acceptodds
Under review as a conference paper at ICLR 2027

DiFuMVC: Distribution Fusion Multi-View Clustering

Abstract

Multi-view clustering (MVC) must decide how to represent each view, how to fuse information across views, and how to cluster. Most MVC pipelines operate on sample relations, point prototypes, or latent codes. Some additionally align or transport distributions in selected representation or fusion operations, while their final clustering remains sample-level. We introduce DiFuMVC, which adopts a distributional perspective through all three stages. Distributional representation retains components in each view and encodes each empirical distribution by a kernel mean embedding (KME). Distribution fusion uses shared reference-object identities to compare source-view points with target-view KMEs while preserving view-specific geometry. Distribution-driven clustering assembles all ordered view pairs into an point-distribution bipartite graph and applies transfer cut on its compact distribution side. We prove that the induced cross-view kernel is positive semidefinite, show that point–distribution fusion is an expected cross-view kernel affinity, and derive conditional finite-sample bounds that propagate both KME sampling error and random-reference approximation error through the fused graph to its compact spectral subspace under fixed-rule and split-reference conditions. Benchmark experiments demonstrate the efficiency and effectiveness of our proposed DiFuMVC.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multi-view clustering, data distribution, kernel mean embedding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9Qs4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes DiFuMVC, a multi-view clustering framework that integrates information across views through distribution modeling and performs efficient clustering via spectral computation on a compact graph. The framework has three stages: distributional representation describes groups discovered in each view through kernel mean embeddings; distribution fusion uses shared reference objects to compute cross-view point-to-distribution similarities while preserving view-specific geometry; and distribution-driven clustering assembles these relationships into a bipartite graph, performs spectral computation on its compact distribution side, and transfers the result to samples. Experiments on 15 benchmark datasets report competitive clustering performance and runtime.

Critical strengths
  1. The design makes use of group information while keeping spectral computation efficient. Each group representation aggregates information from all its members, allowing cross-view fusion to use distributional information that a single representative point may miss. Performing the spectral step on a small matrix also reduces its computational cost.

  2. The reported clustering performance and runtime are encouraging. The method achieves competitive clustering results with fast reported runtimes across multiple benchmarks. The ablation results favoring group representations over single-point representatives also support the use of information from group members.

Critical weaknesses
  1. The explanation that only one Food101 view contains useful information lacks direct evidence. To examine whether fusion across views is beneficial, Section 4.3 compares the full model with a variant that removes connections between views. The latter performs slightly better on some Food101 clustering metrics, which the authors explain by stating that only one view provides useful information while the other two introduce noise. However, the method is intended to improve clustering by exploiting information from multiple views. When adding cross-view fusion instead reduces performance on some metrics, the authors attribute this to noise in the other two views. Could this instead reflect the proposed fusion mechanism's failure to exploit useful information from those views, rather than those views being inherently uninformative?

  2. The presentation of the method is poor. The authors organize the method into distributional representation, distribution fusion, and distribution-driven clustering, providing an overall framework from input data to final clusters. However, the detailed description in Section 3 introduces many mathematical symbols in succession. Readers must repeatedly revisit their definitions to understand what each step takes as input and produces as output, making it difficult to connect the equations into a complete procedure. Consequently, readers struggle to grasp the main ideas the authors intend to convey.

Questions for authors
  1. What rule is used to assign the remaining samples to the retained groups? Section 3.2.2 performs this assignment before constructing the distribution representations. The rule determines which samples contribute to each representation and is therefore necessary for reproducing the method.

  2. What values of ψ and τ were used in the experiments? The main text does not provide the values of the reference-sample count ψ and grouping threshold τ, or show how the clustering results change when these parameters vary. Their impact on the reported performance is therefore difficult to assess.

1 comment

We thanks for your review and feedback.

W1: Among the 86,796 samples in the Food101, all class-text rows are identical and description text has only 397 distinct feature rows. Any deterministic classifier using only one view must predict identically for identical inputs. Its accuracy is therefore bounded by the sum of majority-class counts within feature-equivalence groups, divided by sample count. This gives ceilings of 1.09% for class text and 8.45% for description text. These are direct representation bottlenecks, independent of DiFuMVC’s fusion mechanism. The phrase “both text views are noise” nevertheless needs precision. Description-text empirical mutual information with labels is 0.363 nats, above the maximum of 0.131 across 100 label permutations, so it contains nonzero information (low information). Table 13 uses Full to compare the variants: Full NMI is 52.84, self-only and image-only are 53.22 and 53.30, and Full differs by −0.18, −0.14, and +1.67 points in NMI, ARI, and F1, while exceeding fixed Point by 8.93 NMI points. On the other three datasets, the Full reference exceeds fixed self-only by 4.13, 1.68, and 4.20 NMI points; on Food101, it exceeds the best tested LMVSC by 11.11 points.

W2: Thank you for your suggestion; we have added the description to the revised version.

Q1: Assignment uses initial KMEs computed from retained components, not the final KMEs that have yet to be estimated. The order is: retain components; average their member features; assign remaining samples by maximum inner product; recompute KMEs from all members of each expanded nonempty group; construct the fusion graph.

Q2: The appendix already contains parameter and sensitivity studies.

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

This paper presents DIFUMVC, which explicitly works with distributions throughout the representation, fusion, and clustering stages of multi-view clustering. It first uses isolation features to identify components in each view and represents the resulting empirical distributions using kernel mean embeddings. It then constructs sample-to-distribution affinities for all ordered view pairs, producing a compact bipartite graph. Finally, a transfer cut is performed on the distribution side, and the resulting spectral representation is transferred back to the samples. The paper also analyzes how KME estimation error and random-reference approximation error propagate to the fused graph and its spectral subspace, and evaluates the method on standard benchmarks and spatial transcriptomics datasets.

Critical strengths
  1. Distributions serve as the basic objects throughout all three stages of the method. The KME representation, point-to-distribution fusion, and distribution-side spectral clustering form a coherent pipeline rather than a conventional framework with a distributional regularizer added on top.
  2. The method avoids constructing a full sample-to-sample graph and performs eigendecomposition on the distribution side, which has size at most V2kV^2k. It achieves the best runtime on all 15 datasets, which represents a clear efficiency advantage over all baselines.
  3. Beyond a KME concentration bound, the paper studies random-reference approximation, affinity-ranking preservation, perturbation of the fused graph, and stability of the compact spectral subspace. It also clearly states that these results do not automatically imply stability of the final K-means labels. This careful delineation of the theoretical scope is commendable.
Critical weaknesses
  1. The theoretical analysis assumes that the component-discovery rule is independent of the samples used for KME estimation, or that an independent pilot/evaluation split is used. The actual algorithm, however, uses the same data and reference partitions for these steps. The current theory therefore does not directly cover the algorithm evaluated in the experiments.
  2. Sharing the identities of reference objects across views makes the cross-view inner products mathematically well-defined, but it does not by itself establish that isolation regions in different views have comparable semantics. A direct comparison among shared references, independent references, and randomly matched coordinates is missing, leaving the proposed cross-view fusion mechanism insufficiently validated.
Questions for authors
  1. In Stage I, how are samples outside the retained connected components, including those in smaller components, reassigned in practice? What rules are used when all similarities are zero or when there are ties? These choices directly affect the KMEs, but the current description is not detailed enough for full reproducibility.
  2. Could the authors report empirical KME separation, affinity margins, and the eigengap between the kk-th and (k+1)(k+1)-th eigenvalues on at least a subset of the datasets? These values would help assess whether the stability conditions used in the theory are meaningful under the experimental conditions.
1 comment

We thanks for your review and suggestions.

W1: Our probabilistic analysis gives stability bounds under fixed-rule and required independence conditions; its scope should be distinguished from the default data-reuse implementation. The submitted abstract and Section 3.5 already state this scope and explain why conditioning on groups learned from the same observations does not restore independence. The analysis describes how KME and reference-approximation errors propagate to the graph and compact spectral subspace. Positive semidefiniteness, the empirical KME identity, and compact spectral correspondence are deterministic results under their respective conditions. We tested an executable independent design using disjoint 50% pilot, 25% estimation, and 25% query samples. The pilot fits the map and assignment rule; independent estimation samples construct the KMEs under that fixed rule; query samples form the evaluation graph. A second variant retains the pilot rule but independently redraws the representation reference bank from the same pilot pool. Query NMI is 61.68, 88.60, 62.34, and 54.96 with the pilot map, and 61.63, 88.55, 61.77, and 55.02 with redrawn references. Both designs operate on all four datasets and yield similar mean performance.

W2: Shared identities use the same paired objects as cross-view references. We completed shared-bank, independent-bank, and independent-coordinate-permutation controls. In the permutation control, memberships are fixed and coordinates are shuffled within each partition, with the same shuffle applied to a view’s samples and KMEs. This isolates the contribution of cross-view matching.

The mean NMI differences from independent permutations are 3.86, 2.16, 0.83, and 0.20 points. The first two datasets differences are larger; Caltech7 has greater permutation variability and Food101 has a small difference. In Figure 5, the same Full reference exceeds the variant removing cross-view blocks by 4.21, 2.22, and 1.57 points on the first three datasets. Together, these results support useful cross-view correspondence from shared identities.

Q1: We retain the largest available components of the subset graph and compute initial KMEs from their members. Retained members receive group indices; all remaining samples, including those in small components and outside the subset, are assigned by maximum inner product with an initial KME. If the similarity is 0 or the values ​​are identical (rarely happens), assign randomly. Each nonempty group KME is then recomputed from all assigned members for fusion.

Q2: The minimum empirical KME distance is positive on Flower17, ALOI, and Caltech7. Positive top-two affinity margins occur in 99.9967%, 99.9993%, and 99.9883% of eligible sample-block cases, respectively, and 96.8784% on Food101. Distances use Ψ=Φ/R\Psi=\Phi/\sqrt{R}, counting each distribution node once before ordered-view replication. Margins are the largest minus second-largest raw kernel similarities, with a positivity threshold 10−1210^{-12}. Percentages are averaged over five runs and ordered-view blocks with at least two target nodes. The eigengap is computed from the actual compact normalized operator.

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

This paper proposes a distribution fusion-based multi-view clustering method (DiFuMVC), aiming to unify representation, fusion, and clustering within a distributional framework. Specifically, the proposed method first constructs kernel mean embeddings (KMEs) for each view using isolation kernel features and connected components. It then builds a fused bipartite graph based on cross-view point–distribution similarities and finally employs Transfer Cut to achieve efficient spectral clustering. Furthermore, the authors provide theoretical analyses of the positive semidefiniteness of the kernel, distribution estimation errors, and spectral subspace stability. Extensive experiments demonstrate the effectiveness and computational efficiency of the proposed method.

Critical strengths

This paper presents a multi-view clustering framework from a distributional perspective, incorporating distributional representations throughout the three stages of feature representation, cross-view fusion, and final clustering, demonstrating a certain degree of novelty.

Critical weaknesses

The distribution construction process relies on heuristic strategies, such as similarity thresholds and connected component selection, which may be sensitive to variations in data density, noise, and cluster sizes, without sufficient guarantees regarding the reliability of the resulting distributional representations.

Questions for authors
  1. Kernel mean embedding is essentially a mean representation in the feature space. Could the authors further clarify the fundamental differences between the distribution nodes constructed in this paper and traditional clustering prototypes or anchors?
  2. Could the authors provide additional comparisons with traditional efficient anchor graph-based clustering methods using the same isolation kernel features to further validate the independent contribution of distributional representations over point prototype representations?
1 comment

We thanks for your review and valuable comments.

W1: We tested the complete pipeline from perturbed inputs, rather than only perturbing already discovered memberships. On the originally class-balanced Flower17 and ALOI datasets, we add Gaussian noise with SD equal to 5%, 10%, or 20% of each view’s centered feature RMS, then rebuild shared references, features, groups, and clustering. Separately, we construct target largest-to-smallest class-size ratios of 1, 2, and 4 while preserving total sample count: 680 for Flower17 and 5,400 for ALOI. At the strongest prespecified noise level, Full retains NMI advantages of 8.41, 19.30 points over point representatives. At target class-size ratio 4, the corresponding advantages are 6.60, 16.35 points. These results support effective aggregation even when empirical groups are not true classes. The separate membership stress test is retained: at 20% sample selection for label permutation, KME exceeds the same-group point variant by 8.56–28.48 NMI points.

Q1: The distinction concerns the information represented by a node and the graph built from it. Each isolation-KME coordinate measures the distribution by group members. A query-to-node inner product equals the average kernel similarity to all members. A single-member anchor retains only one sample response, and mapping an input-space centroid through a nonlinear feature map generally differs from averaging mapped samples. The KME therefore carries group distribution information into cross-view affinities. We added fixed-parameter Point and self-only ablations. For a consistent baseline, the mean NMI of DiFuMVC differs from Point by 8.81, 19.72, 25.96, and 8.93 points. KME exceeds Point in NMI and ARI in all datasets. Table 4 also reports self-only graphs to distinguish aggregation from cross-view contributions. The arithmetic mean of the same features and members is numerically identical to that empirical KME. Our contribution is the effective combination of distribution nodes, ordered cross-view affinities through shared reference identities, and compact spectral clustering.

Q2: We completed an independent LMVSC anchor-graph control using the isolation features. The mean NMI differences from the measured best tested LMVSC candidates are 12.29, 26.90, 24.04, and 11.11 points on Flower17, ALOI, Caltech7, and Food101. Reusing isolation features, therefore, does not recover the performance of the complete distribution-fusion pipeline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.