acceptodds
Under review as a conference paper at ICLR 2027

Flow-Embedded Multimodal Variational Autoencoders for Incomplete Multi-view Clustering

Abstract

Incomplete multi-view clustering aims to uncover the underlying cluster structure from data observed across multiple views with missing modalities. Recently, generative modeling has emerged as a promising paradigm for this task, as it provides a principled framework for capturing complex data distributions and inferring unobserved information from the available views. In this paper, we propose **F**low-**E**mbedded **M**ultimodal **VA**riational auto**E**ncoders (FEMVAE) with Gaussian mixture priors to learn expressive latent distributions and a coherent shared clustering structure from incomplete multi-view data. Specifically, we integrate normalizing flows into variational autoencoders to construct flexible posterior distributions within a unified framework, while leveraging a Gaussian mixture prior to impose a cluster-aware structure on the learned latent space. Moreover, to fully align the view-specific latent representations, we employ a pairwise optimal transport strategy to promote distributional consistency across views. This enables the complementary information from different views to be effectively integrated despite missing modalities. Once trained, our model can naturally recover missing data modalities. Extensive experiments on several benchmark incomplete multi-view datasets demonstrate the effectiveness of the proposed method, which consistently achieves competitive clustering performance against state-of-the-art approaches under different missing-view settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete Multi-View Clustering, Normalizing Flows, Variational Autoencoders

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_y3hEOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FEMVAE, a method for incomplete multi-view clustering. It uses normalizing flows to model view-specific posterior distributions and combines the available representations through adaptive fusion. A Gaussian mixture prior is introduced to capture cluster structure, while pairwise optimal transport is used to reduce distributional differences across views. The model also reconstructs missing views using inverse flows and view-specific decoders. The authors evaluate the method on five datasets with different missing rates and conduct ablation studies to examine the contributions of its main components.

Critical strengths
  1. The integration of view-conditioned normalizing flows into variational inference enhances the expressiveness of posterior distributions, allowing the model to capture more complex latent structures than conventional Gaussian assumptions.
  2. The pairwise optimal transport strategy promotes consistency among view-specific latent distributions without requiring complete cross-view sample correspondences, making it suitable for incomplete multi-view settings.
  3. The proposed framework combines adaptive fusion, a Gaussian mixture prior, and inverse flow-based reconstruction to integrate complementary information, encourage cluster-aware latent representations, and facilitate missing-view recovery.
Critical weaknesses
  1. Theorem 2.1 establishes a pairwise upper bound based on a weighted mixture of view-specific distributions. However, this mixture reference does not necessarily coincide with the distribution induced by the adaptive fusion mechanism in Eq. (3), leaving a potential gap between the theoretical analysis and the actual fusion process.
  2. The proposed method employs pairwise optimal transport to align view-specific latent distributions, rather than directly aligning corresponding sample representations as in conventional instance-level alignment methods. Although this strategy avoids the requirement for complete cross-view sample correspondences, the distinctions between distribution-level alignment and instance-level alignment, as well as their implications for incomplete multi-view clustering, are not sufficiently discussed.
  3. Although the ablation study demonstrates the contribution of normalizing flows, the influence of flow depth on clustering performance has not been examined. It remains unclear whether increasing posterior expressiveness consistently benefits the proposed framework.
  4. The mathematical formulation introduces numerous symbols, superscripts, subscripts, and distribution-related notations across different sections, making the methodology relatively difficult to follow. In particular, the notation for view-specific and unified latent variables, probability distributions, fusion weights, and loss terms increases the complexity of understanding the proposed framework.
  5. Several notation inconsistencies and minor formatting errors remain in the manuscript. For example, Eq. (2) uses inconsistent encoder notation, Eq. (8) contains mismatched brackets in the KL divergence, and Eq. (11) employs a view-indexing range inconsistent with the observed-view setting. Additionally, Figure 1 contains garbled or improperly rendered mathematical symbols, which affects its readability.
Questions for authors
  1. The authors should clarify how the theoretical upper bound in Theorem 2.1 relates to the actual adaptive fusion process in Eq. (3), given that the weighted mixture reference distribution does not necessarily coincide with the distribution of the unified latent representation.
  2. The authors should further clarify the motivation for adopting distribution-level alignment instead of conventional instance-level alignment, and explain its potential advantages in handling incomplete observations and heterogeneous view availability.
  3. The authors should conduct additional experiments with different numbers of flow transformations to examine how flow depth affects clustering performance and whether deeper flows consistently lead to improvements.
  4. The authors should consider including a notation table summarizing the key mathematical symbols and their definitions to improve readability and help readers better understand the relationships among different components of the framework.
  5. The authors should carefully review and standardize the mathematical notation throughout the manuscript and correct the rendering issues in Figure 1 to improve clarity and presentation quality.
3 comments

We sincerely thank you for your constructive comments. Your suggestions have helped us clarify the theoretical motivation for alignment, expand the experimental analysis of alignment methods and flow depth, and improve the consistency and readability of the mathematical notation. All revisions in the manuscript are highlighted in blue. Q1, W1: Relationship between Theorem 2.1 and Adaptive Fusion.\textbf{Q1, W1: Relationship between Theorem 2.1 and Adaptive Fusion.}

We agree that the weighted mixture reference distribution in Theorem 2.1 does not generally coincide with the distribution of the unified latent representation produced by Eq. (3). Specifically, the former is a mixture of view-specific distributions weighted by the expected fusion weights, whereas the latter is induced by the weighted sum of view-specific latent representations.

Theorem 2.1 establishes that the discrepancy between each view-specific distribution and the mixture reference is bounded by pairwise discrepancies among the view-specific distributions. It therefore motivates the pairwise alignment term as a way to reduce marginal distributional differences among the representations being fused. The theorem does not directly bound the discrepancy between a view-specific distribution and the distribution of the unified latent representation.

We have revised the discussion following Theorem 2.1 in Section 2.4 to clarify the distinction between the mixture reference and the fused distribution, as well as the role of the theorem in motivating pairwise alignment.

Q2, W2: Distribution-Level vs. Instance-Level Alignment.\textbf{Q2, W2: Distribution-Level vs. Instance-Level Alignment.}

We adopt distribution-level alignment because it does not require matched cross-view observations for every sample. Unlike instance-level alignment, which relies on available cross-view pairs, it compares view-specific distributions using the observed samples from each view, including those without a corresponding observation in the other view.

We further compared distribution-level alignment with paired Euclidean and contrastive alignment on CUB and Handwritten at four missing rates. As shown in Table R3.1, distribution-level alignment achieves the highest ACC, NMI, and ARI across all evaluated settings. These results support its effectiveness under incomplete observations.

We have clarified the motivation for distribution-level alignment in Section 2.4 and added comparisons with instance-level alignment methods to the ablation study in Section 3.4, with detailed results and analysis in Appendix A.11.

Table R3.1. Comparison of alignment methods on CUB and Handwritten. Each entry reports ACC/NMI/ARI (%).

DatasetAlignment methodη=0.1\eta=0.1η=0.3\eta=0.3η=0.5\eta=0.5η=0.7\eta=0.7
CUBPaired Euclidean74.80/73.02/60.5172.10/70.80/56.9071.30/66.25/53.2866.50/62.77/49.03
Contrastive74.57/72.94/60.4371.03/70.73/56.5470.27/66.14/52.5864.47/62.57/48.44
Distribution alignment75.08/73.55/62.1076.00/71.11/59.7572.37/66.46/54.3666.83/63.02/50.08
HandwrittenPaired Euclidean93.10/86.01/81.3791.65/84.33/83.4590.64/83.06/81.5790.35/81.74/79.85
Contrastive91.51/83.80/82.4289.79/80.99/79.1890.35/81.95/80.2390.08/80.72/79.48
Distribution alignment94.15/88.61/83.9292.04/84.87/84.6691.46/83.96/82.3990.68/82.22/80.68

Q3, W3: Sensitivity to Flow Depth.\textbf{Q3, W3: Sensitivity to Flow Depth.}

To assess how flow depth affects clustering performance, we varied the number of affine coupling transformations from one to eight on CUB and Handwritten. Table R3.2 reports the ACC, NMI, and ARI at missing rates η∈{0.1,0.3,0.5,0.7}\eta\in \lbrace 0.1,0.3,0.5,0.7 \rbrace.

Clustering performance varies nonmonotonically with flow depth. On CUB, three layers achieve the highest scores on all three metrics at η=0.1\eta=0.1 and η=0.5\eta=0.5, whereas one layer achieves the highest scores at η=0.7\eta=0.7. On Handwritten, two layers achieve the highest scores on all three metrics at η=0.3\eta=0.3, 0.50.5, and 0.70.7. At η=0.1\eta=0.1, increasing depth from four to eight reduces ACC from 94.15% to 87.68%. Across the eight combinations of dataset and missing rate, the default four-layer configuration achieves an ACC within 2.64 percentage points of the highest ACC among the tested depths. These results support four layers as a practical default for ACC on the two evaluated datasets and show that the best-performing depth depends on the dataset, missing rate, and metric.

In the revised manuscript, we summarize these findings in Section 3.3 and include the full results and detailed analysis in Appendix A.10.

Table R3.2. Sensitivity of clustering performance to flow depth on CUB and Handwritten. Each entry reports ACC/NMI/ARI (%). Boldface marks the highest score across the tested depths for each dataset, missing rate, and metric.

DatasetFlow layersη=0.1\eta=0.1η=0.3\eta=0.3η=0.5\eta=0.5η=0.7\eta=0.7
CUB175.63/73.56/61.0674.93/71.32/58.3072.87/68.41/55.0369.47/65.67/52.39
276.03/74.30/61.6773.77/70.93/57.9370.57/67.23/53.2569.00/64.81/51.38
376.80/74.58/62.3173.90/70.51/57.8773.57/69.26/56.9169.27/64.22/51.51
4 (default)75.08/73.55/62.1076.00/71.11/59.7572.37/66.46/54.3666.83/63.02/50.08
574.63/73.19/60.1574.60/70.72/57.9172.27/68.30/54.9266.07/64.34/49.80
676.03/73.37/61.4375.17/71.85/59.1172.00/68.10/54.5966.67/63.47/49.57
776.27/74.34/62.1075.33/71.09/58.9571.07/67.64/54.0669.17/64.26/50.38
876.37/74.26/62.1575.00/70.95/58.2072.50/67.36/53.9466.43/64.04/50.05
Handwritten180.44/75.48/69.6678.57/69.35/62.7385.30/75.33/71.5182.88/73.21/68.83
293.61/86.81/86.4093.46/86.42/86.1092.25/84.49/83.7091.35/82.88/81.88
382.85/78.03/72.5679.57/74.05/68.1387.18/77.65/74.6387.63/77.49/74.85
4 (default)94.15/88.61/83.9292.04/84.87/84.6691.46/83.96/82.3990.68/82.22/80.68
590.88/83.16/81.3689.49/80.97/78.8086.54/79.34/75.8186.03/76.29/72.90
687.01/83.77/80.1790.21/81.90/79.8289.28/80.98/79.0888.95/79.60/77.14
788.66/82.49/79.8186.55/80.44/76.6389.06/79.38/77.5386.65/77.80/74.23
887.68/82.93/79.2686.86/79.34/76.2087.37/79.88/76.3886.80/76.84/73.61

Q4, W4: Improving Readability through a Notation Table.\textbf{Q4, W4: Improving Readability through a Notation Table.}

We have added a notation table in the revised manuscript to summarize the key mathematical symbols, their meanings, and their spaces. An excerpt is provided below; the complete notation table can be found in Table 7 of the revised manuscript.

Table R3.3 Excerpt from the notation table.

SymbolMeaningSpace
XvX^vObserved data in view vvRNv×Dv\mathbb{R}^{N_v \times D_v}
xivx_i^vii-th instance in view vvRDv\mathbb{R}^{D_v}
xvx^vA generic sample when the sample index is clearRDv\mathbb{R}^{D_v}
MMView-availability matrix{0,1}N×V\lbrace 0,1 \rbrace^{N \times V}
MivM_{iv}Availability indicator of view vv for instance ii{0,1}\lbrace 0,1 \rbrace
Vi\mathcal{V}_i, V\mathcal VSet of observed views for instance ii⊆{1,…,V}\subseteq \lbrace1,\ldots,V\rbrace
uivu_i^v, uvu^vBase noise variable of instance ii in view vvRd\mathbb{R}^d
zivz_i^v, zvz^vView-specific latent representation of instance iiRd\mathbb{R}^d
ziz_i, zzUnified latent representation of instance iiRd\mathbb{R}^d
ZvZ^vMatrix of view-specific latent representationsRN×d\mathbb{R}^{N \times d}
ZZMatrix of unified latent representationsRN×d\mathbb{R}^{N \times d}

Q5, W5: Notation and Equation Corrections.\textbf{Q5, W5: Notation and Equation Corrections.}

We have corrected the notation and formatting in Eqs. (2), (8), and (11), as well as Fig. 1 and its caption. Specifically, we have clarified the view-specific encoder as Ev(xv)E^v(x^v), corrected the KL divergence brackets, and restricted the posterior product to observed views. We have also corrected the dataset and fusion notation in Fig. 1. For the reconstruction loss, we have corrected the signs in Eqs. (8) and (11) and included the omitted Dvlog⁡σ2vD_v\log\sigma_2^v term in Eq. (11). Finally, we have checked the use of XvX^v and xvx^v, along with the relevant equations, figures, and captions, for consistency throughout the manuscript.

All of my issues have been addressed. I will maintain my positive score in support of this paper.

Reviewer_kE9kOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies incomplete multi-view clustering using FEMVAE, which combines normalizing flows with variational inference. The method uses adaptive fusion and a Gaussian mixture prior for latent representation learning, together with optimal transport for cross-view alignment. The authors report clustering results on five datasets under different missing rates.

Critical strengths
  1. The paper focuses on incomplete multi-view clustering and offers an approach that combines latent representation learning, cross-view information integration, and missing-view recovery.

  2. The use of view-conditioned normalizing flows allows for more flexible posterior modeling, while the Gaussian mixture prior helps organize the learned representations into clusters.

  3. The pairwise optimal transport strategy aligns latent distributions using available observations from different views, without relying on complete sample correspondences.

  4. The experiments cover five datasets and different missing rates, with ablation analysis providing further support for the proposed design.

Critical weaknesses
  1. The roles of the view-level KL regularization Lzv\mathcal L_{z^v}, shared GMM regularization Lz\mathcal L_z, and cross-view alignment Lalign\mathcal L_{\mathrm{align}} are not sufficiently distinguished. The existing ablations evaluate the flow posterior, GMM prior, and alignment module, but do not isolate the contribution of Lzv\mathcal L_{z^v}.

  2. The experimental tables report performance values without standard deviations, making it difficult to assess result stability and the reliability of small improvements.

  3. Important implementation details remain unclear, including the computation of adaptive fusion weights, missing-view mask generation, and inference-time sampling.

  4. Figure 1 and its caption contain malformed mathematical notation. In addition, the sign of the reconstruction loss Lrec\mathcal L_{\mathrm{rec}} is inconsistent with its definition as a negative expected log-likelihood and with the minimization objective.

Questions for authors
  1. The distinct role of Lzv\mathcal L_{z^v} relative to Lz\mathcal L_z and Lalign\mathcal L_{\mathrm{align}} deserves further discussion. A controlled comparison with and without this term would help clarify its contribution beyond the components already evaluated in the ablation studies.

  2. The number of independent runs should be specified, together with the mean and standard deviation of the reported results. These statistics would help assess performance stability and the reliability of relatively small improvements.

  3. Additional details are needed regarding the computation of adaptive fusion weights, the generation of missing-view masks, and the sampling strategy used during inference. Clarifying these choices would improve reproducibility and facilitate fair comparisons.

  4. The malformed notation in Figure 1 and its caption should be corrected. The signs in Lrec\mathcal L_{\mathrm{rec}} also require checking to ensure consistency among its negative expected log-likelihood definition, expanded expression, and minimization objective.

3 comments

We appreciate the time and care you have taken in reviewing our work. Your comments have helped us strengthen the ablation study and stability analysis, clarify implementation details for reproducibility, and correct the notation and reconstruction loss expressions. All revisions in the manuscript are highlighted in blue.

Q1, W1: Ablation Study on View-Level KL Regularization.\textbf{Q1, W1: Ablation Study on View-Level KL Regularization.}

The view-level KL term Lzv\mathcal L_{z^v} regularizes each q(zv∣xv)q(z^v\mid x^v) toward a standard Gaussian prior. The shared KL term Lz\mathcal L_z regularizes the fused posterior toward a GMM prior to encourage cluster structure. The alignment term Lalign\mathcal L_{\mathrm{align}} aligns the empirical latent distributions of different views using a pairwise Wasserstein cost.

The original Gaussian-prior ablation evaluates the prior choice for Lz\mathcal L_z but does not isolate Lzv\mathcal L_{z^v}. We therefore added an ablation that removes Lzv\mathcal L_{z^v} while retaining the flow posterior, shared GMM regularization, and cross-view alignment.

As shown in Table R2.1, the full model achieves higher scores in all 24 comparisons across both datasets and four missing rates. At η=0.5\eta=0.5, including Lzv\mathcal L_{z^v} improves ACC/NMI/ARI by 16.14/10.81/14.95 percentage points on CUB and 0.15/0.13/1.30 percentage points on Handwritten. These results support the contribution of view-level KL regularization.

We have updated Section 3.4 of the manuscript to summarize the view-level KL ablation and its main findings. Appendix A.11 describes the ablation setup and provides a detailed analysis of the results. Table 2 now reports all ACC, NMI, and ARI values as percentages to two decimal places, including the results of the previously reported ablation variants.

Table R2.1. Ablation study on view-level KL regularization. Values are reported as percentages.

DatasetMissing rate (η\eta)AlignFlowGMMView-LevelACC (%)NMI (%)ARI (%)
CUB0.1✓✓✓70.5769.2856.19
✓✓✓✓75.0873.5562.10
0.3✓✓✓60.8360.8445.08
✓✓✓✓76.0071.1159.75
0.5✓✓✓56.2355.6539.41
✓✓✓✓72.3766.4654.36
0.7✓✓✓53.8753.1236.28
✓✓✓✓66.8363.0250.08
Handwritten0.1✓✓✓93.6487.4681.71
✓✓✓✓94.1588.6183.92
0.3✓✓✓91.7984.2883.78
✓✓✓✓92.0484.8784.66
0.5✓✓✓91.3183.8381.09
✓✓✓✓91.4683.9682.39
0.7✓✓✓89.9781.9879.55
✓✓✓✓90.6882.2280.68

Q3, W3: Implementation Details and Reproducibility.\textbf{Q3, W3: Implementation Details and Reproducibility.}

Adaptive fusion weights. We introduce a learnable parameter ava^v for each view and compute the weights using a softmax function: αv=exp⁡(av)∑j=1Vexp⁡(aj).\alpha^v=\frac{\exp(a^v)}{\sum_{j=1}^{V}\exp(a^j)}. These parameters are jointly optimized with the model. For an incomplete sample ii, we use an observation mask MivM_{iv}, where Miv=1M_{iv}=1 indicates that view vv is observed, and normalize the weights over its observed views: $ \alpha_i^v= \frac{M_{iv}\alpha^v}{\sum_{j=1}^{V}M_{ij}\alpha^j}.$

Missing-view masks. We randomly select a proportion η\eta of samples to be incomplete. For each selected sample, we generate an observation mask such that at least one view is observed and at least one view is missing.

Sampling strategy. During training, we draw one noise vector uiv∼N(0,I)u_i^v\sim\mathcal N(0,I) for each observed view and construct the view-level latent variable as $ z_i^v=F^v\left(E^v(x_i^v)+\sigma_1^v u_i^v\right). $ After training, we set the noise vector uivu_i^v to the zero vector when obtaining latent representations for clustering, avoiding random sampling at this stage.

We have added these implementation details to Appendix A.8 of the revised manuscript.

Q2, W2: Experimental Stability and Statistical Reliability.\textbf{Q2, W2: Experimental Stability and Statistical Reliability.}

We have clarified in Appendix A.8 that each method was evaluated on each dataset at each missing rate over five independent runs using five random seeds. Table 1 reports the mean ACC, NMI, and ARI across these runs. We have added Table 6 in Appendix A.12 to report the corresponding standard deviations for all five datasets. Table R2.2 below presents the results for CUB and Handwritten.

FEMVAE has standard deviations of at most 2.00 percentage points in 48 of the 60 combinations of dataset, missing rate, and metric. On CUB at η=0.1\eta=0.1, FEMVAE and DCG achieve mean ACC values of 75.08% and 74.57%, respectively. The 0.51-percentage-point gap is small relative to their run-to-run standard deviations of 1.53 and 3.34 percentage points. Across the five runs, FEMVAE shows modest variability in most settings, with its largest standard deviations observed on Caltech7-5V at missing rates of 0.5 and 0.7. The standard deviations put small differences in mean performance in the context of the variation observed across runs.

Table R2.2. Standard deviations of ACC, NMI, and ARI across five independent runs on CUB and Handwritten. Values are in percentage points. Parentheses indicate the missing rate.

DatasetMethodACC (0.1)NMI (0.1)ARI (0.1)ACC (0.3)NMI (0.3)ARI (0.3)ACC (0.5)NMI (0.5)ARI (0.5)ACC (0.7)NMI (0.7)ARI (0.7)
CUBAPADC1.333.433.272.050.501.962.882.261.822.992.321.88
RPCIC5.432.602.266.904.215.283.003.512.901.670.920.51
COMPLETER3.912.214.334.863.074.713.941.493.602.552.500.31
HSACC2.411.802.783.073.516.161.795.603.991.446.167.35
RGCL2.731.191.951.310.590.810.790.990.842.351.702.10
DGIMVCM3.891.601.631.681.391.931.001.611.481.671.401.34
DCG3.342.052.883.190.671.566.533.195.263.353.023.51
BRIDGE2.452.883.923.812.112.976.215.406.402.371.701.06
BURG3.533.550.854.165.964.521.642.402.423.455.892.28
FEMVAE1.531.893.041.400.971.982.001.342.422.481.461.91
HandwrittenAPADC3.303.167.502.162.433.454.252.925.111.001.011.59
RPCIC5.013.214.445.213.144.363.732.502.272.641.501.49
COMPLETER6.365.197.391.421.353.394.843.528.050.981.291.46
HSACC2.411.234.751.762.624.523.272.055.432.352.243.20
RGCL0.640.881.040.861.311.650.931.421.820.981.471.71
DGIMVCM1.340.580.540.500.510.931.130.690.863.381.573.12
DCG2.972.804.275.864.056.336.205.017.126.025.076.96
BRIDGE0.700.700.464.263.584.825.034.665.353.892.422.43
BURG6.144.067.054.793.995.502.601.272.724.583.465.72
FEMVAE0.640.651.260.470.590.900.821.101.650.801.451.51

Q4, W4: Notation and Equation Corrections.\textbf{Q4, W4: Notation and Equation Corrections.}

We have corrected the notation in Fig. 1 and its caption, including {Xv}v=1V\lbrace X^v\rbrace_{v=1}^{V}, and revised the fusion equation to Z=∑v∈VAvZvZ=\sum_{v\in\mathcal V}A^vZ^v. We have also corrected the signs in Eqs. (8) and (11). The revised reconstruction loss is Lrec=−∑v∈VEq(z∣{zv})∏v∈Vq(zv∣xv)[log⁡p(xv∣z)]=∑v∈V[Eq(z∣{zv})∏v∈Vq(zv∣xv)[∥Gv((Fv)−1(z))−xv∥22(σ2v)2]+Dvlog⁡σ2v]\mathcal L_{\mathrm{rec}} = -\sum_{v\in\mathcal V} \mathbb E_{q(z\mid \lbrace z^v \rbrace ) \prod_{v \in \mathcal V}q(z^v \mid x^v)} [\log p(x^v\mid z) ] = \sum_{v\in\mathcal V}[ \mathbb E_{q(z\mid \lbrace z^v \rbrace ) \prod_{v \in \mathcal V}q(z^v \mid x^v)}[ \frac{\|G^v((F^v)^{-1}(z))-x^v\|^2} {2(\sigma_2^v)^2} ] +D_v\log\sigma_2^v]. This correction includes the previously omitted ∑v∈VDvlog⁡σ2v\sum_{v \in \mathcal V}D_v\log\sigma_2^v term and ensures consistency between the negative expected log-likelihood and the minimization objective. We have also checked the relevant equations, figure, and caption for consistency.

Thanks for your response. I have read the comments and the newly added experimental results. The authors have addressed my concerns, and I will maintain my positive score.

Reviewer_e1z6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a generative framework for incomplete multi-view clustering. By enhancing the expressiveness of view-specific latent distributions and aligning their distributions in a shared latent space, FEMVAE aims to improve clustering performance under incomplete view settings. Experiments on five benchmark datasets with varying missing rates evaluate the clustering performance of FEMVAE . However, the probabilistic formulation and the role of view-specific latent variables require further clarification. The experimental analysis and mathematical notation could be further improved.

Critical strengths
  1. This paper points out that missing views reduce the information available for learning shared representations. FEMVAE combines information from visible views and imputes missing views to address this challenge.
  2. The method achieves promising clustering performance across five benchmark datasets and different missing rates.
Critical weaknesses
  1. Equation (4) factorizes the joint distribution into independent priors over the shared latent variable zz and the view-specific latent variables zvz^v, together with observation likelihoods conditioned on zz. This factorization implicitly assumes prior independence between zz and zvz^v, as well as conditional independence between the observations and the view-specific latent variables given zz. However, the view-specific latent representations are inferred from their corresponding observations through observation-conditioned normalizing flows. The authors should clarify the probabilistic assumptions underlying this factorization and explain how the generative model and variational posterior are consistently defined, particularly regarding the dependencies among xvx^v, zvz^v, and zz.
  2. The role of the view-specific latent variables zvz^v in the probabilistic formulation requires further justification. Since these variables are used to construct the shared representation zz, it is unclear why they must be modeled as variational latent variables with independent priors and additional KL regularization, rather than serving as intermediate representations. The authors should clarify the necessity of this design.
  3. The latent dimensionality dd is fixed to the number of clusters KK in the experiments. However, the influence of different latent dimensions on clustering performance has not been investigated, making the robustness of this design choice unclear.
  4. In Line 78, the subscript in the equation is wrong. In addition, the notation Cvi,vjC^{v_i,v_j} in Figure 1 is inconsistently presented as a superscript and a subscript. The authors should carefully check and correct these issues.
Questions for authors

See weaknesses

3 comments

We are grateful for your thoughtful review and valuable suggestions. Your comments have helped us clarify the probabilistic assumptions and view-level regularization, examine the effect of latent dimensionality, and improve the consistency of the notation and equations. All revisions in the manuscript are highlighted in blue.

W1: Probabilistic Assumptions and Variable Dependencies.\textbf{W1: Probabilistic Assumptions and Variable Dependencies.}

Equation (4) defines the generative model with independent priors over the unified latent variable zz and the view-level latent variables {zv}\lbrace z^v \rbrace. It assumes that the observations are conditionally independent across views given zz, and that p(xv∣z,{zv})=p(xv∣z)p(x^v\mid z, \lbrace z^v \rbrace)=p(x^v\mid z). Thus, each view is generated from zz through its corresponding inverse flow and decoder, without additional conditioning on {zv}\lbrace z^v \rbrace.

The inference process follows a different dependency structure, specified by q(z,{zv}∣{xv})=q(z∣{zv})∏vq(zv∣xv)q(z,\lbrace z^v\rbrace \mid\lbrace x^v\rbrace)=q(z\mid\lbrace z^v\rbrace )\prod_v q(z^v\mid x^v). Specifically, for each observed view xvx^v, an observation-conditioned normalizing flow transforms a noise variable into a view-level latent variable zvz^v, defining q(zv∣xv)q(z^v\mid x^v). The resulting view-level variables are then adaptively aggregated according to Eq. (3) to construct the unified representation zz.

These dependencies are introduced by the variational family for approximate inference and do not necessarily coincide with the dependencies in the exact posterior under the generative model. The generative model and the variational inference process are connected through the ELBO formulation in Eq. (8). We have revised the paragraph preceding Eq. (4) in Section 2.2 to clarify these assumptions and the distinction between the generative and inference processes. W2: Rationale for View-Level Variational Modeling and KL Regularization.\textbf{W2: Rationale for View-Level Variational Modeling and KL Regularization.}

Our motivation is to explicitly model and regularize the distributions of the representations used for fusion. Specifically, q(zv∣xv)q(z^v\mid x^v) provides an observation-dependent distribution that supports sampling, density evaluation, and distributional alignment at the view level. This formulation allows us to regularize a conditional distribution for each observation, rather than using only a deterministic intermediate feature.

The view-level KL term directly regularizes each conditional distribution toward a fixed Gaussian prior. This serves a different purpose from regularizing the fused representation or aligning view-level distributions: the former does not uniquely determine the individual distributions, while the latter does not constrain them to a fixed reference. The factorized prior enables separate KL regularization for each view without implying that their learned semantic information is independent.

As shown in Table 2, removing the view-level KL term leads to a decline in clustering performance, supporting the contribution of direct view-level regularization. We have expanded the discussion of Lzv\mathcal L_{z^v} in Section 2.3, following Eq. (8), to explain the motivation for direct view-level regularization and the role of the factorized priors.

W3: Sensitivity to the latent dimension d.\textbf{W3: Sensitivity to the latent dimension d.}

To assess the choice of d=Kd=K, we varied the latent dimension while keeping the number of clusters fixed. We evaluated $d\in \lbrace 3,5,10,15,35 \rbrace $ on CUB and Handwritten at missing rates $\eta\in\lbrace 0.1,0.3,0.5,0.7\rbrace $. Both datasets have K=10K=10 clusters, so d=10d=10 corresponds to the original setting. Table R1.1 reports the ACC, NMI, and ARI for each setting.

Setting d=3d=3 consistently yields lower scores than d=Kd=K on both datasets, indicating that an excessively small latent dimension can limit clustering performance. Across all tested missing rates, the ACC at d=Kd=K is within 2.30 percentage points of the highest ACC among the tested dimensions. Although larger dimensions yield better results on Handwritten, these findings suggest that d=Kd=K is a reasonable default, while tuning dd may further improve performance.

In the revised manuscript, we summarize these findings in Section 3.3. Appendix A.10 provides the detailed analysis, and Table 3 reports the full results.

Table R1.1 Sensitivity of clustering performance to the latent dimension dd on CUB and Handwritten. Each entry reports ACC/NMI/ARI (%). Boldface marks the highest score across the tested dimensions for each dataset, missing rate, and metric. KK denotes the number of clusters.

Datasetddη=0.1\eta=0.1η=0.3\eta=0.3η=0.5\eta=0.5η=0.7\eta=0.7
CUB373.40/73.05/60.7573.90/70.14/57.4867.67/65.24/50.6064.83/60.95/46.92
575.93/74.03/61.7972.80/70.37/57.7970.50/67.46/53.8669.13/63.95/50.52
10 (d=Kd=K)75.08/73.55/62.1076.00/71.11/59.7572.37/66.46/54.3666.83/63.02/50.08
1576.63/74.63/62.2275.03/71.01/58.8274.20/69.27/56.3969.13/64.71/51.09
3577.23/74.71/62.8075.43/70.68/59.2473.30/69.25/55.5868.83/65.21/51.54
Handwritten379.04/74.77/67.3582.48/75.47/70.4781.54/75.59/69.7483.15/74.37/69.23
592.35/85.43/84.0191.29/83.34/81.9386.73/79.43/76.3488.71/79.47/77.14
10 (d=Kd=K)94.15/88.61/83.9292.04/84.87/84.6691.46/83.96/82.3990.68/82.22/80.68
1594.80/88.93/88.8593.14/86.11/85.5192.87/85.22/84.8791.78/83.48/82.72
3595.31/90.04/89.8993.71/86.90/86.5593.32/86.16/85.7592.44/84.77/83.98

W4: Notation and Equation Corrections.\textbf{W4: Notation and Equation Corrections.}

We have corrected the summation subscript in the fusion equation on Line 78, which now reads Z=∑v∈VAvZvZ=\sum_{v\in\mathcal V}A^vZ^v, and replaced C1vC_{1v} with C1vC^{1v} in Fig. 1. We have checked Fig. 1 and its caption, as well as the mathematical notation throughout the manuscript, for consistency across the equations, figures, and text. As part of this consistency check, we have also added the missing negative sign before the expected log-likelihood in Eq. (8) and removed the incorrect negative sign before the squared reconstruction error in Eq. (11). We have included the term ∑v∈VDvlog⁡σ2v\sum_{v\in\mathcal V}D_v\log\sigma_2^v in Eq. (11), omitting constants independent of the model parameters.

The authors have addressed my concerns, and I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.