acceptodds
Under review as a conference paper at ICLR 2027

One-Step Consistency-aware Graph Dictionary Learning for Incomplete Multi-View Clustering

Abstract

Missing samples may degrade the clustering performance of incomplete multiview clustering (IMVC), where one-step graph-based methods show great potential. However, existing graph learning methods for IMVC typically fail to exploit consistent and complementary information simultaneously at the graph embedding level. Moreover, they often impose low-rank constraints on high-dimensional affinity tensors with high computational complexity. To this end, we propose One-Step Consistency-aware Graph Dictionary Learning (OCGDL), which employs a partially shared graph dictionary to capture both consistency and complementarity simultaneously from the similarity graph obtained from the incomplete multiview data. We first decompose each partially observed similarity graph into shared and view-specific atoms under self-orthogonality and orthogonal-complement constraints, while a missing-indicator mask confines reconstruction to observed samples. Subsequently, we relax the strict constraint on decomposed graph features by considering graph features as dictionary atoms for affinity reconstruction, and row-wise coefficients encourage differentiated atom selection across views. Furthermore, we employ the tensor nuclear norm on the concatenated multiview full graph dictionary tensor of size , which can effectively reduce computational complexity and also encourage high-order cross-view correlations at the embedding level. Discrete cluster indicators are obtained directly in the same objective via spectral rotation with adaptive view weights, forming one-step clustering. Extensive experiments on six benchmark multiview datasets validate the effectiveness of OCGDL under various view missing ratios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
One-step Incomplete Multi-View Clustering, Graph Learning, Consistency, Group Sparsity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_SZmdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes One-Step Consistency-aware Graph Dictionary Learning (OCGDL) for incomplete multi-view clustering. It reconstructs partially observed affinity graphs using shared and view-specific orthogonal dictionaries, imposes row-wise ℓ2,1\ell_{2,1} regularization on reconstruction coefficients, and applies a tensor nuclear norm to the stacked dictionaries. Spectral rotation with adaptive view weights connects the dictionaries to discrete clustering assignments. The objective is optimized using an ADMM-style alternating procedure. Experiments cover six benchmark datasets, multiple missing rates, ablations, parameter sensitivity, and runtime comparisons.

Critical strengths

The paper reports competitive clustering results, including clear improvements in several settings on Caltech101-20 and ALOI, with repeated-run statistics. This provides some empirical evidence for the usefulness of the proposed combination.

Critical weaknesses

1. The model lacks a convincing justification, and its claimed spectral foundation does not support the final formulation.

The paper does not identify a sufficiently precise limitation of existing models that necessitates this particular objective. Shared-specific decomposition, group-sparse dictionary learning, tensor regularization, and spectral rotation are established ingredients. The main contribution appears to be their integration, while the advantage over a simpler shared basis with view-dependent spectra remains insufficiently established. The shared-dimension sensitivity study and V5 provide partial evidence, but do not replace a matched comparison that isolates the benefit of the proposed decomposition and relaxation.

The theoretical motivation for Eq. (2) is also problematic. The Davis–Kahan bound in Appendix A.2, Eq. (23), controls distances between spectral subspaces; it does not imply the exact shared block asserted in Eq. (24). Two one-dimensional eigenspaces can be arbitrarily close while having no common nonzero vector. Thus, the existence of an exactly shared nonzero HcH_c requires additional assumptions. Subspace closeness alone also does not establish the particular block-diagonal spectral representation in Eq. (2). These are modeling assumptions requiring justification, rather than consequences of the stated perturbation bound.

The relaxation from Eq. (2) to Eq. (3) further disconnects the final model from this motivation. Replacing ΛH⊤\Lambda H^\top with free coefficients preserves the column orthogonality of Hv=[Hc,Hsv]H^v=[H_c,H_s^v], but does not generally preserve symmetry of the reconstructed affinity or ensure that HvH^v contains its eigenvectors. The claim that the orthogonal-complement constraint retains a genuine eigendecomposition is therefore unsupported. For the unmasked, unregularized reconstruction term with a fixed orthonormal HH, the least-squares solution is A=H⊤WA=H^\top W, giving a standard projection interpretation. The paper should explain what additional graph-spectral property survives the relaxation and why the rows of HH, rather than the coefficient columns, are the appropriate clustering features.

The remaining regularizers are insufficiently motivated. The row-wise ℓ2,1\ell_{2,1} penalty promotes atom selection within each view, but does not itself explicitly encourage different views to select different atoms. The tensor nuclear norm adds another form of cross-view coupling on top of the explicitly shared dictionary and common clustering indicator, without clearly identifying its distinct role. V4 relocates the tensor penalty to affinity representations rather than removing it, so it does not establish that this additional coupling is necessary. Overall, the objective appears to accumulate familiar terms without sufficiently explaining why they form a coherent and necessary model.

The optimization does not resolve these concerns. In particular, Eq. (8) uses a row-wise argmax, whereas the objective contains Y~=Y(Y⊤Y)−1/2\widetilde Y=Y(Y^\top Y)^{-1/2}, whose normalization depends on cluster sizes. The claimed update does not follow directly from this coupled objective and needs a derivation or a descent guarantee. Standard alternating updates and empirical residual plots do not provide a substantive new theoretical contribution or establish convergence to a stationary solution of the stated problem.

2. The efficiency claims omit unresolved graph-construction costs and are not supported by sufficiently controlled comparisons.

Appendix A.3 analyzes optimization given a preconstructed sparse graph with a constant average number of nonzero entries per row. Its simplification to O(tnlog⁡n)O(tn\log n) additionally requires treating the relevant dimensional parameters as constants; k≪nk\ll n and V≪nV\ll n alone do not justify that simplification. More importantly, the analysis does not account for obtaining the input graph from the original data. If all pairwise similarities are computed before sparsification, graph construction already incurs quadratic work in the sample size. Subsequent sparsification does not remove that cost. Approximate nearest-neighbor search may reduce preprocessing, but changes the graph, and the paper would need to verify the reported clustering performance using that actual construction. The graph-building and sparsification procedures are not specified sufficiently to establish an end-to-end near-linear claim.

The storage implications also remain unresolved. An explicitly stored dense n×nn\times n graph requires quadratic space per view, while a sparse representation can reduce this cost only under an explicit sparsity and storage assumption. Reporting a smaller k×V×nk\times V\times n regularization tensor does not establish the total memory complexity of the graph, coefficients, auxiliary variables, and intermediate computations. A complete space-complexity analysis and peak-memory measurements are missing.

Graph sparsity is also an input assumption from which compatible competing methods can benefit. A fair efficiency claim therefore requires consistent accounting of sparsity, preprocessing, and implementation across methods. It cannot be established merely by using sparse-graph bounds for the proposed algorithm and dense-graph bounds for competitors. Compact embedding-tensor regularization can reduce the cost of that particular module, as the comparison with V4 suggests, but this saving must be distinguished from savings attributable to graph sparsification.

The comparison with anchor-based methods needs similar care. Sample-level graphs and anchor-based representations differ in both retained relationships and construction costs. Higher accuracy than an anchor-based method, combined with lower runtime than a dense-graph method, does not by itself establish a superior accuracy–efficiency trade-off. The paper should use common graph inputs where the methods support them and report end-to-end accuracy, runtime, and peak memory over graph-density or anchor-budget settings. Different representations are legitimate algorithmic choices, but their costs and approximation effects need to be exposed rather than left as confounding factors.

Section 4.5 and Figure 6 do report runtime; the concern is what those timings include and whether the comparisons support the claimed advantage. Five benchmarks have at most 2,386 samples, the largest has 10,800, and missingness is generated by random sample removal. Without a controlled scaling study, explicit graph-construction and parameter-selection protocols, and consistent end-to-end timing, these experiments provide limited support for broad scalability or robustness claims.

Questions for authors
  1. What assumptions justify the exact shared block in Eq. (24) and the decomposition in Eq. (2)? After the coefficient relaxation, what spectral property is retained, and what evidence establishes an advantage over a shared basis with view-dependent spectra?

  2. What happens when the tensor nuclear norm is removed while retaining the shared dictionary and spectral-rotation term? Does it provide an independent benefit beyond these existing forms of cross-view coupling?

  3. What exact graph-construction and sparsification procedure is used in the reported experiments? Does it compute all pairwise affinities, and do the reported accuracy and runtime use the same exact or approximate graph-building pipeline?

  4. Does Figure 6 include preprocessing for every method? Are sparse operations used for compatible baselines? Please clarify the comparison with anchor-based methods through end-to-end time, peak memory, parameter selection, and the accuracy–efficiency trade-off across representation budgets.

  5. How is Eq. (8) derived while accounting for the cluster-size dependence of (Y⊤Y)−1/2(Y^\top Y)^{-1/2}? Is the update guaranteed to decrease the stated objective, and how are empty clusters handled?

6 comments
AuthorsOpenReview ↗

W1 Effect of row-wise penalty

We thank the reviewer for the careful and detailed review.

W1: row-wise ℓ2,1\ell_{2,1} penalty promotes atom selection within each view, but does not itself explicitly encourage different views to select different atoms.

AW1: We thank the reviewer for the valuable comments. Although the row-wise penalty does not directly promote diversity across views, sparsity can remove redundant and non-essential coefficients during reconstruction and then induce view-dependent atom energies and selections in both the shared and the specific blocks, promoting cross-view different dictionaries in Hsv\mathbf{H}_s^v and different atom selection in Hc\mathbf{H}_c.

To help understand the effect of the group sparsity, we have added an intuitive comparison of atom activation energy (row ℓ2\ell_2 norm) and coefficient heatmaps for Caltech101-20 and ALOI in Figures 11--13 of the Appendix 4. Compared with V2 (no sparsity) which retains redundant atoms and V3 (ℓ1\ell_1) which produces element-wise sparsity without clear row-wise selection patterns, the row-wise ℓ2,1\ell_{2,1} group sparsity can encourage more differences between views in both shared and view-specific coefficients. Together with the ablation results in Table 2, these results support the effectiveness of group sparsity regularization in encouraging diverse graph dictionaries, reducing redundancy, and improving clustering performance. Please check the complete results and discussion in the revised manuscript.

AuthorsOpenReview ↗

Q1: Assumption and corollary justification

Q1: What justifies the exact shared block (Eq. 24) and decomposition (Eq. 2)? After coefficient relaxation, what spectral property remains, and what evidence supports an advantage over a shared basis with view-dependent spectra?

A1: We thank the reviewer. The revised corollary links Eqs. (24) and (2): under small perturbations, each view's leading kck_c eigenvectors Hcv\mathbf{H}_{c}^{v} can approach Hc\mathbf{H}_{c} after orthogonal alignment (Davis--Kahan).

Corollary: Shared and view-specific spectral factors

Let W⋆∈Rn×n\mathbf{W}^{\star}\in\mathbb{R}^{n\times n} be a symmetric latent consensus affinity, with ordered eigenvalues λ1⋆≥⋯≥λn⋆\lambda_{1}^{\star}\ge\cdots\ge\lambda_{n}^{\star} and gap γ⋆=λkc⋆−λkc+1⋆>0\gamma^{\star}=\lambda_{k_c}^{\star}-\lambda_{k_c+1}^{\star}>0. Let Hc∈Rn×kc\mathbf{H}_{c}\in\mathbb{R}^{n\times k_c} collect the corresponding orthonormal eigenvectors and set Λc⋆=diag(λ1⋆,…,λkc⋆)\boldsymbol{\Lambda}_{c}^{\star}=\mathrm{diag}(\lambda_{1}^{\star},\ldots,\lambda_{k_c}^{\star}), so W⋆Hc=HcΛc⋆\mathbf{W}^{\star}\mathbf{H}_{c}=\mathbf{H}_{c}\boldsymbol{\Lambda}_{c}^{\star}. For each view, let Wv=W⋆+Nv\mathbf{W}^{v}=\mathbf{W}^{\star}+\mathbf{N}^{v} be symmetric, and let Hcv∈Rn×kc\mathbf{H}_{c}^{v}\in\mathbb{R}^{n\times k_c} collect the leading orthonormal eigenvectors of Wv\mathbf{W}^{v}, with Λcv=diag(λ1v,…,λkcv)\boldsymbol{\Lambda}_{c}^{v}=\mathrm{diag}(\lambda_{1}^{v},\ldots,\lambda_{k_c}^{v}). These columns are the view-wise counterpart of Hc\mathbf{H}_{c}:

WvHcv=HcvΛcv.\mathbf{W}^{v}\mathbf{H}_{c}^{v}=\mathbf{H}_{c}^{v}\boldsymbol{\Lambda}_{c}^{v}.

Davis--Kahan gives

∥sin⁡Θ(Hc,Hcv)∥F≤2min⁡{kc,∥Nv∥2,∥Nv∥F}γ⋆.\bigl\lVert\sin\boldsymbol{\Theta}(\mathbf{H}_{c},\mathbf{H}_{c}^{v})\bigr\rVert_{F} \le \frac{2\min\bigl\lbrace\sqrt{k_c}\\,\lVert\mathbf{N}^{v}\rVert_{2},\lVert\mathbf{N}^{v}\rVert_{F}\bigr\rbrace}{\gamma^{\star}}.

Thus, for some orthogonal Rv∈Rkc×kc\mathbf{R}^{v}\in\mathbb{R}^{k_c\times k_c},

∥HcvRv−Hc∥F2=2∑i=1kc(1−cos⁡θi)≤2∥sin⁡Θ(Hc,Hcv)∥F2,\bigl\lVert\mathbf{H}_{c}^{v}\mathbf{R}^{v}-\mathbf{H}_{c}\bigr\rVert_{F}^{2} = 2\sum_{i=1}^{k_c}(1-\cos\theta_i) \le 2\bigl\lVert\sin\boldsymbol{\Theta}(\mathbf{H}_{c},\mathbf{H}_{c}^{v})\bigr\rVert_{F}^{2},

where θi∈[0,π/2]\theta_i\in[0,\pi/2] are the principal angles between the two subspaces.

Also,

∥Hcv(Hcv)⊤−HcHc⊤∥F=2,∥sin⁡Θ(Hc,Hcv)∥F.\bigl\lVert\mathbf{H}_{c}^{v}(\mathbf{H}_{c}^{v})^{\top}-\mathbf{H}_{c}\mathbf{H}_{c}^{\top}\bigr\rVert_{F} = \sqrt{2}\\,\bigl\lVert\sin\boldsymbol{\Theta}(\mathbf{H}_{c},\mathbf{H}_{c}^{v})\bigr\rVert_{F}.

Let k=kc+ksk=k_c+k_s and Hv=[Hcv,Hsv]\mathbf{H}^{v}=[\mathbf{H}_{c}^{v},\mathbf{H}_{s}^{v}] be the leading kk orthonormal eigenvectors of Wv\mathbf{W}^{v} (descending eigenvalues), with Λv=diag(Λcv,Λsv)\boldsymbol{\Lambda}^{v}=\mathrm{diag}(\boldsymbol{\Lambda}_{c}^{v},\boldsymbol{\Lambda}_{s}^{v}).

The exact partial eigen-expansion is

HvΛv(Hv)⊤=HcvΛcv(Hcv)⊤+HsvΛsv(Hsv)⊤,\mathbf{H}^{v}\boldsymbol{\Lambda}^{v}(\mathbf{H}^{v})^{\top} = \mathbf{H}_{c}^{v}\boldsymbol{\Lambda}_{c}^{v}(\mathbf{H}_{c}^{v})^{\top} + \mathbf{H}_{s}^{v}\boldsymbol{\Lambda}_{s}^{v}(\mathbf{H}_{s}^{v})^{\top},

and Wv=HvΛv(Hv)⊤+W⊥v\mathbf{W}^{v}=\mathbf{H}^{v}\boldsymbol{\Lambda}^{v}(\mathbf{H}^{v})^{\top}+\mathbf{W}_{\perp}^{v}. If λkcv=λkc+1v\lambda_{k_c}^{v}=\lambda_{k_c+1}^{v}, the ordered split is not unique and the two blocks can mix inside the tied subspace. Set Λ~cv=(Rv)⊤ΛcvRv\tilde{\boldsymbol{\Lambda}}_{c}^{v}=(\mathbf{R}^{v})^{\top}\boldsymbol{\Lambda}_{c}^{v}\mathbf{R}^{v}.

Substituting the aligned basis gives

HvΛv(Hv)⊤=HcΛ~cvHc⊤+HsvΛsv(Hsv)⊤+Ev,\mathbf{H}^{v}\boldsymbol{\Lambda}^{v}(\mathbf{H}^{v})^{\top} = \mathbf{H}_{c}\tilde{\boldsymbol{\Lambda}}_{c}^{v}\mathbf{H}_{c}^{\top} + \mathbf{H}_{s}^{v}\boldsymbol{\Lambda}_{s}^{v}(\mathbf{H}_{s}^{v})^{\top} + \mathbf{E}^{v},

where ∥Ev∥F≤2∥Λcv∥2,∥HcvRv−Hc∥F\lVert\mathbf{E}^{v}\rVert_{F}\le 2\lVert\boldsymbol{\Lambda}_{c}^{v}\rVert_{2}\\,\lVert\mathbf{H}_{c}^{v}\mathbf{R}^{v}-\mathbf{H}_{c}\rVert_{F}.

With a positive eigengap and small perturbations, Davis--Kahan bounds subspace misalignment. This motivates shared Hc\mathbf{H}_c and view-specific Hsv\mathbf{H}_s^v in Eq. (2).

After coefficient relaxation (Eq. (3)), Hv⊤Hv=I\mathbf{H}^{v\top}\mathbf{H}^v=\mathbf{I} stays symmetric and positive semidefinite for any coefficient, while Av\mathbf{A}^v serves as reconstruction coefficients of the affinity columns for flexible graph modeling and do not force to be orthogonal, so Hv\mathbf{H}^v are the appropriate clustering features.

In the revised manuscript, we include the Learned per-view graphs on Caltech101-20 and ALOI at a 50% missing rate in Figures 14 and 15 respectively to show that OCGDL and learn effective graphs after the dictionary relaxation, which approximate class-aligned blocks on Caltech101-20 and within-class diagonals on ALOI, with varying contrast across views. At 50% missingness, OCGDL averages 91.69% ACC across six datasets versus 86.25% for V5, supporting coefficient relaxation.

AuthorsOpenReview ↗

Q2: w/o TNN ablation & Q3: graph-construction procedure & Q4: preprocessing and end-to-end time and peak memory

Q2: What happens when the tensor nuclear norm is removed while retaining the shared dictionary and spectral-rotation term? Does it provide an independent benefit beyond these existing forms of cross-view coupling?

A2: We thank the reviewer for the valuable suggestion. We have added a new ablation variant V6 (w/o TNN) in the revised Table 2, removing only the TNN regularization term while keeping all other components and experimental settings unchanged.

TNN captures high-order cross-view correlations on the compact k×V×nk\times V\times n graph dictionary tensor. At a 50 missing rate, removing TNN reduces the average ACC across six datasets from 91.69 to 63.42, demonstrating its additional contribution beyond the shared dictionary and spectral rotation. See our response to Reviewer BkvR (Q2) for ablation results and complete results in the revised manuscript.

Q3: What exact graph-construction and sparsification procedure is used in the reported experiments? Does it compute all pairwise affinities, and do the reported accuracy and runtime use the same exact or approximate graph-building pipeline?

A3: We thank the reviewer for the valuable suggestions. OCGDL employs kNN to construct similarity graphs by computing all pairwise affinities in ts nvn_v observed samples with the computational complexity O(nv2dv+nv2log⁡nv)\mathcal{O}(n_v^2 d_v+n_v^2\log n_v) for one view. The corresponding edge weights are computed using a self-tuning Gaussian kernel, followed by symmetrization and normalization to obtain a sparse similarity graph. All reported accuracy and runtime of OCGDL use the same graph initialization by setting k=10.

The compared methods follow the graph initialization strategies of their public implementations, including kNN (LSIMVC, OAGL, HCLS_CGL, OIMC-IAGD), PCA (SCSL), Zeros (CAL). To help understand the effects of different initialization strategies, we compare three graph initialization methods, including kNN, self-representation-based L1SC, and PCA, refer to A5 of Reviewer gLYR. The results show that PCA is an possible alternative of kNN, thus all reported accuracy and runtime of different methods except for CAL use the same exact or approximate graph-building pipeline.

Q4: Does Figure 6 include preprocessing for every method? Are sparse operations used for compatible baselines? Please clarify the comparison with anchor-based methods through end-to-end time, peak memory, parameter selection, and the accuracy--efficiency trade-off across representation budgets.

A4: We thank the reviewer for the valuable comments. Figure 6 in the original manuscript reports runtime including preprocessing and optimization. Figure 6 in the original manuscript reports runtime including preprocessing and optimization. In the revised manuscript, Figure 6(a) distinguishes preprocessing and end-to-end runtime, and Figure 6(b) additionally reports peak memory usage.

The compared methods use their public implementations and their parameter search ranges are added in the revised manuscript in Appendix A.4. OCGDL attains a competitive end-to-end runtime and peak memory usage comparable to anchor-based methods. Applying the tensor nuclear norm to a compact graph dictionary keeps the computational and memory overhead low while capturing low-rank correlations across views, so OCGDL offers a favorable accuracy–efficiency trade-off.

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

The manuscript proposes a graph-based incomplete multi-view clustering framework that integrates shared and view-specific graph dictionary learning, group-sparse reconstruction, tensor nuclear norm regularization, and one-step discrete clustering. The proposed framework is evaluated on six benchmark datasets under different missing rates.

Critical strengths
  1. The proposed framework integrates graph representation learning and discrete clustering into a unified objective, avoiding a separate post-processing clustering stage.
  2. Applying tensor nuclear norm regularization to compact graph-dictionary embeddings instead of full affinity tensors can reduce computational costs.
  3. The experiments cover six benchmark datasets, multiple missing rates, and several competitive baselines. The reported results are generally promising.
Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including shared-private dictionary learning, group sparsity, tensor nuclear norm regularization, and spectral rotation. Although applying tensor regularization to shared-private graph dictionaries is interesting, similar ideas have been explored in previous studies. The authors should better explain what is new compared with closely related methods.
  2. In Eq. (8), cluster labels are assigned using a row-wise argmax operation. However, the normalized indicator matrix depends on cluster sizes through (Y^TY), so the optimization cannot generally be solved independently for each sample. It is unclear whether the proposed update actually solves the stated subproblem. The authors should clarify this point and explain how empty clusters are handled.
  3. The authors claim that row-wise (\ell_{2,1}) regularization encourages different views to select different dictionary atoms. However, this regularization is applied separately to each view and mainly encourages sparsity within each view. It does not directly promote diversity across views. This claim needs further justification.
  4. The claimed near-linear computational complexity is based on assuming that several parameters remain small constants. Also, the analysis does not clearly include the cost of constructing similarity graphs. The authors should clarify the overall computational complexity.
  5. Section 3.1 discusses several ways to construct similarity graphs, but it is unclear which method is actually used in the experiments. The authors should provide more details about the graph construction process and parameter settings, as these choices can significantly affect clustering performance.
Questions for authors
  1. What are the main methodological differences between the proposed framework and existing shared-private representation learning and embedding-level tensor regularization methods? What specific advantages does the proposed integration provide?
  2. In Eq. (8), how does the row-wise argmax operation solve the discrete clustering subproblem, given that the normalized indicator matrix depends on cluster sizes through YTYY^TY? Is this an approximate update, and how are empty clusters handled?
  3. How does the row-wise ℓ2,1\ell_{2,1} regularization explicitly encourage different views to select different dictionary atoms, given that it is applied independently to each view?
  4. Does the reported computational complexity include similarity graph construction? Under what assumptions does the claimed near-linear complexity hold?
  5. Which similarity graph construction method is used in the experiments, and how are its parameters selected? Are the same graph construction settings used for all compared methods?
4 comments
AuthorsOpenReview ↗

Q1: Methodological differences and advantages over existing methods

We thank the reviewer for the thoughtful review.

Q1: What are the main methodological differences between the proposed framework and existing shared-private representation learning and embedding-level tensor regularization methods? specific advantages does the proposed integration provide?

A1: We thank the reviewer for the valuable comments. Existing IMVC methods exploit cross-view consistency and complementarity in different ways. HCLS_CGL [Wen et al., 2023] learns a consensus graph from high-confidence local structures; CAL [Liu et al., 2025] decomposes affinity representations into consensus and diverse components and models high-order correlations through tensor regularization; OAGL [Chen et al., 2025] captures both types of information through adaptive graphs and spectral embedding outer products. ARSL-IMVC [Chen et al., 2026] learns shared and view-specific subspace representations based on imputation and considers shared representation as a global view in the low-rank tensor space. However, these methods do not explicitly distinguish shared and view-specific spectral directions at the graph embedding level, which may affect the accuracy of modeling cross-view structures in IMVC. Motivated by Davis–Kahan perturbation theory, spectral subspaces from different views may share approximately aligned directions, so OCGDL uses Hc\mathbf{H}_c to model these shared directions for cross-view consistency and Hsv\mathbf{H}_s^v to model the remaining view-specific directions to capture complementarity. The V1(w/o shared dictionary, kc=0k_c=0) ablation in Table 2 supports the role of Hc\mathbf{H}_c. The per-view visualizations of Hv(Hv)⊤\mathbf H^v(\mathbf H^v)^\top in the revised Appendix further show class-related structures and differences across views.

Since strict spectral reconstruction can limit the modeling of incomplete similarity graphs, OCGDL further introduces missing-indicator masks and free reconstruction coefficients Acv\mathbf{A}_c^v and Asv\mathbf{A}_s^v to relax this restriction through graph dictionary learning. The V5(w/o relaxation) ablation supports this design. In addition, row-wise ℓ2,1\ell_{2,1} group sparsity is applied to the reconstruction coefficients to encourage view-adaptive atom selection, and diversity in the learned graph dictionaries. The V2(w/o sparsity) and V3(ℓ1\ell_1 sparsity) ablations and the coefficient heatmaps in Figures 11–13 further support its effectiveness.

Moreover, unlike CAL,ARSL-IMVC and OAGL, which impose tensor regularization on high-dimensional relation tensors with O((V+1)n2)\mathcal{O}((V+1)n^2) or O(Vn2)\mathcal{O}(Vn^2) respectively, OCGDL applies TNN to a compact graph dictionary tensor, reducing the tensor size to O(Vkn)\mathcal{O}(Vkn) and thereby lowering the computational and storage overhead of tensor regularization. The V4 and V6 ablations further support this design. In the revised manuscript, we revise the statement and discussion in the introduction to clarify these differences.

AuthorsOpenReview ↗

Q2: Indicator Y Update & Q3: Effect of row-wise penalty & Q4: Preprocessing and end-to-end time

Q2: In Eq.~(8), how does the row-wise argmax operation solve the discrete clustering subproblem, given that the normalized indicator matrix depends on cluster sizes through (Y⊤Y)−1/2(\mathbf{Y}^{\top}\mathbf{Y})^{-1/2}? Is this an approximate update, and how are empty clusters handled?

A2: We thank the reviewer for the valuable comments and we have revised the optimization step of Y\mathbf{Y}. The indicator Y\mathbf{Y} is updated based on the approximated solution. After the argmax assignment, Y\mathbf{Y} is normalized by scaling existing column by nj−1/2n_j^{-1/2} on each occupied cluster jj, with size nj=(Y⊤Y)jj>0n_j=(\mathbf{Y}^\top\mathbf{Y})_{jj}>0. The greedy solution may contain empty clusters. An alternative choice is to use the incremental rule to ascend the normalized score one row at a time (Tang et al., 2023; Chen et al., 2025), where samples only move to the other cluster when the score strictly increases, and keep a singleton stay to avoid undefined Y\mathbf{Y}.

The tables in A5 of Reviewer SZmd report clustering performances of the two variants in our OCGDL and the view-specific graph learning OAGL. Overall, argmax and incremental are competitive. When clustering accuracy is close to 100%, incremental likely outperforms argmax (e.g., Handwritten) because every cluster is already occupied, and the correct label already leads the size-normalized score on most rows. A simultaneous argmax still rebuilds every row from a frozen (Y⊤Y)−1/2(\mathbf{Y}^\top\mathbf{Y})^{-1/2}, so several boundary samples can flip together; incremental accepts a row only when the normalized score strictly increases and therefore keeps that assignment. But when clustering with a not well-learned block-diagonal graph in Caltech101-20 and ALOI, argmax tends to perform better, because few single-row moves strictly raise the objective, so incremental leaves most samples in place, whereas argmax resets every row from λvHvRv\lambda_v\mathbf{H}^{v}\mathbf{R}^{v} in one step. For the complete results, please check the revised manuscript.

Q3: How does the row-wise (ℓ2,1\ell_{2,1}) regularization explicitly encourage different views to select different dictionary atoms, given that it is applied independently to each view?

A3: We thank the reviewer for the valuable comments. Although the row-wise penalty does not directly promote diversity across views, sparsity can remove redundant and non-essential coefficients during reconstruction and then induce view-dependent atom energies and selections in both the shared and the specific blocks, promoting cross-view different dictionaries in Hsv\mathbf{H}_s^v and different atom selection in Hs\mathbf{H}_s.

To help understand the effect of the group sparsity, we have added an intuitive comparison of atom activation energy (row ℓ2\ell_2 norm) and coefficient heatmaps for Caltech101-20 and ALOI in Figures 11--13 of the Appendix 4. Compared with V2 (no sparsity) which retains redundant atoms and V3 (ℓ1\ell_1) which produces element-wise sparsity without clear row-wise selection patterns, the row-wise ℓ2,1\ell_{2,1} group sparsity can encourage more differences between views in both shared and view-specific coefficients. Together with the ablation results in Table 2, these results support the effectiveness of group sparsity regularization in encouraging diverse graph dictionaries, reducing redundancy, and improving clustering performance. Please check the complete results and discussion in the revised manuscript.

Q4: Does the reported computational complexity include similarity graph construction? Under what assumptions does the claimed near-linear complexity hold?

A4: We thank the reviewer for the valuable comments. Figure 6 in the original manuscript reports runtime including preprocessing and optimization. In the revised manuscript, Figure 6(a) distinguishes preprocessing and end-to-end runtime, while Figure 6(b) reports peak memory usage. OCGDL shows reasonable end-to-end runtime and peak memory usage, as TNN on a compact graph dictionary helps reduce computational and memory overhead.

The near-linear complexity O(tnlog⁡n)\mathcal{O}(tn\log n) applies only to optimization with preconstructed similarity graphs, assuming VV, kk, and, dd are constants. The graph is constructed by kNN which computes all pairs of squared Euclidean distances and sorting every row in one view. The computational complexity is O(nv2dv+nv2log⁡nv)\mathcal{O}(n_v^2 d_v+n_v^2\log n_v), which depends on the number of observed samples nvn_v in each view and generally decreases as the missing rate increases.

AuthorsOpenReview ↗

Q5: Graph-construction procedure

Q5: Which similarity graph construction method is used in the experiments, and how are its parameters selected? Are the same graph construction settings used for all compared methods?

A5: We thank the reviewer for the valuable suggestions. OCGDL employs kNN to construct similarity graphs. For each view vv, we compute all pairwise squared Euclidean distances among its nvn_v observed samples and retain the k=10k=10 nearest neighbors for each sample. The corresponding edge weights are computed using a self-tuning Gaussian kernel, followed by symmetrization and normalization to obtain a sparse similarity graph.

The compared methods follow the graph initialization strategies of their public implementations, including kNN (LSIMVC, OAGL, HCLS_CGL, OIMC-IAGD), PCA (SCSL), Zeros (CAL). To help understand the effects of different initialization strategies, we compare three graph initialization methods, including kNN, self-representation-based L1SC, and PCA in the following table, PreTime refers to the similarity graph construction time costs.

\hline & & \mathrm{MSRCv1} & & \mathrm{Caltech101\text{-}20} & & \mathrm{ALOI} & \\\\ \mathrm{Rate} & \mathrm{Variant} & \mathrm{ACC} & \mathrm{PreTime} & \mathrm{ACC} & \mathrm{PreTime} & \mathrm{ACC} & \mathrm{PreTime} \\\\ \hline 0.1 & \mathrm{kNN} & \mathbf{99.67}\pm\mathbf{0.45} & \mathbf{0.026}\pm\mathbf{0.015} & \mathbf{66.30}\pm\mathbf{3.66} & \mathbf{0.527}\pm\mathbf{0.028} & \underline{85.44}\pm\underline{1.08} & \underline{6.346}\pm\underline{0.040} \\\\ & \mathrm{L1SC} & 98.67\pm1.12 & 0.082\pm0.006 & 61.75\pm2.24 & 2.596\pm0.044 & 82.30\pm1.15 & 8.623\pm0.051 \\\\ & \mathrm{PCA} & \underline{99.19}\pm\underline{0.64} & \underline{0.039}\pm\underline{0.045} & \underline{61.99}\pm\underline{2.80} & \underline{1.073}\pm\underline{0.019} & \mathbf{86.38}\pm\mathbf{1.44} & \mathbf{5.819}\pm\mathbf{0.028} \\\\ \hline 0.3 & \mathrm{kNN} & \mathbf{99.33}\pm\mathbf{0.68} & \mathbf{0.017}\pm\mathbf{0.001} & \underline{66.58}\pm\underline{1.57} & \mathbf{0.366}\pm\mathbf{0.022} & \mathbf{91.39}\pm\mathbf{1.09} & \underline{3.870}\pm\underline{0.017} \\\\ & \mathrm{L1SC} & 98.19\pm0.59 & 0.062\pm0.007 & \mathbf{66.99}\pm\mathbf{2.99} & 1.939\pm0.034 & 84.58\pm1.02 & 6.005\pm0.069 \\\\ & \mathrm{PCA} & \underline{98.29}\pm\underline{0.88} & \underline{0.018}\pm\underline{0.003} & 62.51\pm2.47 & \underline{0.820}\pm\underline{0.012} & \underline{89.84}\pm\underline{1.25} & \mathbf{3.693}\pm\mathbf{0.026} \\\\ \hline 0.5 & \mathrm{kNN} & \mathbf{98.86}\pm\mathbf{0.68} & \underline{0.013}\pm\underline{0.001} & \underline{65.61}\pm\underline{2.58} & \mathbf{0.229}\pm\mathbf{0.014} & \underline{91.11}\pm\underline{2.32} & \mathbf{1.947}\pm\mathbf{0.013} \\\\ & \mathrm{L1SC} & 94.29\pm8.46 & 0.042\pm0.005 & \mathbf{67.36}\pm\mathbf{2.66} & 1.314\pm0.025 & 85.65\pm1.53 & 3.909\pm0.038 \\\\ & \mathrm{PCA} & \underline{97.76}\pm\underline{1.17} & \mathbf{0.012}\pm\mathbf{0.001} & 59.87\pm3.57 & \underline{0.544}\pm\underline{0.009} & \mathbf{91.30}\pm\mathbf{1.01} & \underline{2.071}\pm\underline{0.019} \\\\ \hline 0.7 & \mathrm{kNN} & \mathbf{95.00}\pm\mathbf{8.40} & \underline{0.009}\pm\underline{0.000} & \underline{62.23}\pm\underline{1.75} & \mathbf{0.195}\pm\mathbf{0.006} & \underline{86.24}\pm\underline{2.34} & \mathbf{0.703}\pm\mathbf{0.009} \\\\ & \mathrm{L1SC} & \underline{91.67}\pm\underline{11.70} & 0.025\pm0.002 & \mathbf{62.60}\pm\mathbf{1.73} & 0.889\pm0.022 & 80.59\pm2.26 & 2.125\pm0.029 \\\\ & \mathrm{PCA} & 75.00\pm18.10 & \mathbf{0.007}\pm\mathbf{0.000} & 60.21\pm1.90 & \underline{0.429}\pm\underline{0.007} & \mathbf{86.50}\pm\mathbf{1.98} & \underline{0.835}\pm\underline{0.009} \\\\ \hline \end{array}$$ For complete comparison, please check the revised manuscript. kNN is the most consistent of the three. PCA is a close alternative with similar reasonable preprocessing time costs. L1SC only improves on Caltech101-20 with high missing ratios, but the preprocessing time costs are consistently higher because of the iteration steps until convergence. Therefore, we select kNN as the default graph initialization of OCGDL. Moreover, all reported accuracy and runtime of different methods except for CAL use the same exact or approximate graph-building pipeline.

I thank the authors for the clarifications and additional experimental results. While several of my technical concerns have been addressed, I still have concerns about the novelty of the proposed method, as several key components have already been explored in existing work. Therefore, I will maintain my original score.

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

The paper presents One-Step Consistency-aware Graph Dictionary Learning (OCGDL), which employs a partially shared graph dictionary to capture both consistency and complementarity simultaneously from the similarity graph obtained from the incomplete multiview data. Extensive experiments on six benchmarks under diverse missing rates demonstrate that OCGDL achieves competitive or superior clustering performance against representative IMVC baselines in different metrics.

Critical strengths

1.This paper proposes OCGDL, a framework for incomplete multi-view clustering (IMVC) that unifies partially shared graph dictionaries, masked reconstruction, tensor nuclear norm (TNN), and one-step spectral rotation fusion in a single objective. 2.Extensive experiments on six benchmarks under diverse missing rates demonstrate that OCGDL achieves competitive or superior clustering performance against representative IMVC baselines in different metrics.

Critical weaknesses

1.The paper states that existing graph learning methods for IMVC fail to exploit consistency and complementarity simultaneously at the graph embedding level. What are the essential differences between the proposed method and existing methods? 2.Why is graph dictionary learning chosen as the core tool, instead of a more mature IMVC paradigm such as subspace decomposition? Please elaborate on the unique benefits brought by dictionary learning. 3.The paper summarizes the limitations of existing IMVC methods as three issues: “inconsistency between the stages at which consistency and complementarity are modeled, error accumulation in two-step methods, and the high computational cost of low-rank modeling.” Please explain more specifically which module of OCGDL directly addresses which issue.

Questions for authors

1.The authors select several representative methods for comparison. How do the performance advantages of the proposed OCGDL method compare with the additional state-of-the-art IMVC baselines? 2.Ablation studies are only conducted under the 50% missing rate. Do the contributions of individual components change under different missing rates? 3.The reproducibility statement says the code will be provided in the supplementary materials, but the manuscript does not include a code link, a complete parameter table, preprocessing details, or run scripts.

4 comments
AuthorsOpenReview ↗

W1: Key differences & W2: advantages over subspace decomposition & W3: Correspondence between limitations and modules

W1: row-wise ℓ2,1\ell_{2,1} penalty promotes atom selection within each view, but does not itself explicitly encourage different views to select different atoms.

AW1: We thank the reviewer for the valuable comments. Although the row-wise penalty does not directly promote diversity across views, sparsity can remove redundant and non-essential coefficients during reconstruction and then induce view-dependent atom energies and selections in both the shared and the specific blocks, promoting cross-view different dictionaries in Hsv\mathbf{H}_s^v and different atom selection in Hc\mathbf{H}_c.

To help understand the effect of the group sparsity, we have added an intuitive comparison of atom activation energy (row ℓ2\ell_2 norm) and coefficient heatmaps for Caltech101-20 and ALOI in Figures 11--13 of the Appendix 4. Compared with V2 (no sparsity) which retains redundant atoms and V3 (ℓ1\ell_1) which produces element-wise sparsity without clear row-wise selection patterns, the row-wise ℓ2,1\ell_{2,1} group sparsity can encourage more differences between views in both shared and view-specific coefficients. Together with the ablation results in Table 2, these results support the effectiveness of group sparsity regularization in encouraging diverse graph dictionaries, reducing redundancy, and improving clustering performance. Please check the complete results and discussion in the revised manuscript.

W2: Why is graph dictionary learning chosen as the core tool, instead of a more mature IMVC paradigm such as subspace decomposition? Please elaborate on the unique benefits brought by dictionary learning.

AW2: We thank the reviewer for the valuable comments. V5 is an ablation variant based on shared and view-specific spectral subspace decomposition, which reconstructs the similarity graph as Wv≈HcΛ~cvHc⊤+HsvΛ~sv(Hsv)⊤.\mathbf{W}^{v}\approx \mathbf{H}_{c}\tilde{\boldsymbol{\Lambda}}_{c}^{v}\mathbf{H}_{c}^{\top} + \mathbf{H}_{s}^{v}\tilde{\boldsymbol{\Lambda}}_{s}^{v} (\mathbf{H}_{s}^{v})^{\top}. This formulation restricts the reconstruction coefficients to view-dependent spectral matrices. To better represent incomplete graph structures and learn discriminative representations for clustering, OCGDL relaxes this formulation to a graph dictionary learning problem as W^v≈HcAcv+HsvAsv,\hat{\mathbf{W}}^{v}\approx \mathbf{H}_{c}\mathbf{A}_{c}^{v} + \mathbf{H}_{s}^{v}\mathbf{A}_{s}^{v}, where orthogonality constraints preserve the shared and view-specific graph embedding structure, and row-wise ℓ2,1\ell_{2,1} group sparsity suppresses redundant graph dictionary atom activations to encourage diversity across views.

As shown in Table~2, V5 performs worse than the full OCGDL model on most datasets. At a 50% missing rate, ACC on MSRCv1 decreases from 98.86% (OCGDL) to 91.86% (V5), supporting the benefit of relaxed graph dictionary reconstruction.

W3: The paper summarizes the limitations of existing IMVC methods as three issues: inconsistency between the stages at which consistency and complementarity are modeled, error accumulation in two-step methods, and the high computational cost of low-rank modeling. Which module of OCGDL directly addresses which issue?

AW3: We thank the reviewer for the valuable suggestion. We have clarified the correspondence between the three limitations and the main components of OCGDL in the revised Introduction, as follows:

(1) Consistency and complementarity: OCGDL employs a partially shared graph dictionary to explicitly distinguish shared and view-specific spectral directions at the graph embedding level, and applies the tensor nuclear norm (TNN) on the graph embedding tensor to capture high-order cross-view correlations.

(2) Error accumulation in two-step clustering: OCGDL employs spectral rotation with adaptive view weights to directly learn the discrete cluster indicator matrix Y\mathbf{Y} from shared and view-specific graph embeddings, avoiding a separate post-hoc kk-means step.

(3) Computational cost of low-rank modeling: OCGDL applies TNN to a compact graph dictionary tensor of size k×V×nk\times V\times n, rather than a full affinity tensor of size n×V×nn\times V\times n,reducing the tensor storage complexity from O(Vn2)\mathcal{O}(Vn^2) to O(Vkn)\mathcal{O}(Vkn) and alleviating the computational and storage overhead associated with tensor regularization.

AuthorsOpenReview ↗

Q1: Comparison with additional SOTA IMVC baselines

Q1: How do the performance advantages of OCGDL compare with additional state-of-the-art IMVC baselines?

A1: We thank the reviewer for the valuable suggestion. We have added two representative IMVC baselines: HCLS_CGL [Wen et al., 2023], which learns a consensus graph from high-confidence local structures, and ARSL-IMVC [Wang et al., 2026], which models shared and private low-rank subspace structures. The following table presents their comparison with OCGDL in terms of ACC. HCLS_CGL is not available on ALOI.

\hline & \mathrm{MSRCv1} & & & \mathrm{Caltech101\text{-}20} & & & \mathrm{ALOI} & & \\\\ \hline & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} \\\\ \hline 0.1 & 68.19\pm4.81 & 94.86\pm2.64 & \mathbf{99.67}\pm\mathbf{0.45} & 53.04\pm2.45 & 47.01\pm0.98 & \mathbf{66.30}\pm\mathbf{3.66} & \text{N/A} & 83.94\pm1.91 & \mathbf{85.44}\pm\mathbf{1.08} \\\\ 0.3 & 71.52\pm5.48 & 95.76\pm2.12 & \mathbf{99.33}\pm\mathbf{0.68} & 55.66\pm2.25 & 43.15\pm1.04 & \mathbf{66.58}\pm\mathbf{1.57} & \text{N/A} & 82.30\pm1.18 & \mathbf{91.39}\pm\mathbf{1.09} \\\\ 0.5 & 67.10\pm7.58 & 91.20\pm7.18 & \mathbf{98.86}\pm\mathbf{0.68} & 54.67\pm3.04 & 43.58\pm1.07 & \mathbf{65.61}\pm\mathbf{2.58} & \text{N/A} & 83.50\pm1.27 & \mathbf{91.11}\pm\mathbf{2.32} \\\\ 0.7 & 52.00\pm4.78 & 93.70\pm2.29 & \mathbf{95.00}\pm\mathbf{8.40} & 49.69\pm5.72 & 45.81\pm1.00 & \mathbf{62.23}\pm\mathbf{1.75} & \text{N/A} & 84.25\pm2.17 & \mathbf{86.24}\pm\mathbf{2.34} \\\\ \hline \end{array}$$ $$\begin{array}{l|rrr|rrr|rrr} \hline & \mathrm{Handwritten} & & & \mathrm{ORL40} & & & \mathrm{BBCSport} & & \\\\ \hline & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} & \mathrm{HCLS_CGL} & \text{ARSL-IMVC} & \mathrm{OCGDL} \\\\ \hline 0.1 & 85.66\pm0.21 & 99.64\pm0.11 & \mathbf{99.76}\pm\mathbf{0.13} & 76.98\pm1.57 & 91.76\pm1.38 & \mathbf{99.78}\pm\mathbf{0.18} & 81.64\pm4.21 & 95.47\pm2.30 & \mathbf{99.87}\pm\mathbf{0.15} \\\\ 0.3 & 85.15\pm1.96 & 99.75\pm0.10 & \mathbf{100.00}\pm\mathbf{0.00} & 72.25\pm1.65 & 90.41\pm1.83 & \mathbf{97.42}\pm\mathbf{2.02} & 77.30\pm7.96 & 87.51\pm10.92 & \mathbf{99.15}\pm\mathbf{0.26} \\\\ 0.5 & 83.08\pm0.71 & 99.67\pm0.12 & \mathbf{100.00}\pm\mathbf{0.00} & 59.48\pm2.11 & 92.10\pm0.90 & \mathbf{96.75}\pm\mathbf{2.37} & 44.76\pm2.08 & 78.43\pm9.50 & \mathbf{97.79}\pm\mathbf{5.06} \\\\ 0.7 & 80.14\pm1.28 & \mathbf{99.33}\pm\mathbf{0.15} & 98.90\pm3.15 & -- & -- & -- & -- & -- & -- \\\\ \hline \end{array}$$ These results further support the performance advantages of OCGDL. For complete comparisons, please refer to Table~1 of the revised manuscript.
AuthorsOpenReview ↗

Q2: Ablation across different missing rates

Q2: Ablation studies are only conducted under the 5050\\% missing rate.

A2: We thank the reviewer for the valuable suggestion. Due to space limitations, the original ablation study only reported results at a 50% missing rate. We have now added experimental results for 10%, 30%, and 70% missing rates in the revised manuscript. The following table presents the ACC comparison between OCGDL and its variants.

\hline \mathrm{MSRCv1} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & \mathbf{99.67}\pm\mathbf{0.45} & 99.52\pm0.39 & 98.05\pm1.87 & 98.14\pm1.59 & 97.67\pm3.73 & 97.76\pm3.44 & 75.43\pm2.49 \\\\ 0.3 & \mathbf{99.33}\pm\mathbf{0.68} & \mathbf{99.33}\pm\mathbf{0.68} & 99.10\pm0.47 & 98.24\pm4.58 & 98.38\pm1.33 & 96.71\pm1.70 & 71.24\pm3.97 \\\\ 0.5 & \mathbf{98.86}\pm\mathbf{0.68} & 93.33\pm7.32 & 90.33\pm10.90 & 96.74\pm2.38 & 94.19\pm6.90 & 91.86\pm6.95 & 67.19\pm4.16 \\\\ 0.7 & \mathbf{95.00}\pm\mathbf{8.40} & 82.48\pm12.67 & 90.52\pm7.61 & 85.38\pm10.09 & 72.14\pm5.94 & 51.33\pm13.88 & 54.24\pm4.42 \\\\ \hline \end{array}$$ $$\begin{array}{l|rrrrrrr} \hline \mathrm{Caltech101\text{-}20} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & \mathbf{66.30}\pm\mathbf{3.66} & 62.27\pm3.54 & 61.22\pm2.75 & 62.29\pm2.17 & 56.26\pm2.35 & 65.49\pm4.00 & 43.85\pm4.58 \\\\ 0.3 & 66.58\pm1.57 & 65.20\pm2.15 & 62.97\pm2.91 & 66.07\pm3.76 & 59.64\pm5.47 & \mathbf{67.23}\pm\mathbf{3.56} & 48.87\pm2.85 \\\\ 0.5 & \mathbf{65.61}\pm\mathbf{2.58} & 56.13\pm2.68 & 62.28\pm2.46 & 62.82\pm5.50 & 59.06\pm3.49 & 62.51\pm2.94 & 39.01\pm2.30 \\\\ 0.7 & \mathbf{62.23}\pm\mathbf{1.75} & 55.37\pm3.66 & 60.35\pm4.10 & 43.07\pm2.84 & 60.60\pm3.48 & 59.90\pm3.18 & 39.84\pm3.41 \\\\ \hline \end{array}$$ $$\begin{array}{l|rrrrrrr} \hline \mathrm{ALOI} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & \mathbf{85.44}\pm\mathbf{1.08} & 84.73\pm1.20 & 83.03\pm1.71 & 82.73\pm1.69 & 84.51\pm1.19 & 84.92\pm1.73 & 77.00\pm2.16 \\\\ 0.3 & \mathbf{91.39}\pm\mathbf{1.09} & 88.19\pm1.26 & 91.13\pm1.36 & 90.08\pm1.19 & 82.45\pm1.94 & 80.03\pm1.52 & 71.98\pm1.69 \\\\ 0.5 & \mathbf{91.11}\pm\mathbf{2.32} & 89.02\pm0.78 & 89.73\pm2.00 & 88.41\pm1.87 & 80.35\pm1.25 & 87.20\pm5.67 & 69.13\pm1.65 \\\\ 0.7 & \mathbf{86.24}\pm\mathbf{2.34} & 84.36\pm2.39 & 85.54\pm2.57 & 84.21\pm2.36 & 80.90\pm1.88 & 74.34\pm3.75 & 49.54\pm1.83 \\\\ \hline \end{array}$$ $$\begin{array}{l|rrrrrrr} \hline \mathrm{Handwritten} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & 99.76\pm0.13 & 98.63\pm3.34 & 98.67\pm3.40 & 99.74\pm0.13 & \mathbf{99.98}\pm\mathbf{0.03} & 99.75\pm0.12 & 96.59\pm0.26 \\\\ 0.3 & \mathbf{100.00}\pm\mathbf{0.00} & 99.98\pm0.04 & \mathbf{100.00}\pm\mathbf{0.02} & \mathbf{100.00}\pm\mathbf{0.00} & 97.99\pm4.44 & 99.74\pm0.20 & 93.16\pm4.44 \\\\ 0.5 & \mathbf{100.00}\pm\mathbf{0.00} & 99.97\pm0.03 & 99.98\pm0.05 & 99.99\pm0.02 & 99.99\pm0.02 & 99.61\pm0.22 & 88.28\pm4.57 \\\\ 0.7 & 98.90\pm3.15 & 99.84\pm0.11 & 98.89\pm3.14 & 99.56\pm0.33 & \mathbf{99.92}\pm\mathbf{0.07} & 94.54\pm6.30 & 74.76\pm3.91 \\\\ \hline \end{array}$$ $$\begin{array}{l|rrrrrrr} \hline \mathrm{ORL40} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & \mathbf{99.78}\pm\mathbf{0.18} & 98.78\pm1.07 & 98.93\pm1.12 & 99.58\pm0.76 & 94.83\pm3.77 & 98.60\pm2.72 & 80.05\pm2.00 \\\\ 0.3 & \mathbf{97.42}\pm\mathbf{2.02} & 95.90\pm2.41 & 96.08\pm1.86 & 96.75\pm1.99 & 91.85\pm1.90 & 92.93\pm5.55 & 73.40\pm2.25 \\\\ 0.5 & \mathbf{96.75}\pm\mathbf{2.37} & 91.33\pm3.67 & 96.28\pm1.29 & 96.18\pm2.62 & 94.90\pm1.55 & 84.25\pm6.49 & 57.15\pm2.90 \\\\ \hline \end{array}$$ $$\begin{array}{l|rrrrrrr} \hline \mathrm{BBCSport} & \mathrm{Full} & \mathrm{V1} & \mathrm{V2} & \mathrm{V3} & \mathrm{V4} & \mathrm{V5} & \mathrm{V6} \\\\ \hline 0.1 & \mathbf{99.87}\pm\mathbf{0.15} & 98.64\pm4.11 & 97.70\pm5.86 & 98.05\pm4.70 & 95.06\pm7.29 & 85.90\pm8.41 & 79.54\pm8.25 \\\\ 0.3 & \mathbf{99.15}\pm\mathbf{0.26} & 97.30\pm4.00 & 98.25\pm2.84 & 99.03\pm0.41 & 98.46\pm1.17 & 96.43\pm6.54 & 82.85\pm5.55 \\\\ 0.5 & \mathbf{97.79}\pm\mathbf{5.06} & 92.35\pm10.49 & 92.26\pm12.27 & 96.91\pm6.08 & 91.38\pm10.72 & 92.06\pm10.54 & 59.76\pm6.95 \\\\ \hline \end{array}$$ These results further confirm the overall performance advantage of the full OCGDL model over its variants. Complete results are provided in Table 2 of the revised manuscript.
AuthorsOpenReview ↗

Q3: Reproducibility

Q3: The reproducibility statement says the code will be provided in the supplementary materials, but the manuscript does not include a code link, a complete parameter table, preprocessing details, or run scripts.

A3: We thank the reviewer for the valuable suggestion. Due to the double-blind review policy, we have provided the implementation code, detailed parameter settings, and run scripts in the anonymous supplementary materials, and will release the code and scripts on GitHub after the review process.

We have also clarified the preprocessing details and initialization comparisons in Appendix A.3, provided the parameter search ranges in Appendix A.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.