6.28\\\\ \text{ASCR}&66.82\pm1.87&78.12\pm1.88&\underline{53.82\pm7.60}&20.34\pm5.53\\\\ \text{JVILG}&67.50\pm1.29&59.50\pm0.72&47.40\pm13.10&27.26\pm1.80\\\\ \text{DIMVC-HIA}&67.38\pm1.29&\underline{80.21\pm1.75}&37.51\pm5.03&24.93\pm6.95\\\\ \text{ARSL-IMVC}&67.53\pm0.68&79.71\pm0.90&40.06\pm3.09&20.25\pm1.34\\\\ \text{EAA}&\mathbf{74.95\pm2.14}&\mathbf{81.67\pm2.58}&\mathbf{55.31\pm8.85}&\mathbf{34.04\pm1.77}\\\\ \hline \end{array}$$ Bold and underline indicate the highest and second-highest mean NMI within each column, respectively. EAA achieves the highest mean NMI on all four datasets under this alternative common decoder. Its gains over the strongest competing acquisition policy are 1.87, 1.46, 1.49, and 2.22 percentage points, respectively. The mean performance advantage therefore persists when the final clustering procedure is changed.

Thanks for the authors' experiments to address my concerns. I will keep my score.

Thank you for reviewing our additional experiments and for your constructive feedback. We appreciate your time and consideration.

Reviewer_8VW1OpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Evidence-Adaptive Acquisition (EAA) for budgeted incomplete multi-view clustering. EAA adapts its acquisition granularity to the available evidence: it ranks individual missing entries by typed conditional information gain when evidence is sufficient, and falls back to residual-based view selection when dense views are under-supported, updating the model and candidates over multiple rounds. Experiments show consistent gains over comparison methods across datasets.

Critical strengths

Unlike the traditional missing multi-view learning mechanism, this study no longer treats the missing data items as permanently unavailable. Instead, it focuses on how to obtain additional observed data within a limited budget. This problem setting has practical application value and distinguishes this paper from the traditional incomplete multi-view clustering methods that mainly focus on the recovery of missing view data.

Critical weaknesses

CIG is estimated through cluster condition generation models. However, if the model setting is inaccurate or there is a deviation in the posterior clustering allocation, it may lead to the obtained scores not truly reflecting the clustering benefits brought by obtaining a certain data item. The existing clustering results can demonstrate that the overall acquisition strategy has a certain effect, but this core assumption has not been directly verified yet.

Questions for authors
  1. Could the author test a few other thresholds, such as alpha=0.5 and alpha=0.2, on the ORL, Yale, and representative dense datasets, and report the corresponding results?

  2. Could the setting of the missing rate be explained more clearly?

  3. Is the estimated CIG able to reliably reflect the actual utility of obtaining the missing data items?

4 comments
AuthorsOpenReview ↗

Threshold Sensitivity

Dear Reviewer 8VW1,

Thank you for your careful review and suggestions.

Q1. Could the author test a few other thresholds, such as alpha=0.5 and alpha=0.2, on the ORL, Yale, and representative dense datasets, and report the corresponding results?

Response. Both requested thresholds, α\alpha = 0.20 and 0.50, give the same branches and results as the default α\alpha = 1.00 on all four datasets. We evaluated α∈0.10,0.15,0.20,0.50,1.00,2.00\alpha\in\\{0.10,0.15,0.20,0.50,1.00,2.00\\} on ORL, Yale, COIL20, and Mfeat, using the same five fixed seeds, initial masks, 30% acquisition budget, three acquisition rounds, and final clustering procedure.

The threshold α\alpha determines the branch before acquisition begins. For these dense datasets, EAA uses Typed CIG when the initial observed support reaches α\alpha, and Residual selection otherwise. The selected branch remains fixed throughout acquisition; α\alpha does not enter its scoring or optimization.

Table R1. Clustering performance at different support thresholds.

Mean ± sample standard deviation over five seeds (%). CIG/Residual gives the number of seeds assigned to each branch.

\begin{array}{lccccc} \hline \text{Dataset} & \alpha & \text{CIG/Residual} & \text{ACC} & \text{NMI} & \text{ARI} \\\\ \hline \text{ORL} & 0.10 & 5/0 & 64.70\pm 1.92 & 78.71\pm 1.06 & 49.86\pm 2.52 \\\\ \text{ORL} & 0.15 & 4/1 & 66.10\pm 3.81 & 79.58\pm 2.44 & 51.87\pm 5.21 \\\\ \text{ORL} & 0.20 & 0/5 & 72.60\pm 2.40 & 83.79\pm 1.14 & 60.76\pm 2.55 \\\\ \text{ORL} & 0.50 & 0/5 & 72.60\pm 2.40 & 83.79\pm 1.14 & 60.76\pm 2.55 \\\\ \text{ORL} & 1.00 & 0/5 & 72.60\pm 2.40 & 83.79\pm 1.14 & 60.76\pm 2.55 \\\\ \text{ORL} & 2.00 & 0/5 & 72.60\pm 2.40 & 83.79\pm 1.14 & 60.76\pm 2.55 \\\\ \text{Yale} & 0.10 & 5/0 & 58.30\pm 3.96 & 61.02\pm 3.81 & 38.21\pm 5.48 \\\\ \text{Yale} & 0.15 & 5/0 & 58.30\pm 3.96 & 61.02\pm 3.81 & 38.21\pm 5.48 \\\\ \text{Yale} & 0.20 & 0/5 & 62.55\pm 3.72 & 63.98\pm 3.09 & 43.83\pm 3.86 \\\\ \text{Yale} & 0.50 & 0/5 & 62.55\pm 3.72 & 63.98\pm 3.09 & 43.83\pm 3.86 \\\\ \text{Yale} & 1.00 & 0/5 & 62.55\pm 3.72 & 63.98\pm 3.09 & 43.83\pm 3.86 \\\\ \text{Yale} & 2.00 & 0/5 & 62.55\pm 3.72 & 63.98\pm 3.09 & 43.83\pm 3.86 \\\\ \text{COIL20} & 0.10 & 5/0 & 74.35\pm 2.78 & 83.43\pm 0.85 & 69.89\pm 3.09 \\\\ \text{COIL20} & 0.15 & 5/0 & 74.35\pm 2.78 & 83.43\pm 0.85 & 69.89\pm 3.09 \\\\ \text{COIL20} & 0.20 & 5/0 & 74.35\pm 2.78 & 83.43\pm 0.85 & 69.89\pm 3.09 \\\\ \text{COIL20} & 0.50 & 5/0 & 74.35\pm 2.78 & 83.43\pm 0.85 & 69.89\pm 3.09 \\\\ \text{COIL20} & 1.00 & 5/0 & 74.35\pm 2.78 & 83.43\pm 0.85 & 69.89\pm 3.09 \\\\ \text{COIL20} & 2.00 & 0/5 & 68.18\pm 2.26 & 80.41\pm 0.72 & 61.87\pm 1.47 \\\\ \text{Mfeat} & 0.10 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \text{Mfeat} & 0.15 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \text{Mfeat} & 0.20 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \text{Mfeat} & 0.50 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \text{Mfeat} & 1.00 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \text{Mfeat} & 2.00 & 5/0 & 82.72\pm 0.44 & 87.29\pm 0.61 & 80.76\pm 0.54 \\\\ \hline \end{array}$$ Identical rows arise when the thresholds select the same branch for every seed. With the data, seeds, masks, and budget fixed, these settings have identical downstream computations, so we reuse the corresponding branch runs. For ORL at $\alpha$ = 0.15, one seed has initial support 0.14766 and uses Residual, whereas the other four have support 0.15000–0.15391 and use CIG under the $\geq\alpha$ rule. The reported mean and sample standard deviation are computed over these five runs. Lower thresholds switch ORL and Yale to CIG and reduce mean performance. At $\alpha$ = 2.00, COIL20 switches to Residual, reducing mean NMI from 83.43 to 80.41. Mfeat uses CIG throughout. Thus, performance is stable around the default threshold, with changes occurring when the threshold crosses a branch boundary.
AuthorsOpenReview ↗

Missingness and CIG Utility

Q2. Could the setting of the missing rate be explained more clearly?

Response. A missing entry is the complete feature vector of one sample in one view, i.e., a sample–view entry. For N samples and V views, the observation mask M∈0,1N×VM\in\\{0,1\\}^{N\times V} records whether each entry is observed (1) or missing (0). The initial missing rate is the number of missing sample–view entries divided by NVNV.

The default initial missing rate is 50%. We visit sample–view entries in a random order and hide a total of

m0=round⁡(0.5NV)m_0=\operatorname{round}(0.5NV)

entries, while retaining at least one observed view per sample. One fixed mask is generated for each dataset and seed and shared by all compared methods.

The acquisition budget is defined relative to the initially missing entries:

B=round⁡(0.3m0)B=\operatorname{round}(0.3m_0)

Thus, the default budget acquires 30% of the initially missing entries. It is divided across three rounds, with any integer remainder allocated to earlier rounds, so exactly BB entries are acquired in total. Each sample can acquire at most one new view per round. Acquisition reveals the original feature values; sample class labels are not used to select entries.

For example, with 1,000 sample–view entries, 500 are initially hidden and the acquisition budget is 150. After three rounds, 350 entries remain missing, giving a final missing rate of approximately 35%.

Section 4.4 also evaluates initial missing rates of 30%, 50%, and 65%, with the acquisition budget fixed at 30% of the initially missing entries.

Q3. Is the estimated CIG able to reliably reflect the actual utility of obtaining the missing data items?

Response. We conducted equal-budget comparisons to examine the effect of CIG-selected batches on actual clustering performance.

Table R2 reports results on Mfeat, COIL20, Reuters-1200, scene15, and Caltech101-20. Initial missingness is 50%, and the total budget is 30% of the initially missing entries across three rounds. Before each EAA acquisition round, we compare selecting entries with the highest CIG, the lowest CIG, or at random. All alternatives start from the same observed data and model state, with identical batch sizes, feasibility constraints, and downstream clustering. We reveal the selected entries and rerun clustering, using five fixed seeds.

Table R2. Clustering performance after equal-budget batch acquisition.

NMI (%) reported as mean ± sample standard deviation. Differences are in percentage points. Best values are bold; second-best values are underlined.

\begin{array}{lccccc} \hline \text{Dataset} & \text{Round} & \text{High CIG} & \text{Low CIG} & \text{Random} & \begin{array}{c}\text{High CIG}\\\\-\text{Random}\end{array} \\\\ \hline \text{Mfeat} & 1 & \mathbf{85.93\pm 1.37} & 82.34\pm 0.87 & \underline{84.05\pm 1.28} & +1.88 \\\\ \text{Mfeat} & 2 & \mathbf{86.51\pm 1.10} & 85.27\pm 1.47 & \underline{86.32\pm 1.37} & +0.19 \\\\ \text{Mfeat} & 3 & \mathbf{87.29\pm 0.61} & 86.66\pm 1.23 & \underline{86.76\pm 0.59} & +0.53 \\\\ \text{COIL20} & 1 & \mathbf{81.41\pm 0.41} & 77.61\pm 0.51 & \underline{79.97\pm 0.81} & +1.44 \\\\ \text{COIL20} & 2 & \mathbf{83.59\pm 0.88} & 81.08\pm 0.60 & \underline{82.07\pm 0.36} & +1.52 \\\\ \text{COIL20} & 3 & \mathbf{83.43\pm 0.85} & 82.27\pm 1.15 & \underline{83.06\pm 0.94} & +0.37 \\\\ \text{Reuters-1200} & 1 & \mathbf{26.54\pm 0.94} & 21.37\pm 3.11 & \underline{26.23\pm 1.28} & +0.31 \\\\ \text{Reuters-1200} & 2 & \mathbf{26.73\pm 1.26} & 23.24\pm 3.76 & \underline{24.99\pm 1.83} & +1.74 \\\\ \text{Reuters-1200} & 3 & \mathbf{27.22\pm 1.64} & 22.15\pm 5.27 & \underline{27.09\pm 1.49} & +0.13 \\\\ \text{scene15} & 1 & \mathbf{37.92\pm 0.75} & 34.74\pm 0.53 & \underline{35.72\pm 0.56} & +2.20 \\\\ \text{scene15} & 2 & \mathbf{40.02\pm 0.75} & 38.41\pm 0.54 & \underline{38.58\pm 1.67} & +1.44 \\\\ \text{scene15} & 3 & \mathbf{42.08\pm 0.43} & 40.66\pm 0.43 & \underline{41.16\pm 0.79} & +0.92 \\\\ \text{Caltech101-20} & 1 & \mathbf{56.76\pm 0.89} & 52.73\pm 1.39 & \underline{53.83\pm 0.92} & +2.93 \\\\ \text{Caltech101-20} & 2 & \mathbf{58.71\pm 0.60} & \underline{56.72\pm 1.15} & 56.67\pm 1.65 & +2.04 \\\\ \text{Caltech101-20} & 3 & \mathbf{59.12\pm 0.97} & 57.63\pm 0.97 & \underline{58.73\pm 0.74} & +0.39 \\\\ \hline \end{array}$$ Across the five datasets shown in Table R2, High CIG achieves higher mean NMI than both Low CIG and equal-budget random acquisition in all three rounds. Gains over Random range from 0.13 to 2.93 percentage points. Starting from the same state and acquiring the same number of entries allows this difference to measure the additional benefit of CIG-based batch selection. These results support the practical value of CIG ranking on the five datasets shown.

Thank you for your response. I think this work can be of interest to the user community.  I have increased my score accordingly.

Thank you very much for your careful reconsideration and the updated score. We are grateful that you find our work relevant to the community. Your constructive feedback has helped us a lot in improving this manuscript, and we will keep refining it carefully.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.