acceptodds
Under review as a conference paper at ICLR 2027

Semantic-Guided Cross-View Latent Transport via Flow Matching for Incomplete Multi-View Clustering

Abstract

This paper presents TRACE, a semantic-guided cross-view latent transport frame- work via conditional flow matching for incomplete multi-view clustering. Specifi- cally, TRACE first learns a stable dual-level representation that separates compact semantic codes from view-specific complete latent representations, providing clustering-oriented semantic conditions for cross-view transport while preserving view-specific information. Building on these semantic conditions, we introduce a target-excluded conditional flow matching strategy that temporarily treats an observed view as a pseudo-missing target and learns source-to-target transport routes in the frozen latent space without accessing target-view information during conditioning. Furthermore, TRACE employs an observed-priority multi-source recovery mechanism, in which each missing view is independently recovered from all authentic observed views and the resulting target candidates are aggregated, while generated views are prevented from participating in subsequent recovery. This design avoids recursive error propagation and preserves the information of genuinely observed views. Extensive experiments on five datasets demonstrate that TRACE achieves competitive or superior performance against recent state-of-the- art methods.

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

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

This paper proposes TRACE, a flow-matching-based framework for incomplete multi-view clustering. The method learns view-specific latent representations and compact semantic codes in a first stage, then trains a shared velocity field to transport representations from observed views toward pseudo-missing views. During inference, representations generated from multiple observed views are aggregated to recover missing views, while the original observed representations are preserved.

Critical strengths

The two-stage design is reasonable and helps separate representation learning from missing-view recovery. Excluding the target view's semantic code from the transport condition is also a sensible choice to reduce potential information leakage during training.

Critical weaknesses
  1. The ablation study is limited to one dataset and one missing-view ratio. It is therefore difficult to determine whether the contribution of each component remains consistent under other settings.

  2. The method combines several recovered representations during inference, but the effect of this aggregation strategy is not examined separately.

Questions for authors
  1. Would the main conclusions of the ablation study remain the same under a different missing-view ratio?

  2. How does the performance change when the method uses only one observed view to recover a missing view instead of combining multiple candidates?

1 comment

Q1: Would the main conclusions of the ablation study remain the same under a different missing-view ratio?

A1: We thank the reviewer for this question on the robustness of the ablation. We repeated the HandWritten ablation at 10%, 30%, 50%, and 70% with the masks and seeds of the corresponding Table 1 entries, so the Full row equals Table 1 at every rate. Entries are ACC / NMI / ARI (%), averaged over five seeds.

Variant10%30%50%70%
Full TRACE93.07 ± 1.26 / 87.18 ± 1.63 / 85.38 ± 2.3792.77 ± 0.82 / 86.54 ± 1.10 / 84.77 ± 1.6092.78 ± 0.73 / 86.62 ± 0.84 / 84.81 ± 1.2892.95 ± 1.38 / 86.78 ± 1.67 / 85.17 ± 2.56
w/o Unified Semantic Objective89.29 ± 1.77 / 81.88 ± 1.26 / 78.44 ± 2.8089.97 ± 2.95 / 82.78 ± 3.19 / 79.77 ± 4.9088.14 ± 2.23 / 80.51 ± 2.31 / 76.43 ± 3.6891.30 ± 0.92 / 83.88 ± 1.36 / 81.90 ± 1.73
w/o Semantic Guidance92.78 ± 0.79 / 86.59 ± 1.12 / 84.79 ± 1.4591.91 ± 2.20 / 85.70 ± 2.35 / 83.30 ± 3.8292.33 ± 1.52 / 85.79 ± 1.82 / 83.95 ± 2.7392.38 ± 1.41 / 85.89 ± 1.69 / 84.06 ± 2.57
w/o Flow Matching92.33 ± 1.16 / 86.08 ± 1.32 / 84.08 ± 2.0792.64 ± 1.50 / 86.53 ± 1.44 / 84.65 ± 2.5592.52 ± 1.24 / 86.38 ± 1.26 / 84.40 ± 2.1892.67 ± 1.41 / 86.51 ± 1.70 / 84.61 ± 2.58
w/o Observed Priority90.10 ± 0.88 / 81.80 ± 1.05 / 79.51 ± 1.6489.71 ± 0.90 / 81.55 ± 0.94 / 78.87 ± 1.6589.12 ± 1.11 / 80.52 ± 1.49 / 77.63 ± 2.1290.16 ± 1.04 / 81.93 ± 1.39 / 79.67 ± 1.91

The conclusions hold across rates: Full is best in all three metrics at every rate. The 70% results are in Table 2 and all four rates, with standard deviations, in Table 4 (Appendix B). This confirms that every component contributes at all missing rates.

Q2: How does the performance change when the method uses only one observed view to recover a missing view instead of combining multiple candidates?

A2: We compared TRACE's multi-source aggregation with a single-source variant that keeps one authentic observed source per missing target, chosen by a fixed rule (sample index, target view, and seed). Both keep the same target-excluded semantic condition built from all available views, so the comparison isolates candidate aggregation; encoders, velocity field, integration budget, fusion, and clustering are unchanged. Entries average the four missing rates and five seeds per rate, as in Table 1. R²_rec = 1 − MSE_rec / MSE_mean is measured on genuinely hidden views against mean imputation; hidden values are used only for evaluation.

DatasetACC singleACC multiR²_rec singleR²_rec multi
BDGP39.4839.57-0.2700.038
BBC74.2074.530.1280.466
OutdoorScene76.7676.760.4870.621
HandWritten92.4492.890.4350.615
ALOI-10084.3984.610.6110.708

Aggregation gives higher recovery R² on all five datasets. The comparison is reported in Table 7 (Appendix B). Multi-source aggregation therefore consistently improves recovery quality and never lowers the average ACC on any dataset, supporting the use of all authentic observed views.

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

This paper studies incomplete multi-view clustering from the perspective of cross-view latent transport. Instead of directly completing missing features in the original space, TRACE learns a flow-based model to generate latent representations for unobserved views. The model is trained using observed views as temporary targets, and multiple source views are used to estimate each missing representation at test time. The paper evaluates the approach on five datasets and compares it with several SOTA methods.

Critical strengths
  1. Excluding the target view from the semantic condition is a reasonable design. It allows the target representation to provide training supervision without requiring its semantic code at inference.

  2. The two-stage framework is easy to follow. Learning representations first and then freezing them gives the transport model a fixed latent space to work with.

Critical weaknesses
  1. The motivation for using a shared velocity field across different views could be further clarified, particularly when the views have distinct feature distributions and semantic characteristics.

  2. The method averages generated representations from different observed views with equal weights, but its behavior under substantial differences in view quality remains unclear.

  3. The comparison methods are not sufficiently up to date. The paper lacks comparisons with some methods from 2026, making it difficult to establish the effectiveness of TRACE.

  4. Some numbers differ across the main results, ablation study, and convergence analysis. The paper does not explain whether these differences come from random seeds, training configurations, or reporting procedures.

Questions for authors
  1. Why is a shared velocity field suitable for modeling transitions across different views?

  2. How does TRACE handle differences in information quality across observed views?

  3. Since different views may have different feature dimensions and distributions, could the authors clarify how they are mapped into a shared latent space?

3 comments

Q1: Why is a shared velocity field suitable for modeling transitions across different views?

A1: We thank the reviewer for the careful questions on our design. The field is shared but conditioned on the current latent, time, the target-excluded semantic condition and learned source and target view embeddings (Equation 10). The embeddings identify the ordered route, so one network represents different transformations for different view pairs. Stage I maps views to latents of a common dimension that keep view-specific information through reconstruction, and Stage II trains the field on frozen source and target endpoints, so view identities and route-specific supervision account for view differences. Sharing learns all routes jointly without a network per ordered pair.

Q2: How does TRACE handle differences in information quality across observed views?

A2: Each source passes through its own encoder and is transported with source- and target-specific conditioning, so different sources can yield different candidates for the same target. The semantic condition is the L2-normalized mean of the available semantic codes (Equation 8); candidates from all authentic sources are averaged with equal weights in the target latent space, not in raw feature space (Equation 13), and fill only the missing slot, keeping observed representations (Equation 14). Equal averaging needs no external reliability ranking; the weights are fixed, not learned from view-quality estimates.

Q3: Since different views may have different feature dimensions and distributions, could the authors clarify how they are mapped into a shared latent space?

A3: Each view is first processed by a view-specific preprocessor fitted only on its observed rows (Section 3.1); together with the view-specific encoder E_v, this maps each view, whatever its dimension D_v, to a full latent of common dimension d, and a semantic head H_v gives a compact semantic code (Equation 1). The full latents keep view-specific information through reconstruction, while the semantic codes are aligned across views, so a common dimension does not imply identical distributions. Stage II then learns the remaining source-to-target transformation between the frozen latents, guided by the semantic condition and view embeddings.

Weakness 3: The comparison methods are not sufficiently up to date. The paper lacks comparisons with some methods from 2026, making it difficult to establish the effectiveness of TRACE.

Response 3: Thank you for this suggestion. In addition to OPTION (ICML 2026), which was already included, the revised Table 1 now includes two further 2026 methods with public code: ACOVA (ICML 2026) and ERL-Net (IEEE TNNLS 2026).

DatasetMethod10%30%50%70%
BDGPACOVA23.30 / 1.12 / 0.4523.32 / 0.82 / 0.3922.90 / 0.71 / 0.3923.01 / 0.76 / 0.46
BDGPERL-Net30.51 / 5.49 / 4.4329.27 / 4.64 / 3.5227.66 / 3.79 / 2.5126.90 / 3.55 / 2.18
BDGPTRACE (Ours)41.10 / 16.40 / 13.3039.15 / 15.29 / 12.0839.02 / 13.62 / 11.5239.00 / 13.04 / 11.06
BBCACOVA54.66 / 41.49 / 27.7456.76 / 39.95 / 28.1854.04 / 40.05 / 28.0657.34 / 42.05 / 29.50
BBCERL-Net63.50 / 43.23 / 39.6557.34 / 39.91 / 35.2260.18 / 43.63 / 39.6552.15 / 29.90 / 26.55
BBCTRACE (Ours)71.42 / 63.17 / 55.1771.30 / 63.24 / 56.1774.72 / 66.62 / 62.1680.70 / 68.80 / 66.51
OutdoorSceneACOVA73.49 / 58.68 / 51.4372.51 / 59.01 / 51.6473.79 / 59.20 / 52.2472.54 / 59.07 / 51.58
OutdoorSceneERL-Net76.76 / 63.45 / 56.2476.38 / 62.63 / 55.9376.56 / 62.62 / 55.8874.76 / 60.75 / 53.65
OutdoorSceneTRACE (Ours)77.72 / 62.86 / 57.2877.46 / 62.34 / 56.7476.96 / 61.83 / 55.9474.88 / 60.47 / 53.81
HandWrittenACOVA92.41 / 85.82 / 84.3292.33 / 85.24 / 83.9192.40 / 85.17 / 84.0292.44 / 85.70 / 84.31
HandWrittenERL-Net82.18 / 80.42 / 73.0783.61 / 81.65 / 74.8382.16 / 80.76 / 73.8382.55 / 81.38 / 74.51
HandWrittenTRACE (Ours)93.07 / 87.18 / 85.3892.77 / 86.54 / 84.7792.78 / 86.62 / 84.8192.95 / 86.78 / 85.17
ALOI-100ACOVA29.24 / 48.86 / 15.5128.01 / 47.00 / 14.3923.61 / 43.25 / 10.9422.15 / 42.53 / 10.14
ALOI-100ERL-Net77.17 / 90.92 / 73.7576.70 / 90.90 / 73.9775.93 / 90.26 / 72.9277.80 / 90.46 / 74.47
ALOI-100TRACE (Ours)86.77 / 90.85 / 79.3086.00 / 90.30 / 78.4284.72 / 89.65 / 77.0280.62 / 87.97 / 72.60

TRACE achieves the highest ACC in all 20 dataset–rate settings and the highest ARI in 19 of them, which supports its effectiveness against these recent methods.

Weakness 4: Some numbers differ across the main results, ablation study, and convergence analysis. The paper does not explain whether these differences come from random seeds, training configurations, or reporting procedures.

Response 4: We thank the reviewer for catching these inconsistencies, and we apologize for them. We traced every number to its run. Table1,Table 2 and the numbers in Section 4.5 came from earlier development runs with different seeds and a smaller training budget, which the paper did not state. We reran them with the configuration, seeds, and masks of Table 1.

  • Table 1, HandWritten, TRACE (ACC / NMI / ARI): 10% 93.07 / 87.18 / 85.38; 30% 92.77 / 86.54 / 84.77; 50% 92.78 / 86.62 / 84.81; 70% 92.95 / 86.78 / 85.17.
  • Table 2, HandWritten, 70% (ACC / NMI / ARI): Full TRACE 92.95 / 86.78 / 85.17; w/o Unified Semantic Objective 91.30 / 83.88 / 81.90; w/o Semantic Guidance 92.38 / 85.89 / 84.06; w/o Flow Matching 92.67 / 86.51 / 84.61; w/o Observed Priority 90.16 / 81.93 / 79.67.
  • Section 4.5, HandWritten, 50%: during Stage I, ACC / NMI / ARI 91.30 / 83.94 / 81.94 → 92.52 / 86.38 / 84.40 and the representation objective 1.5438 → 1.1946; after 1,800 Flow updates, ACC / NMI / ARI 92.78 / 86.60 / 84.82 and the Flow Matching objective 0.01728 → 0.00196.

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_88PLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of clustering data when some views are missing. The method learns a representation for each view and uses a flow-matching model to transfer information from observed views to missing ones. The model is trained by treating some available views as if they were missing. At test time, it estimates the missing representations and uses them together with the observed representations to perform clustering. The experiments compare TRACE with several existing methods on five datasets.

Critical strengths

The paper provides a clear pipeline for handling missing views and evaluates the proposed method on multiple datasets. The paper also presents the main procedure in a relatively accessible way, making the proposed approach easier to follow.

Critical weaknesses

The actual efficiency of TRACE has not been fully evaluated. Comparing the running time with related benchmark methods can help assess its practical applicability.

The evaluation mainly focuses on downstream clustering performance, while the quality of the generated representations is not directly examined.

Questions for authors
  1. How does the inference time of TRACE compare with that of other relevant baselines?

  2. Could the authors provide an additional analysis of the generated representations to clarify their contribution to the final clustering performance?

  3. Since the inference procedure generates candidates from multiple observed views, how does its computational cost change as the number of views increases?

  4. The flowchart shows three steps, while the paper describes two stages. Could the authors clarify this inconsistency?

3 comments

Q1: How does the inference time of TRACE compare with that of other relevant baselines?

A1: We thank the reviewer for suggesting an inference-time comparison. We benchmarked TRACE, OPTION, and DCG on one RTX 3090 at a 50% missing rate (5 warm-up runs and 30 timed repetitions per seed, CUDA synchronization), reporting the mean of per-seed medians over the full dataset. Feature inference runs from preprocessed inputs to clustering features (encoding, recovery, fusion, and data transfer); inference with K-means also runs K-means.

DatasetFeature inference (ms): TRACE / OPTION / DCGWith K-means (s): TRACE / OPTION / DCG
BDGP21.19 ± 1.38 / 676.96 ± 84.15 / 230.97 ± 22.410.191 ± 0.018 / 0.767 ± 0.086 / 0.326 ± 0.016
BBC49.06 ± 4.59 / 217.01 ± 9.60 / 187.00 ± 2.570.112 ± 0.010 / 0.258 ± 0.012 / 0.238 ± 0.006
OutdoorScene38.56 ± 0.37 / 706.44 ± 70.78 / 220.14 ± 3.500.547 ± 0.036 / 0.840 ± 0.019 / 0.358 ± 0.013
HandWritten88.93 ± 9.14 / 739.77 ± 90.58 / 206.40 ± 0.920.668 ± 0.018 / 0.814 ± 0.084 / 0.348 ± 0.014
ALOI-10042.59 ± 0.48 / 2856.25 ± 266.21 / 387.29 ± 11.4816.009 ± 1.042 / 9.594 ± 0.608 / 7.444 ± 0.265

In feature inference, TRACE is 2.3–10.9× faster than DCG and 4.4–67× faster than OPTION. With K-means, it is fastest on BDGP and BBC, faster than OPTION but slower than DCG on OutdoorScene and HandWritten, and slowest on ALOI-100. Table 5 (Appendix B) reports both scopes with standard deviations.

Q2: Could the authors provide an additional analysis of the generated representations to clarify their contribution to the final clustering performance?

A2: Following this suggestion, we analyzed the generated representations. Both analyses use a 50% missing rate and five seeds, as in Table 1. (i) Recovery quality: on genuinely missing views, we compare the generated latent with the frozen encoder's output on the held-out view, used only as an evaluation target, via R²_rec = 1 − MSE_gen / MSE_mean, where the reference imputes the mean observed latent. (ii) Contribution: with the same trained model, we compare TRACE with leaving missing slots as zero placeholders; all other modules are unchanged.

DatasetR²_recACC without recoveryACC with TRACEGain
BDGP0.03536.0539.02+2.97
BBC0.44568.8574.72+5.87
OutdoorScene0.61870.4576.96+6.50
HandWritten0.60492.5292.78+0.26
ALOI-1000.70364.8884.72+19.84

Recovery beats mean imputation and improves ACC on all five datasets. See Table 6 (Appendix B). This confirms that the generated representations are informative and directly contribute to clustering, with the largest gain on ALOI-100 (+19.84 ACC).

Q3: Since the inference procedure generates candidates from multiple observed views, how does its computational cost change as the number of views increases?

A3: A sample with o_i observed and m_i missing views needs one transport per source–target pair, i.e., o_i·m_i transports of K Euler steps; generated views are never reused as sources, so the cost scales with K·Σ_i o_i·m_i.

To measure the cost of extra sources, we fixed the sample count, and increased the number of observed sources with the trained checkpoints (RTX 3090, 5 warm-up and 30 timed runs per seed). Times (ms) cover recovery from cached latents only.

DatasetN1 source2 sources3 sources4 sources5 sources
BDGP2,5003.00 ± 0.035.14 ± 0.10–––
BBC6853.91 ± 0.037.05 ± 0.1210.06 ± 0.14––
OutdoorScene2,6883.07 ± 0.095.15 ± 0.057.35 ± 0.12––
HandWritten2,0003.02 ± 0.065.09 ± 0.067.19 ± 0.099.27 ± 0.0911.90 ± 1.13
ALOI-10010,8003.78 ± 0.056.44 ± 0.109.03 ± 0.10––

Recovery time grows approximately linearly over one to five sources for a fixed target. Full-dataset inference batches by view pair, so its runtime also depends on batch sizes and GPU utilization and need not be linear in V. See Table 8 (Appendix B). Each additional source adds only about 2–3 ms, so aggregating all observed views remains computationally cheap.

Q4: The flowchart shows three steps, while the paper describes two stages. Could the authors clarify this inconsistency?

A4: Thank you for pointing this out, and we apologize for the confusion. The reviewer is right: the submitted Figure 1 was inconsistent with the text. TRACE has two training stages: Stage I trains the representation parameters θ, and Stage II freezes θ and trains only the velocity field and view embeddings φ; nothing is trained afterwards. However, the submitted Figure 1(a), which followed the inference pipeline, mislabelled its blocks as stages: its "Stage II" block actually showed the trained field recovering missing views at inference (its training is in Figure 1(b)), and its "Stage III" block only applies the frozen heads, fusion, and K-means, with no trainable parameters. We have corrected this in the revision: these blocks are now labelled "Inference · Multi-source recovery" and "Inference · Fusion and clustering", Figure 1(b) is labelled "Stage II · Conditional flow matching training", and Section 3.1 states the two training stages followed by inference. Accordingly, we will update the flowchart (Figure 1) in the revised version of the paper.

Thank you to the authors for their detailed response. They have addressed my concerns, and I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.