Pith. sign in

REVIEW 3 major objections 7 minor 51 references

Straight-Path Flow Matching for Incomplete Multi-View Clustering

T0 review · 3 major / 7 minor · reviewed 2026-07-08 · glm-5.2

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Straight-line flow paths beat diffusion for clustering with missing views

desk verdict The theory has a real gap (non-convex cluster sets), but the method works well in practice and the flow-matching-for-IMVC idea is a genuine new application. read the letter →

arxiv 2607.06281 v1 pith:SITJ3AED submitted 2026-07-07 cs.CV

Yiteng Yuan , Junyan Wang , Zheyuan Liu , Hong Jia , Lei Fan , Zhulin Tao , Lianbo Guo This is my paper · ORCID
classification cs.CV
keywords flowmatchingincompletemulti-viewclusteringdeterministicODEclusterseparabilityviewcompletiongenerative
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that when recovering missing views in multi-view clustering, the generative transport mechanism matters as much as the model capacity. The authors replace the standard diffusion-based approach—which starts from cluster-agnostic Gaussian noise and follows a stochastic denoising path—with a deterministic flow-matching formulation that transports an observed view representation to its missing counterpart along a straight-line interpolation path. The central claim is twofold: (1) this straight path is the shortest possible route between paired representations, and its length shrinks adaptively as cross-view alignment improves during training, unlike diffusion whose path length is fixed by the distance from noise to data; and (2) because the interpolation stays within the cluster region of the observed view, trajectories from different clusters never intersect at any finite number of integration steps, whereas diffusion only achieves cluster separability asymptotically (when noise variance drops below half the inter-cluster distance), a condition frequently unmet in finite-step inference. The paper builds an end-to-end architecture combining this flow-matching completion with cluster-level contrastive alignment and entropy-weighted probability fusion, and reports state-of-the-art results on five benchmarks.

What carries the argument

Flow matching with linear interpolation path; deterministic ODE vector field trained via mean-squared velocity matching; bidirectional reversibility (reverse vector field is negation of forward); cluster-level InfoNCE contrastive alignment; entropy-weighted centered log-ratio fusion for probability distributions; completion consistency loss enforcing bidirectional reconstruction fidelity.

What would settle it

If one could construct a multi-view dataset where same-sample view representations naturally fall in different cluster regions (violating view consistency), the straight-line interpolation would pass through inter-cluster space and the finite-step separability guarantee would fail.

Watch

Extended reading notes

Core claim

The core object is the linear interpolation path x_t = (1-t)z1 + t*z2 between paired latent representations of two views. The paper proves two corollaries about this path under three shared assumptions (cluster separability, view consistency, sufficient model approximation). Corollary 1 establishes that the straight-line path is the Euclidean shortest path between paired views, with length ||z2 - z1|| that adapts as training aligns the representations—contrasted with diffusion, whose path from noise to data is fixed and cluster-agnostic. Corollary 2 establishes that because both endpoints lie in the same cluster set (by the view consistency assumption), the ODE uniqueness theorem guarantees:

Load-bearing premise

The load-bearing premise is Assumption (ii), View Consistency: that different-view representations of the same sample already belong to the same cluster. The finite-step separability guarantee of Corollary 2 depends on both z1 and z2 residing in the same cluster set S_{c(z1)}, so that the straight-line interpolation between them stays within that cluster. If view representations for the same sample are not already cluster-consistent—a common situation in real multi-view data—

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes a flow-matching framework for incomplete multi-view clustering (IMVC) that replaces diffusion-based stochastic trajectories with deterministic ODE flows using linear interpolation between paired view representations. The authors provide a theoretical analysis arguing that flow matching is structurally better aligned with clustering objectives than diffusion, supported by two corollaries: (1) flow matching constructs an optimal shortcut path, and (2) flow matching achieves cluster separability within any finite number of steps. The method integrates straight-path completion with cluster-level contrastive alignment and entropy-based alignment in an end-to-end architecture. Experiments on five benchmarks demonstrate state-of-the-art performance across multiple missing rates.

Significance. The paper addresses a relevant problem in IMVC and the application of flow matching to cross-view completion is a reasonable and novel formulation. The bidirectional reversibility property (Appendix D), which allows training a single vector field for both directions, is a practical advantage. The ablation study (Table 2) is informative, particularly the dramatic drop when L_cluster is removed, which empirically validates the importance of cross-view alignment. The flow-step analysis (Fig. 5b) provides useful empirical support for the finite-step separability claim. The experimental coverage across five datasets and three missing rates is adequate.

major comments (3)
  1. Corollary 2's proof (Appendix C, Eq. 27-28) contains a geometric gap that undermines the central theoretical claim of finite-step cluster separability. The proof claims that for z1, z2 in S_{c(z1)}, the interpolation path x_t = (1-t)z1 + t*z2 satisfies x_t in S_{c(z1)} for all t in [0,1]. However, Assumption (i) only states that cluster sets S_k are compact subsets of R^d with positive minimum inter-cluster distance delta. Compactness does not imply convexity. For a non-convex compact cluster set, the Euclidean line segment between two interior points can exit the set and potentially enter another cluster's region S_j. The proof needs either an explicit convexity assumption on S_k or a different argument for why the linear interpolation path remains within the cluster set. This is load-bearing because Corollary 2 is one of the two main theoretical contributions and is cited as motivation
  2. Assumption (ii) (View Consistency, Sec. 3.1) states that different-view representations of the same sample belong to the same cluster, i.e., c(z_v) = c(z_v'). This assumption is circular with respect to the IMVC objective: the method's goal is to recover missing-view representations such that both views correspond to the same cluster, yet the theoretical guarantees (both corollaries) require this property to already hold. The authors should clarify whether Assumption (ii) is intended to hold for the training data only (where paired complete samples are available) or for the representations at inference time. If the latter, the assumption presupposes the method's success. This should be explicitly acknowledged as a limitation or the scope of the theoretical claims should be restricted accordingly.
  3. The proof of Corollary 2 (Appendix C) invokes the ODE uniqueness theorem to claim non-intersection of trajectories from different clusters. However, ODE uniqueness prevents two trajectories with different initial conditions from crossing at the same time t, but it does not, by itself, prevent a trajectory from leaving its cluster set. The non-intersection argument conflates trajectory uniqueness (a property of the ODE solution) with cluster-set containment (a geometric property of the interpolation path). These are distinct claims and the proof does not adequately bridge them. The authors should separate these two aspects and clarify what role ODE uniqueness actually plays.
minor comments (7)
  1. Table 3 (Appendix F.1) shows that the hyperparameters lambda_1 and lambda_2 vary by orders of magnitude across datasets (e.g., lambda_1 ranges from 1.0 to 1000.0). This dataset-specific tuning weakens the claim of method robustness. The authors should discuss what guides these choices or whether a more principled selection procedure exists.
  2. In Sec. 4.1, the method uses a single integration step (Step=1) during both training and inference. While this is computationally efficient, the flow-step analysis in Fig. 5b only covers HandWritten and Synthetic3D. It would strengthen the paper to include flow-step analysis on a more challenging dataset (e.g., CUB or LandUse-21) where cluster structure is less clear.
  3. The paper states (Sec. 3.2) that the vector field F(x_t, t) takes 'the observed view z1 as input,' but Eq. (8) defines the loss as ||F(x_t, t) - (z2 - z1)||^2 without explicitly showing z1 as a conditioning input to F. The architecture of F should be clarified.
  4. In Table 1, some results show the method is outperformed by baselines on specific metrics (e.g., LandUse-21 at tau=0.3, DCG achieves NMI 32.09 vs. ours 31.71). The paper's claim of 'consistently achieves the best or competitive performance' is accurate but could be more precise about where the method does not improve over the best baseline.
  5. The related work section (Sec. 2) does not discuss any existing work applying flow matching to multi-view learning or clustering. If such work exists, it should be cited and differentiated; if not, this absence should be noted as a novelty claim.
  6. Fig. 5a is labeled '(a) t-SNE Visualization' but the subfigure label in the caption refers to '(a) Flow-Step Analysis.' The labels appear swapped or mislabeled.
  7. The paper uses 'c(z1)' notation for cluster assignment but does not formally define c(.) as a function until Sec. 3.1. An earlier forward reference or brief note would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee raises three interconnected concerns about the proof of Corollary 2 (Appendix C), specifically regarding a missing convexity assumption, the circularity of Assumption (ii), and the conflation of ODE uniqueness with cluster-set containment. We agree that the proof as written has genuine gaps. Below we address each point and describe the revisions we will make.

read point-by-point responses

  1. Referee: Corollary 2's proof (Appendix C, Eq. 27-28) contains a geometric gap that undermines the central theoretical claim of finite-step cluster separability. The proof claims that for z1, z2 in S_{c(z1)}, the interpolation path x_t = (1-t)z1 + t*z2 satisfies x_t in S_{c(z1)} for all t in [0,1]. However, Assumption (i) only states that cluster sets S_k are compact subsets of R^d with positive minimum inter-cluster distance delta. Compactness does not imply convexity. For a non-convex compact cluster set, the Euclidean line segment between two interior points can exit the set and potentially enter another cluster's region S_j. The proof needs either an explicit convexity assumption on S_k or a different argument for why the linear interpolation path remains within the cluster set. This is load-bearing because Corollary 2 is one of the two main theoretical contributions and is cited as motivation

    Authors: The referee is correct. Compactness alone does not imply convexity, and the linear interpolation between two points in a non-convex compact set can exit the set. The proof of Corollary 2 as written implicitly relies on this geometric property without stating it. We will revise the manuscript to add an explicit convexity assumption on the cluster sets S_k (or, equivalently, state that the theoretical claims hold under the additional condition that each S_k is convex). We will add this as part of Assumption (i) or as a separate assumption, and we will revise the proof of Corollary 2 accordingly. We will also add a remark acknowledging that convexity of cluster sets is an idealization and that in practice the learned representations may not satisfy this exactly, but that the alignment losses (L_cluster and L_entropy) are designed to encourage approximately convex, well-separated cluster regions, which is empirically supported by the t-SNE visualizations in Fig. 5a. revision: yes

  2. Referee: Assumption (ii) (View Consistency, Sec. 3.1) states that different-view representations of the same sample belong to the same cluster, i.e., c(z_v) = c(z_v'). This assumption is circular with respect to the IMVC objective: the method's goal is to recover missing-view representations such that both views correspond to the same cluster, yet the theoretical guarantees (both corollaries) require this property to already hold. The authors should clarify whether Assumption (ii) is intended to hold for the training data only (where paired complete samples are available) or for the representations at inference time. If the latter, the assumption presupposes the method's success. This should be explicitly acknowledged as a limitation or the scope of the theoretical claims should be restricted accordingly.

    Authors: The referee raises a valid and important point about the scope of Assumption (ii). We agree that the assumption is circular if interpreted as holding at inference time for the recovered representations, since view consistency is precisely what the method aims to achieve. In the revised manuscript, we will clarify that Assumption (ii) is intended to hold for the training data, where paired complete samples are available and the cluster-level contrastive loss (L_cluster) can be applied to enforce cross-view assignment consistency. The theoretical claims in Corollaries 1 and 2 should be understood as characterizing the properties of the flow-matching transport when the source and target representations already belong to the same cluster, which is the regime the method is designed to achieve through its alignment objectives. We will explicitly state that the theoretical guarantees describe the idealized setting and that at inference time the method approximates this regime. We will add this as a stated limitation of the theoretical analysis. revision: yes

  3. Referee: The proof of Corollary 2 (Appendix C) invokes the ODE uniqueness theorem to claim non-intersection of trajectories from different clusters. However, ODE uniqueness prevents two trajectories with different initial conditions from crossing at the same time t, but it does not, by itself, prevent a trajectory from leaving its cluster set. The non-intersection argument conflates trajectory uniqueness (a property of the ODE solution) with cluster-set containment (a geometric property of the interpolation path). These are distinct claims and the proof does not adequately bridge them. The authors should separate these two aspects and clarify what role ODE uniqueness actually plays.

    Authors: The referee is correct that ODE uniqueness and cluster-set containment are distinct properties, and the current proof does not adequately bridge them. ODE uniqueness ensures that trajectories with different initial conditions do not cross at the same time t, but this alone does not guarantee that a trajectory remains within its cluster set. The containment property actually follows from the geometric structure of the linear interpolation path (x_t = (1-t)z1 + t*z2) combined with the convexity of the cluster set, not from ODE uniqueness per se. In the revised proof, we will restructure the argument into two separate steps: (1) a geometric step showing that under the convexity assumption (to be added per the first comment), the linear interpolation path between two points in the same convex cluster set remains within that set, and (2) a separate step invoking ODE uniqueness to show that trajectories originating from different cluster sets cannot merge, which together with the containment property ensures separation. We will clarify that ODE uniqueness plays the role of preventing cross-cluster trajectory merging, while convexity provides the containment guarantee. revision: yes

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; one minor self-citation to [50] for diffusion-in-IMVC applicability, but core flow-matching derivation is self-contained.

full rationale

The paper's central theoretical claims (Corollaries 1 and 2) are derived from standard flow-matching formulations cited to external work [22, 23] and from three explicitly stated assumptions. The proofs in Appendices B–D use standard ODE uniqueness theorems and basic properties of Gaussian processes, not self-cited results. The one self-citation to [50] (Zhang et al., a prior IMVC diffusion paper) is invoked in Appendix A.1 and B.1 to note that diffusion models have been 'empirically adopted' for IMVC, but this is contextual framing, not a load-bearing premise for the flow-matching corollaries. The paper's own theoretical contribution (Theorem 1, Corollaries 1–2) is self-contained against the stated assumptions. The reader's concern about Assumption (ii) (View Consistency) being circular with the IMVC goal is a correctness/validity concern about whether the assumption holds in practice, not a circularity in the derivation itself: the assumption is stated as a premise, not derived from the conclusion. The skeptic's concern about convexity (Corollary 2's proof implicitly requiring cluster-set convexity when only compactness is assumed) is also a correctness gap, not circularity. No step in the derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions, etc.). The free parameters are standard loss weights tuned per dataset. The axioms are domain assumptions standard in clustering literature, though Assumption (ii) creates a mild circularity with the method's goal.

free parameters (7)
  • λ1 (flow prediction loss weight) = 1.0–1000.0 (per dataset)
    Tuned per dataset (Table 3), ranging over three orders of magnitude
  • λ2 (completion consistency loss weight) = 1.0–10.0 (per dataset)
    Tuned per dataset (Table 3)
  • λ3 (cluster-level contrastive loss weight) = 1.0 (all datasets)
    Fixed across datasets (Table 3)
  • λ4 (entropy alignment loss weight) = 1.0 (all datasets)
    Fixed across datasets (Table 3)
  • τ (InfoNCE temperature) = not specified
    Used in Eq. 6 but value not reported
  • learning rate = 1e-3
    Standard choice, not a free parameter in the usual sense
  • number of flow steps = 1
    Single step used for both training and inference
assumptions (3)
  • domain assumption Cluster Separability: cluster sets S_k are compact subsets of R^d with strictly positive minimum inter-cluster distance δ > 0
    Stated in §3.1, Assumption (i). Required for Corollary 2 and the diffusion separability analysis in Appendix A.2, Eq. 20.
  • domain assumption View Consistency: different-view representations of the same sample belong to the same cluster, c(z_v) = c(z_v')
    Stated in §3.1, Assumption (ii). Load-bearing for Corollary 2's proof (Appendix C): the non-intersection of ODE trajectories depends on z1, z2 being in the same cluster set. This is essentially the goal of IMVC, creating a mild circularity.
  • domain assumption Sufficient Model Approximation: the neural network can sufficiently approximate the target function with bounded error
    Stated in §3.1, Assumption (iii). Standard for neural network-based methods but unverified in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Straight-Path Flow Matching for Incomplete Multi-View Clustering." pith.science (2026). https://pith.science/paper/SITJ3AED

@misc{pith2026260706281,
  author       = {Pith},
  title        = {Pith review of: Straight-Path Flow Matching for Incomplete Multi-View Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SITJ3AED}},
  note         = {Machine review of arXiv:2607.06281}
}
read the original abstract

Incomplete Multi-View Clustering addresses the problem of clustering multi-modal data when certain views are missing. Recent end-to-end generative approaches leverage diffusion models to recover missing views via stochastic noise-to-data trajectories. While expressive, such mechanisms are not explicitly designed for clustering, as they initialize from cluster-agnostic noise and rely on stochastic denoising dynamics. In this work, we revisit probability path design in end-to-end generative IMVC. We introduce a flow-matching framework with a linear interpolation path between paired view representations, that replaces diffusion with probability flows between observed and missing views. We provide a formal analysis showing that deterministic ODE flows are inherently better aligned with clustering objectives than diffusion-based stochastic trajectories, especially in terms of transport mechanisms that respect class-conditional data distributions and maintain cluster consistency in finite-step regimes. Building upon this insight, we develop an end-to-end IMVC architecture that integrates straight-path flow-matching view completion with cluster-level and entropy-based alignment to enforce cross-view clustering consistency. Extensive experiments on standard IMVC benchmarks demonstrate that the proposed framework establishes new state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2607.06281 by the authors.

Figure 1
Figure 1. Comparison between Flow Matching and Diffusion trajectories. Without loss of generality, we illustrate a two-view IMVC setting, where z1 denotes an observed view and z2 denotes a missing view. M denotes the region of the repre￾sentation space that contains the cluster data, while z0 represents Gaussian noise. Diffusion trajectory deviates from M due to injected noise, resulting in a long and winding path when recove… view at source ↗
Figure 2
Figure 2. Cluster separability under diffusion noise. (a) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed flow-matching framework for incomplete [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hyperparameter analysis of the loss coefficients on five datasets with [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Compared with the scattered and mixed raw features, the representa [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Representation visualization. (a), (b), (c), and (d) show the t-SNE visual￾izations of CUB, LandUse-21, Synthetic3d, and Fashion datasets, respectively, with a missing rate of 0.3 [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [1]

    Asuncion, A., Newman, D., et al.: Uci machine learning repository (2007)

    work page 2007

  2. [2]

    In: Icdm

    Bickel, S., Scheffer, T.: Multi-view clustering. In: Icdm. vol. 4, pp. 19–26 (2004)

    work page 2004

  3. [3]

    IEEE Trans

    Cai, H., Huang, W., Yang, S., Ding, S., Zhang, Y., Hu, B., Zhang, F., Cheung, Y.M.: Realize generative yet complete latent representation for incomplete multi- view learning. IEEE Trans. Pattern Anal. Mach. Intell.46(5), 3637–3652 (2023)

    work page 2023

  4. [4]

    In: AAAI

    Chao, G., Jiang, Y., Chu, D.: Incomplete contrastive multi-view clustering with high-confidence guiding. In: AAAI. vol. 38, pp. 11221–11229 (2024)

    work page 2024

  5. [5]

    Flow Matching in Latent Space

    Dao, Q., Phung, H., Nguyen, B., Tran, A.: Flow matching in latent space. arXiv preprint arXiv:2307.08698 (2023)

    work page Pith review arXiv 2023

  6. [6]

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Adv. Neural Inform. Process. Syst.34, 8780–8794 (2021)

    work page 2021

  7. [7]

    In: AAAI

    Feng, W., Sheng, G., Wang, Q., Gao, Q., Tao, Z., Dong, B.: Partial multi-view clustering via self-supervised network. In: AAAI. vol. 38, pp. 11988–11995 (2024)

    work page 2024

  8. [8]

    In: ACM Int

    Gao, H., Peng, Y., Jian, S.: Incomplete multi-view clustering. In: ACM Int. Conf. Multimedia. pp. 245–255. Springer (2016)

    work page 2016

Show all 51 references
  1. [9]

    IEEE transactions on knowledge and data engineering35(4), 3313–3332 (2021)

    Gui, J., Sun, Z., Wen, Y., Tao, D., Ye, J.: A review on generative adversarial networks: Algorithms, theory, and applications. IEEE transactions on knowledge and data engineering35(4), 3313–3332 (2021)

    2021

  2. [10]

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Adv. Neural Inform. Process. Syst.33, 6840–6851 (2020)

    2020

  3. [11]

    arXiv preprint arXiv:2410.19310 (2024)

    Huang, Z., Geng, Z., Luo, W., Qi, G.j.: Flow generator matching. arXiv preprint arXiv:2410.19310 (2024)

    2024 arXiv

  4. [12]

    In: IEEE Conf

    Jin, J., Wang, S., Dong, Z., Liu, X., Zhu, E.: Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11600–11609 (2023)

    2023

  5. [13]

    arXiv preprint arXiv:1312.6114 (2013)

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

    2013 arXiv

  6. [14]

    Klein, L., Krämer, A., Noé, F.: Equivariant flow matching. Adv. Neural Inform. Process. Syst.36, 59886–59910 (2023)

    2023

  7. [15]

    Kumar, A., Rai, P., Daume, H.: Co-regularized multi-view spectral clustering. Adv. Neural Inform. Process. Syst.24(2011)

    2011

  8. [16]

    1 kontext: Flow match- ing for in-context image generation and editing in latent space

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dock- horn, T., English, J., English, Z., Esser, P., et al.: Flux. 1 kontext: Flow match- ing for in-context image generation and editing in latent space. arXiv preprint arXiv:2506.15742 (2025)

    2025 arXiv

  9. [17]

    arXiv preprint arXiv:2301.11045 (2023)

    Li, H., Li, Y., Yang, M., Hu, P., Peng, D., Peng, X.: Incomplete multi-view clus- tering via prototype-based imputation. arXiv preprint arXiv:2301.11045 (2023)

    2023 arXiv

  10. [18]

    IEEE Trans

    Li, H., Lin, Y., Hu, P., Yang, M., Peng, X.: Community-aware multi-view repre- sentation learning with incomplete information. IEEE Trans. Pattern Anal. Mach. Intell. (2025)

    2025

  11. [19]

    Information Sciences647, 119418 (2023)

    Lin, R., Du, S., Wang, S., Guo, W.: Consistent graph embedding network with optimal transport for incomplete multi-view clustering. Information Sciences647, 119418 (2023)

    2023

  12. [20]

    IEEE Trans

    Lin, Y., Gou, Y., Liu, X., Bai, J., Lv, J., Peng, X.: Dual contrastive prediction for incomplete multi-view representation learning. IEEE Trans. Pattern Anal. Mach. Intell.45(4), 4447–4461 (2022)

    2022

  13. [21]

    In: IEEE Conf

    Lin, Y., Gou, Y., Liu, Z., Li, B., Lv, J., Peng, X.: Completer: Incomplete multi- view clustering via contrastive prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11174–11183 (2021) 16 Y. Yuan & J. Wang et al

    2021

  14. [22]

    arXiv preprint arXiv:2210.02747 (2022)

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. arXiv preprint arXiv:2210.02747 (2022)

    2022 arXiv

  15. [23]

    arXiv preprint arXiv:2209.03003 (2022)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003 (2022)

    2022 arXiv

  16. [24]

    IEEE Trans

    Liu, X., Li, M., Tang, C., Xia, J., Xiong, J., Liu, L., Kloft, M., Zhu, E.: Efficient and effective regularized incomplete multi-view clustering. IEEE Trans. Pattern Anal. Mach. Intell.43(8), 2634–2646 (2020)

    2020

  17. [25]

    In: AAAI

    Liu, X., Zhu, X., Li, M., Tang, C., Zhu, E., Yin, J., Gao, W.: Efficient and effective incomplete multi-view clustering. In: AAAI. vol. 33, pp. 4392–4399 (2019)

    2019

  18. [26]

    arXiv preprint arXiv:1807.03748 (2018)

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  19. [27]

    arXiv preprint arXiv:2307.01952 (2023)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    2023 arXiv

  20. [28]

    In: IEEE Conf

    Schusterbauer, J., Gui, M., Fundel, F., Ommer, B.: Diff2flow: Training flow match- ing models via diffusion model alignment. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 28347–28357 (2025)

    2025

  21. [29]

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: Int. Conf. Learn. Represent. (2021)

    2021

  22. [30]

    arXiv preprint arXiv:2011.13456 (2020)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

    2011 arXiv

  23. [31]

    Tang, H., Liu, Y.: Deep safe incomplete multi-view clustering: Theorem and algo- rithm. In: Int. Conf. Mach. Learn. pp. 21090–21110. PMLR (2022)

    2022

  24. [32]

    Wah,C.,Branson,S.,Welinder,P.,Perona,P.,Belongie,S.,etal.:Thecaltech-ucsd birds-200-2011 dataset. Tech. rep., Technical Report CNS-TR-2011-001, California Institute of Technology (2011)

    2011

  25. [33]

    IET Computer Vision17(2), 189–202 (2023)

    Wang, J., Xu, Z., Yang, X., Guo, D., Liu, L.: Self-supervised image clustering from multiple incomplete views via constrastive complementary generation. IET Computer Vision17(2), 189–202 (2023)

    2023

  26. [34]

    arXiv preprint arXiv:2003.13088 (2020)

    Wang, Q., Ding, Z., Tao, Z., Gao, Q., Fu, Y.: Generative partial multi-view clus- tering. arXiv preprint arXiv:2003.13088 (2020)

    2003 arXiv

  27. [35]

    IEEE Trans

    Wang, Y., Chang, D., Fu, Z., Wen, J., Zhao, Y.: Incomplete multiview clustering via cross-view relation transfer. IEEE Trans. Circuit Syst. Video Technol.33(1), 367–378 (2022)

    2022

  28. [36]

    In: 2020 IEEE International Conference on Data Mining (ICDM)

    Wei, S., Wang, J., Yu, G., Domeniconi, C., Zhang, X.: Deep incomplete multi- view multiple clusterings. In: 2020 IEEE International Conference on Data Mining (ICDM). pp. 651–660. IEEE (2020)

    2020

  29. [37]

    Wen, J., Deng, S., Wong, W., Chao, G., Huang, C., Fei, L., Xu, Y.: Diffusion-based missing-view generation with the application on incomplete multi-view clustering. In: Int. Conf. Mach. Learn. (2024)

    2024

  30. [38]

    In: Proceedings of the 30th ACM international con- ference on information & knowledge management

    Wen, J., Wu, Z., Zhang, Z., Fei, L., Zhang, B., Xu, Y.: Structural deep incomplete multi-view clustering network. In: Proceedings of the 30th ACM international con- ference on information & knowledge management. pp. 3538–3542 (2021)

    2021

  31. [39]

    IEEE Trans

    Wen, J., Xu, G., Tang, Z., Wang, W., Fei, L., Xu, Y.: Graph regularized and feature aware matrix factorization for robust incomplete multi-view clustering. IEEE Trans. Circuit Syst. Video Technol.34(5), 3728–3741 (2023)

    2023

  32. [40]

    IEEE transactions on cybernetics50(4), 1418–1429 (2018)

    Wen, J., Xu, Y., Liu, H.: Incomplete multiview spectral clustering with adaptive graph learning. IEEE transactions on cybernetics50(4), 1418–1429 (2018)

    2018

  33. [41]

    arXiv preprint arXiv:1708.07747 (2017) Straight-Path Flow Matching for Incomplete Multi-View Clustering 17

    Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for bench- marking machine learning algorithms. arXiv preprint arXiv:1708.07747 (2017) Straight-Path Flow Matching for Incomplete Multi-View Clustering 17

    2017 arXiv

  34. [42]

    Xu, G., Wen, J., Liu, C., Hu, B., Liu, Y., Fei, L., Wang, W.: Deep variational incompletemulti-viewclustering:Exploringsharedclusteringstructures.In:AAAI. vol. 38, pp. 16147–16155 (2024)

    2024

  35. [43]

    IEEE Trans

    Xu, J., Li, C., Peng, L., Ren, Y., Shi, X., Shen, H.T., Zhu, X.: Adaptive feature projection with distribution alignment for deep incomplete multi-view clustering. IEEE Trans. Image Process.32, 1354–1366 (2023)

    2023

  36. [44]

    In: AAAI

    Xu, J., Li, C., Ren, Y., Peng, L., Mo, Y., Shi, X., Zhu, X.: Deep incomplete multi-view clustering via mining cluster complementarity. In: AAAI. vol. 36, pp. 8761–8769 (2022)

    2022

  37. [45]

    In: IEEE Conf

    Yan,W.,Zhang,Y.,Lv,C.,Tang,C.,Yue,G.,Liao,L.,Lin,W.:Gcfagg:Globaland cross-view feature aggregation for multi-view clustering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19863–19872 (2023)

    2023

  38. [46]

    In: Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems

    Yang, Y., Newsam, S.: Bag-of-visual-words and spatial extensions for land-use classification. In: Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems. pp. 270–279 (2010)

    2010

  39. [47]

    IEEE Transactions on Knowledge and Data Engineering35(3), 2671–2682 (2021)

    Yin, J., Sun, S.: Incomplete multi-view clustering with reconstructed views. IEEE Transactions on Knowledge and Data Engineering35(3), 2671–2682 (2021)

    2021

  40. [48]

    Zhang, C., Han, Z., Fu, H., Zhou, J.T., Hu, Q., et al.: Cpm-nets: Cross partial multi-view networks. Adv. Neural Inform. Process. Syst.32(2019)

    2019

  41. [49]

    In: ACM Int

    Zhang, C., Wei, J., Wang, B., Li, Z., Chen, C., Li, H.: Robust spectral embedding completion based incomplete multi-view clustering. In: ACM Int. Conf. Multime- dia. pp. 300–308 (2023)

    2023

  42. [50]

    In: AAAI

    Zhang, Y., Lin, Y., Yan, W., Yao, L., Wan, X., Li, G., Zhang, C., Ke, G., Xu, J.: Incomplete multi-view clustering via diffusion contrastive generation. In: AAAI. vol. 39, pp. 22650–22658 (2025)

    2025

  43. [51]

    In: AAAI

    Zhao, H., Ding, Z., Fu, Y.: Multi-view clustering via deep matrix factorization. In: AAAI. vol. 31 (2017) 18 Y. Yuan & J. Wang et al. Appendix A Limitations of Conditional Diffusion Models on IMVC Here, we establish a formal foundation for the comparative theoretical analysis ...

    2017

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.