acceptodds
Under review as a conference paper at ICLR 2027

Prototype-Routed Flow Matching for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) aims to learn cluster-discriminative representations from partially observed views. Completion-based methods estimate missing-view representations so that observed and completed views can be used together for clustering. However, direct prediction and prototype-based reconstruction typically produce a single estimate for each missing view. When the observed views cannot uniquely determine the missing view, a single estimate may weaken cluster separation. Generative completion can instead model a conditional distribution over missing-view representations, but iterative denoising makes diffusion-based completion costly during training. We propose Prototype-Routed Flow Matching (PRFM), which formulates missing-view completion as clustering-aligned conditional transport in a shared latent space. PRFM uses dynamically updated online prototypes as routing signals during flow matching, transforming missing-view completion from static point prediction into a prototype-routed transport process. At each generation step, the assigned prototype provides a cluster-level direction, while a residual velocity captures sample- and view-specific variation. The same prototypes also prioritize observable candidates for cluster-consistent soft supervision and optimize completed representations in the prototype space. Completed latents are produced through few-step ODE inference. Experiments on multiple IMVC benchmarks show that PRFM achieves competitive clustering performance, remains robust under high missing rates, and offers favorable training and inference efficiency compared with diffusion-based completion. The code will be released upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering, flow matching, prototype learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.40 over 5 reviews, above 99% of the others.
Reviewer_Etb3OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a novel Prototype-Routed Flow Matching (PRFM) for IMVC. PRFM explicitly incorporates dynamically updated clustering prototypes into the velocity field, combining prototype-routed and residual velocities to model the generation of missing-view latent representations. The paper further introduce prototype-prioritized soft-pair supervision and jointly optimize completed representations and clustering structures in shared prototype space. Experiments on multiple IMVC benchmarks demonstrate competitive clustering performance and efficient generation.

Critical strengths

1.The key innovation of this paper is to explicitly incorporate clustering prototypes into the Flow Matching velocity field, rather than using them solely as conditioning inputs or generation targets. The prototype-routed velocity provides cluster-level guidance, while the residual velocity preserves sample- and view-specific information, allowing missing-view completion to be more directly aligned with the clustering objective. This design is well motivated and offers a new perspective on generative incomplete multi-view clustering.

2.The paper cleverly integrates online prototype learning, prototype-guided generation, soft-pair supervision, and clustering optimization into a unified framework. The ablation studies are comprehensive, particularly the comparisons with Standard FM and Prototype-Cond. FM, which help distinguish the contributions of standard flow matching, prototype conditioning, and explicit prototype routing.

3.The paper compares PRFM with seven state-of-the-art IMVC methods across multiple benchmark datasets, demonstrating competitive clustering performance. In addition, PRFM reduces generation costs through few-step ODE inference. The runtime experiments further show clear inference speed advantage over diffusion-based completion methods, demonstrating good balance between clustering performance and computational efficiency.

Critical weaknesses

1.This paper dynamically updates clustering prototypes using EMA and employs them to guide the Flow Matching velocity field. However, it unclear how much additional benefit dynamic prototype updates provide compared with fixed prototypes.

2.Although PRFM explicitly incorporates prototype-routed guidance into the velocity field, the residual velocity may potentially counteract this guidance. Lacks analysis of how these two velocity components interact during inference, making it unclear to what extent prototype guidance is preserved in the resulting velocity field.

3.The existing ablation studies without separately evaluating the contributions of the clustering-alignment loss and confidence-weighted fusion strategy.

Questions for authors

Q1. Could authors provide a comparison with fixed prototypes to assess the contribution of EMA dynamic prototype updates?

Q2. Could authors analyze the relative magnitudes and directional alignment of the prototype-routed and residual velocities at different integration times to clarify how prototype guidance influences the learned velocity field?

Q3. Could authors provide ablation studies without the clustering-alignment loss and confidence-weighted fusion to clarify the individual contributions of these two components?

1 comment

We thank the reviewer for these suggestions. We provide additional ablations for online prototype updates, clustering alignment, and confidence-weighted fusion, together with an analysis of the two velocity components during generation.

Q1: Contribution of online prototype updates

Response: We compared Full PRFM with Frozen Prototypes, which keeps the initialized prototype bank fixed instead of updating it through EMA. Table R1 reports the results on HandWritten and YouTubeFace10 at a 70% missing rate, together with the ablations requested in Q3.

Table R1. Additional ablations at a 70% missing rate.

DatasetSettingACC (%)NMI (%)ARI (%)
HandWrittenFull PRFM84.1472.3567.97
HandWrittenFrozen Prototypes81.9570.5466.51
HandWrittenw/o Lcomp\mathcal{L}_{\mathrm{comp}}83.3271.6567.15
HandWrittenUniform Fusion77.8066.7460.09
YouTubeFace10Full PRFM78.7473.8166.94
YouTubeFace10Frozen Prototypes76.6673.2465.79
YouTubeFace10w/o Lcomp\mathcal{L}_{\mathrm{comp}}78.2473.5365.76
YouTubeFace10Uniform Fusion69.9858.0754.88

Freezing the prototypes reduces ACC by 2.19 and 2.08 percentage points on HandWritten and YouTubeFace10, respectively, with decreases in NMI and ARI as well. These results support updating the prototypes during learning rather than relying solely on their initial values.

Q2: Interaction between prototype-routed and residual velocities

Response: We measured the weighted velocity components along inference trajectories on HandWritten and MSRC-V1 at a 70% missing rate. Here, R(t)R(t) is the prototype-component norm divided by the sum of the two component norms; A(t)A(t) is the cosine similarity between the two components; and P(t)P(t) is the cosine similarity between the total velocity and the direction toward the assigned prototype. The table reports averages over evaluated missing sample-view pairs.

Table R2. Velocity interaction at different integration times.

DatasetttR(t)R(t)A(t)A(t)P(t)P(t)
HandWritten0.10.5094-0.01360.7141
HandWritten0.50.2603-0.22240.1306
HandWritten0.90.0531-0.4877-0.4433
MSRC-V10.10.53850.00710.7598
MSRC-V10.50.2880-0.18090.2204
MSRC-V10.90.0602-0.4254-0.3709

The total velocity has strong positive alignment with the prototype direction early in generation. As integration progresses, the relative prototype magnitude decreases and the residual component becomes dominant. The negative values of A(t)A(t) and P(t)P(t) near the endpoint confirm that the residual can counteract and override the prototype direction.

Thus, prototype guidance is not preserved as a strict directional constraint throughout generation. Instead, the observed behavior is consistent with the time-decaying guidance in Eq. (6): an early prototype-directed bias followed by greater residual flexibility. This is also compatible with Proposition 1, which allows the residual field to compensate for the explicit prototype term.

Q3: Contributions of clustering alignment and confidence-weighted fusion

Response: Table R1 separately removes Lcomp\mathcal{L}_{\mathrm{comp}} and replaces confidence-weighted fusion with uniform averaging of the observed and generated latents.

Removing Lcomp\mathcal{L}_{\mathrm{comp}} produces small decreases across all three metrics on both datasets, including ACC reductions of 0.82 and 0.50 percentage points. This indicates a modest additional benefit from clustering alignment in the evaluated settings.

Uniform Fusion leads to larger reductions: ACC decreases by 6.34 and 8.76 percentage points on HandWritten and YouTubeFace10, respectively, accompanied by lower NMI and ARI. These results support weighting the view contributions rather than treating observed and generated representations equally.

We have added the Frozen Prototypes comparison to Appendix H.1 and the velocity-component analysis to Appendix H.2 of the revised manuscript.

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

The paper proposes a novel framework for incomplete multi-view clustering that models missing-view completion as a conditional transport process guided by clustering structure. Rather than using online clustering prototypes only as conditioning inputs or completion targets, PRFM incorporates them directly into the velocity field. The prototype-routed velocity provides a cluster-level direction, while a learned residual velocity models sample- and view-specific variation. Prototype guidance gradually decays over generation time, shifting the process from cluster-level routing to residual refinement. The same prototypes also provide soft supervision targets when paired observations are unavailable and guide the cluster assignments of completed representations. This connects the generation trajectory, supervision targets, and clustering optimization within a shared prototype space.

Critical strengths

First, the paper introduces an interesting design that makes clustering prototypes an explicit part of the velocity field, rather than merely using them as conditioning inputs. This allows prototypes to directly guide the transport process, while the residual branch retains the flexibility to capture sample-specific variations. This goes beyond simply applying Flow Matching to missing-view completion. Second, prototype-prioritized soft pairing is closely connected to the generation mechanism. The prototypes not only guide the transport direction but also help select supervision targets when paired observations are unavailable. This links the generation process with supervision and distinguishes PRFM from conventional generative completion methods. Third, The experiments comparisons with Standard FM and Prototype-Cond. FM help distinguish the contributions of standard flow matching, prototype conditioning, and explicit prototype routing. The runtime results further demonstrate the inference efficiency of the method.

Critical weaknesses

First, the paper does not clearly specify whether the assigned prototype remains fixed throughout ODE inference or is updated as the latent representation evolves. Second, when no same-prototype candidates are available, the soft-pair module falls back to all samples with the target view observed. However, it is unclear whether these fallback pairs are treated differently in confidence weighting, given their potentially lower semantic consistency. Third, the paper uses online prototypes to guide the clustering of completed representations, but applies K-means to obtain the final cluster assignments. The effectiveness of directly using the learned prototypes for clustering has not been evaluated.

Questions for authors

Please refer to weakness

1 comment

We thank the reviewer for the helpful comments. We clarify prototype handling during inference and fallback weighting, and provide an additional comparison of direct prototype assignment and final K-means.

Q1: Prototype assignments during ODE inference

Response: The assigned prototype remains fixed throughout each ODE trajectory. It is selected from the observed-view consensus before integration and is not reassigned using intermediate generated representations. The prototype bank is also fixed during inference.

A fixed prototype does not imply a constant guidance velocity. The prototype-directed component depends on the evolving latent representation and the time-dependent guidance coefficient. Thus, the prototype provides a consistent reference while the guidance changes during generation. “Online” refers to prototype updates during training, not within an inference trajectory.

Q2: Confidence weighting of fallback pairs

Response: When no same-prototype candidates are available, Eq. (8) expands the candidate pool to all samples in the current mini-batch with the target view observed. Candidates are still ranked by similarity to the assigned prototype, selected using the top-RR rule, and aggregated with the softmax weights in Eq. (9).

Fallback pairs follow the same confidence-weighting rule as regular soft pairs. Their velocity-regression loss is weighted according to the assignment confidence of the anchor and selected candidates, as described in Eq. (10). No additional fallback-specific penalty is applied. The fallback therefore changes the candidate pool, but not the subsequent aggregation or confidence-weighting procedure.

Q3: Direct prototype assignment versus final K-means

Response: We added a comparison on HandWritten and YouTubeFace10 at a 70% missing rate. We report mean results over five random seeds. Both procedures use exactly the same completed representations and differ only in the final assignment rule. Direct prototype assignment assigns each completed representation to the prototype with the highest assignment probability in Eq. (15), while K-means follows the original evaluation procedure.

DatasetFinal clustering procedureACC (%)NMI (%)ARI (%)
HandWrittenDirect prototype assignment53.6550.9137.97
HandWrittenK-means84.1472.3567.97
YouTubeFace10Direct prototype assignment51.0749.2234.80
YouTubeFace10K-means78.7473.8166.94

K-means achieves better performance on the same completed representations across both datasets. The online prototypes guide generation and representation learning, whereas K-means fits the final cluster centers to the completed representations. These results support retaining the final clustering step rather than treating the online routing prototypes as the final cluster centers.

We have added the clarifications on prototype handling during inference and on fallback pairs to Appendix A of the revised manuscript.

Reviewer_kvkwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PRFM, a completion-based method for incomplete multi-view clustering (IMVC). Missing-view latents are generated by conditional flow matching in a shared latent space. The velocity field is decomposed (Eq. 6) into a time-decaying "prototype-routed" term apg(t)(sg(cη)−zt)a_{pg}(t)(\mathrm{sg}(c_\eta) - z_t) that pulls the state toward the assigned online prototype, plus a learned residual velocity. The same prototypes are used (i) to build soft regression endpoints for missing pairs from top-RR same-cluster candidates (Eqs. 8–10), and (ii) to align the completed representations in prototype space (Eqs. 15–16). Proposition 1 shows that, with an unrestricted residual, the prototype term does not change the weighted-regression optimum. Experiments on four benchmarks (MSRC-V1, HandWritten, ALOI-Deep3V, YouTubeFace10) at missing rates of 10–70% report the best ACC in all 16 settings. The paper also includes an ablation at 70% on two datasets, hyperparameter sensitivity studies, and a runtime comparison with two diffusion-based completion methods (DCG, NCD).

Critical strengths
  1. Clear and well-motivated design. Writing the prototype into the velocity parameterization (rather than only feeding it as a condition) is a simple, cheap idea. Figures 1 and 2 and Algorithm 1 make the pipeline easy to follow.
  2. I appreciate this candor. Appendix D.1 states explicitly that Proposition 1 establishes regression compatibility only, and does not guarantee attraction to true clusters or improved expressivity.
  3. Few-step ODE inference is orders of magnitude faster than the diffusion baselines in Table 3 (e.g., 10.82 ms vs. 151.79 ms for DCG and 28,588.76 ms for NCD on HandWritten).
  4. Broad experimental grid and a well-conceived ablation design.
Critical weaknesses
  1. The ablation contradicts the paper's own sensitivity study. By Eqs. 6 and 11, λpg=0\lambda_{pg}=0 and λ=0\lambda=0 equal "Prototype-Cond. FM" and "w/o Soft Pair.". Yet Figure 5 and Table 2 disagree sharply (HandWritten ACC 84.00 vs. 81.10; YouTubeFace10 79.83 vs. 69.63), and the ablated settings match or beat Full PRFM. Because Proposition 1 makes routing neutral at the optimum, both core contributions currently lack empirical support.
  2. Weak statistical rigor. No standard deviations are reported, although many margins are under one point. The ACC values on MSRC-V1 and HandWritten fall on single-run grids (e.g., k/210k/210), which is hard to reconcile with 5-seed means. Hyperparameters appear to be tuned with labels on the reported settings, while baselines use defaults and some look like failed runs.
  3. Inference draws a single sample, and the soft targets are prototype-weighted averages, so the "conditional distribution vs. point estimate" argument is never examined. Without a no-completion control or an evaluation of completion fidelity, it is unclear whether completion itself drives the gains.
  4. The overall pipeline of Figure 2 is fine as is; however, the visual details could be refined, particularly regarding the arrows and line work.
Questions for authors
  1. Are "Prototype-Cond. FM" and λpg=0\lambda_{pg}=0 the same configuration? Are "w/o Soft Pair." and λ=0\lambda=0 the same? If not, what exactly differs? If they are the same, which numbers are correct, and how do the conclusions of Section 4.3 change?
  2. Are the Figure 5 cells single runs or 5-seed means? Why do several distinct (λpg,λ)(\lambda_{pg}, \lambda) settings give identical triples, in particular the λpg=0\lambda_{pg}=0 column on YouTubeFace10?
  3. What is the pseudo-label accuracy on incomplete samples during training?
  4. How were the hyperparameters and the number of Euler steps chosen without labels? Were baselines given a comparable tuning budget? What are the default RR and τp\tau_p, and why do the curves in Figure 4(d)(e) not pass through the default result?
3 comments

We sincerely thank the reviewer for the detailed review, and for recognizing that our design is clear and well motivated. We provide a point-to-point response below and the revised manuscript has also been uploaded.

Q1 & W1: Consistency between the ablation study and hyperparameter sensitivity analysis

Response: Thank you for identifying this inconsistency. We have carefully checked the experimental results and corrected the corresponding figures in the revised manuscript.

We confirm that the configurations are equivalent. Specifically, setting λpg=0\lambda_{\mathrm{pg}}=0 in Eq. (6) removes the explicit prototype-routed velocity while retaining the prototype as a conditioning input, corresponding to Prototype-Cond. FM in Table 2. Similarly, setting λ=0\lambda=0 in Eq. (11) disables soft-pair supervision, corresponding to w/o Soft Pair.

Figures 4 and 5 in the original manuscript presented single-run results. We thank the reviewer for pointing out their inconsistency with the corresponding results in Table 2. We have rechecked the corresponding configurations and updated the sensitivity plots in the revised manuscript to report mean results over five random seeds, while also checking the corresponding entries in Table 2. The revised figures and table now report consistent results for the corresponding configurations.

The updated results continue to support the conclusions of Section 4.3. In particular, removing explicit prototype routing while retaining prototype conditioning reduces ACC by 2.91 and 7.46 percentage points on HandWritten and YouTubeFace10, respectively, supporting the effectiveness of the proposed routing mechanism.

We also clarify that Proposition 1 establishes regression compatibility rather than guaranteed prototype attraction or improved expressivity. As discussed in the manuscript, the benefit of explicit routing arises from its inductive bias under finite-capacity training, which is supported by the controlled ablation experiments.

W2 & Q2: Statistical reporting and multi-seed evaluation

Response: Thank you for pointing out the lack of clarity in our statistical reporting. Each cell in Figure 5 of the original manuscript reported a single-run result rather than a mean over five random seeds. We apologize for not clearly specifying this reporting protocol. In the revised manuscript, Figure 5 has been updated to report mean results over five random seeds, and we have checked its consistency with Table 2 for the corresponding configurations.

Given the rebuttal time constraints, we prioritized supplementary evaluation at the more challenging 70% missing rate, covering PRFM and all seven baselines on all four datasets. The means and sample standard deviations over five random seeds are reported in Appendix H.4, Table 7 of the revised manuscript, and the corresponding Table 1 entries for all methods at this missing rate have been updated to these means. We will add the five-seed means and standard deviations at the other missing rates to the manuscript once these experiments are complete. PRFM achieves the highest mean ACC on all four datasets and the best or second-best mean ARI. These supplementary results report both average performance and variability across runs, providing a more complete basis for comparing the methods.

Q3: Pseudo-label accuracy on incomplete samples

Response: At a 70% missing rate, we form the consensus representation of each incomplete sample using only its observed views and assign it to the most similar online prototype. We evaluate these assignments against the ground-truth classes on the incomplete subset using Hungarian matching. Ground-truth labels are used only for this post-hoc evaluation, not for training or prototype assignment. The pseudo-label accuracies evaluated from saved training checkpoints are reported below.

Dataset# Incomplete samplesPseudo-label accuracy (%)
HandWritten1,40053.50
YouTubeFace1027,05850.99

These moderate accuracies reflect the challenging setting of incomplete observations and the fact that the assignments are obtained from evolving online prototypes during training, rather than from the final completed representations. Importantly, these assignments serve as routing signals rather than ground-truth supervision. Their potential noise is mitigated by confidence-weighted soft-pair supervision, time-decaying prototype guidance, and a learned residual velocity that allows corrections beyond the assigned prototype.

Despite imperfect pseudo-label assignments, the ablation results in Table 2 show that removing soft-pair supervision reduces ACC by 5.32 and 8.40 points on HandWritten and YouTubeFace10, respectively, while removing explicit prototype routing reduces ACC by 2.91 and 7.46 points. These results indicate that both components remain beneficial for clustering despite noisy intermediate assignments.

Q4: Hyperparameter settings, baseline configurations, and Figure 4 consistency

Response: Thank you for these questions. Our default settings are R=5R=5, τp=0.20\tau_p=0.20, four Euler updates, λpg=0.5\lambda_{\mathrm{pg}}=0.5, and λ=0.05\lambda=0.05. We use these settings across all datasets without dataset-specific tuning. Figures 4 and 5 analyze the effects of these parameters, rather than being used to select the best-performing configuration for each dataset.

For the baselines, we use the official implementations and follow the hyperparameter configurations recommended in their original papers.

Figure 4 in the original manuscript presented single-run results. In the revised manuscript, we have updated it to report mean results over five random seeds. The values at the default settings are consistent with those in Table 2.

W3: Conditional distribution modeling and the contribution of completion

Response: We would like to clarify that single-sample inference does not imply deterministic point estimation. The key distinction lies in the learning objective and generation mechanism, rather than the number of samples drawn at inference. Unlike deterministic completion methods that directly regress missing-view representations, our method PRFM learns a conditional velocity field that transports Gaussian noise toward target-view representations. Although the ODE trajectory is deterministic given the initial noise and condition, different noise samples can produce different outputs. We use a single generated representation per missing view for efficient downstream clustering.

Regarding soft-pair supervision, the soft targets are weighted combinations of observed target-view representations, rather than prototypes themselves. Prototypes are used to prioritize candidates and determine their aggregation weights. Moreover, soft-pair supervision complements hard-pair supervision, which uses actual observed target-view representations. Both types of supervision train the conditional velocity field through flow matching rather than direct endpoint regression.

The effectiveness of generative completion is supported by our ablation experiments. As shown in Table 2, PRFM outperforms Deterministic Completion, which replaces flow matching with a direct MLP regressor, by 2.71 and 6.50 ACC points on HandWritten and YouTubeFace10, respectively.

To further examine whether the generated representations themselves contribute to clustering, we conduct an inference-time ablation while keeping the trained model, missing masks, and clustering evaluation unchanged. Specifically, we compare fusing only observed views with fusing both observed and generated views using the confidence-weighted fusion in Eq. (14). The results at a 70% missing rate are:

DatasetFused representationsACCNMIARI
HandWrittenObserved views only80.4071.3866.47
HandWrittenObserved + generated (PRFM)84.1472.3567.97
YouTubeFace10Observed views only75.8970.8662.42
YouTubeFace10Observed + generated (PRFM)78.7473.8166.94

Incorporating the generated views improves ACC, NMI, and ARI by 3.74, 0.97, and 1.50 points on HandWritten, and by 2.85, 2.95, and 4.52 points on YouTubeFace10, respectively. These results provide direct evidence that the generated representations contribute to improved clustering performance beyond the observed views alone.

For completion fidelity, we further examine whether the generated representations are close to the true missing-view representations. Under the same 70% missing-rate setting, we take fully observed samples, mask one view, generate its latent from the remaining views, and compare the generated latent with the encoder latent of the masked view. We report MSE to measure latent reconstruction error and cosine similarity to assess directional agreement between generated and observed target-view representations. We compare Full PRFM with Deterministic Completion from Table 2.

DatasetMethodMSE ↓Cosine similarity ↑
HandWrittenDeterministic Completion0.72260.6248
HandWrittenFull PRFM0.69520.7504
YouTubeFace10Deterministic Completion0.78410.6521
YouTubeFace10Full PRFM0.70600.7679

Full PRFM achieves lower MSE and higher cosine similarity than Deterministic Completion on both datasets. Both analyses have been added to Appendix H.1 of the revised manuscript.

Together, these results support the practical effectiveness of prototype-routed generative completion over direct point prediction and demonstrate the contribution of generated representations to clustering. While we do not claim to recover the true conditional distribution or its multimodal structure, our focus is on learning a conditional generative process that produces cluster-discriminative representations for incomplete multi-view clustering.

W4: Visual clarity of Figure 2

Response: Thank you for the suggestion. We have improved the visual clarity of Figure 2 in the revised manuscript.

Reviewer_qxGEOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on missing-view generation for incomplete multi-view clustering. In particular, it proposes Prototype-Routed Flow Matching (PRFM), which incorporates dynamically updated clustering prototypes directly into the Flow Matching velocity field. The velocity is decomposed into a prototype-routed component that provides cluster-level guidance and a residual component that captures sample- and view-specific information. Moreover, a prototype-prioritized soft-pair supervision strategy is introduced to construct supervision targets when paired observations are unavailable. The observed and generated representations are further combined and optimized in a shared prototype space to improve clustering. Experiments on multiple benchmark datasets under different missing rates demonstrate competitive clustering performance, while few-step ODE inference offers improved efficiency over diffusion-based completion methods.

Critical strengths

The main idea is the explicit use of clustering prototypes to shape the Flow Matching velocity field, rather than simply conditioning the generator.

The residual component allows the transport process to retain flexibility beyond cluster-level guidance.

PRFM uses a shared prototype space to connect missing-view generation, soft-pair supervision, and clustering optimization.

Proposition 1 further clarifies the role of prototype routing as an inductive bias rather than a strict constraint on the velocity field.

Critical weaknesses

The paper explores an interesting approach to develop a prototype-guided generative framework for incomplete multi-view clustering. However, several technical details could be explained more clearly. For example, the paper uses different confidence weights for soft-pair supervision and completed-view fusion, but their respective roles and differences are not sufficiently discussed. Similarly, Eq. (5) introduces a learnable target-view embedding without clearly explaining its parameter sharing and use during inference.

Another point concerns the prototype guidance coefficient: Eq. (6) assumes λpg<1\lambda_{\mathrm{pg}}<1, while the sensitivity analysis also considers λpg=1\lambda_{\mathrm{pg}}=1. Although this is briefly addressed in the appendix, the distinction between the theoretical assumptions and experimental settings could be clearer.

In addition, Eq. (16) applies the clustering-alignment loss to all samples, including those with complete observations, but the motivation for this choice is not fully explained. Finally, the discussion of related methods such as ProImp, RPCIC, DCG, and NCD is somewhat limited, particularly regarding their differences in prototype usage, missing-view generation, and supervision strategies.

Questions for authors

Please refer to the weakness

2 comments

We thank the reviewer for the constructive comments. We clarify the technical details and the relationship to prior work below.

Q1: Confidence weights for supervision and fusion

Response: The two weights serve different purposes. The soft-pair confidence ρiu\rho_i^u in Eq. (10) weights the velocity-regression loss according to the prototype-assignment confidence of the anchor and selected candidates. It controls the contribution of each constructed supervision target.

In contrast, ωiv\omega_i^v in Eq. (14) controls each view's contribution to the completed representation. Observed views use encoder confidence, while generated views are weighted using the mean observed-view confidence and their agreement with the observed consensus and assigned prototype, as specified in Eq. (20). Thus, ρiu\rho_i^u weights supervision, whereas ωiv\omega_i^v weights representation fusion. They are calculated separately, although both use prototype-related information.

Q2: Target-view embedding

Response: The target-view embedding eue_u in Eq. (5) is a learnable vector shared across all samples whose generation target is view uu. It is learned during training and remains fixed during inference. Together with the assigned prototype and observed-view consensus, it forms the condition for the shared velocity network. Its role is to identify which view to generate; it does not contain the missing observation itself.

Q3: The boundary setting λpg=1\lambda_{\mathrm{pg}}=1

Response: The restriction λpg<1\lambda_{\mathrm{pg}}<1 ensures that the residual coefficient 1−apg(t)1-a_{\mathrm{pg}}(t) remains positive throughout generation, so the residual-field expression in Proposition 1 is well-defined. When λpg=1\lambda_{\mathrm{pg}}=1, the forward velocity in Eq. (6) remains defined, but the residual coefficient is zero at t=0t=0. The proposition's residual-field expression therefore does not apply at that boundary point.

As stated in Appendix E, λpg=1\lambda_{\mathrm{pg}}=1 is included only as a sensitivity stress test. The default setting is λpg=0.5\lambda_{\mathrm{pg}}=0.5, and the main comparisons use λpg<1\lambda_{\mathrm{pg}}<1.

Q4: Applying the clustering-alignment loss to complete samples

Response: Lcomp\mathcal{L}_{\mathrm{comp}} optimizes the cluster assignments of fused representations; it is not a reconstruction loss restricted to missing views. For a fully observed sample, Eq. (14) reduces to confidence-weighted fusion of its observed latents, without generating any missing-view representation.

Including complete samples allows all available observations to contribute to the shared clustering objective. The terms in Eq. (16) sharpen the completed assignments and align them with targets derived from observed-view assignments. For complete samples, this provides clustering regularization on the observed representations rather than an additional imputation task.

Q5: Relationship to ProImp, RPCIC, DCG, and NCD

Response: We clarify these relationships below. ProImp completes missing views using target-view prototypes and sample–prototype relationships transferred from observed views. RPCIC further addresses noisy cross-view prototype correspondence. DCG combines diffusion-based generation with cross-view contrastive alignment, while NCD uses neighborhood structure and missingness information to condition denoising.

PRFM differs in how clustering structure enters the generation process. Rather than using prototypes only as reconstruction anchors or conditioning inputs, PRFM incorporates prototypes updated online during training directly into the Flow Matching velocity field. The velocity is decomposed into a prototype-routed component that provides explicit, state-dependent cluster-level guidance and a learned residual component that models sample- and view-specific variation. Prototype guidance is therefore built into the transport dynamics rather than left entirely to the network to learn from conditioning inputs.

The same prototypes also prioritize observable candidates to construct soft supervision targets when paired observations are unavailable. PRFM thus connects prototype-directed transport and prototype-guided supervision, rather than simply replacing diffusion with Flow Matching. The comparisons with Standard FM and Prototype-Cond. FM in Table 2 directly examine the benefit of explicit routing beyond standard flow matching and prototype conditioning.

We have added these clarifications to Appendices A, E, and H.3 of the revised manuscript.

The author did a great job of answering all my questions.

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

This paper proposes Prototype-Routed Flow Matching (PRFM) for incomplete multi-view clustering. Unlike conventional conditional Flow Matching, PRFM explicitly incorporates online clustering prototypes into the velocity field. The prototype-routed velocity provides cluster-level guidance, while the residual velocity captures sample- and view-specific variations.

Critical strengths
  1. Rather than simply using prototypes as conditioning inputs, PRFM explicitly incorporates dynamic cluster-level guidance into the transport process, while the residual branch preserves flexibility for sample-specific variations.

  2. The prototype-prioritized soft-pair supervision instead of directly selecting similar samples for completion, the method first identifies candidates based on prototype assignments and then constructs soft supervision targets through similarity-weighted aggregation. This allows the prototypes to guide both the generation process and the construction of supervision targets when paired observations are unavailable.

  3. The paper conducts extensive comparative and ablation experiments to verify the effectiveness of the method.

Critical weaknesses
  1. PRFM jointly optimizes completed representations through a clustering-alignment loss. However, it is unclear whether gradients from this loss propagate through the ODE generation process to the velocity network.

  2. The method combines hard-pair and soft-pair supervision, but some mini-batches may contain no valid pairs of a particular type, especially at high missing rates. The handling and normalization of these loss terms in such cases are not clearly specified.

  3. The method updates clustering prototypes using EMA, but the exact order of prototype assignment, model optimization, and prototype updates within each mini-batch could be clarified.

Questions for authors

See Weaknesses

2 comments

We sincerely thank the reviewer for the careful and constructive review, and for recognizing the strengths of our prototype-guided design. We clarify the gradient flow, loss normalization, and prototype update sequence below.

Q1: Gradient propagation through ODE generation

Response: During joint training, the Euler updates are unrolled within the computation graph. The clustering-alignment loss Lcomp\mathcal{L}_{\mathrm{comp}} therefore backpropagates through the generated latents and integration steps to the residual velocity network. The assigned prototype and observed-view consensus used as generation conditions are detached, but the generated trajectory is not. Thus, during joint training, clustering supervision can update the residual velocity network through this path when the loss yields nonzero gradients.

Q2: Empty pairs and loss normalization

Response: Hard- and soft-pair losses are computed separately for each target view. A view contributes to hard supervision only when at least two samples have both the target view and another observed view. For soft supervision, a view is skipped if it has no missing samples or no observed candidates. For a missing pair without same-prototype candidates, the candidate pool falls back to all samples in the current mini-batch with the target view observed, following Eq. (8).

Within each valid view, the hard loss in Eq. (7) is averaged over the valid hard pairs, and the soft loss in Eq. (10) is averaged over the valid soft pairs, with each pair weighted by its confidence ρiu\rho_i^u. The soft loss is not divided by the sum of the confidence weights. Each pair-type loss is then averaged across its respective valid views. If no view provides valid supervision for a pair type, its loss is set to zero, avoiding empty-set normalization.

Q3: Order of prototype updates

Response: Observed-view assignments are first computed using the current prototype bank. These assignments are then used for a gradient-free EMA update within the forward pass, before the optimizer step. Missing-view generation uses the prototypes selected before this update, whereas completed representations are assigned using the updated bank. Hard-pair conditions use a snapshot of the pre-update bank, with assignments recomputed from the remaining observed views after excluding the target view.

After computing the losses, the model parameters are updated by the optimizer.

We have added these clarifications to Appendix A and updated Algorithm 1 accordingly in the revised manuscript.

Reviewer_UMHqOpenReview ↗

My concerns had been resolved.

All my concerns have been addressed, and I am happy to recommend acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.