acceptodds
Under review as a conference paper at ICLR 2027

Cross-View Granular-Ball Structure Transfer for Incomplete Multi-View Clustering

Abstract

Incomplete multi-view clustering (IMVC) groups unlabeled samples when one or more views are unavailable. Existing graph-based methods typically recover missing structure through instance-level pairwise neighborhoods, which are vulnerable to view-specific geometry and local noise. Granular balls instead aggregate locally coherent samples into mesoscopic units that preserve neighborhood organization while reducing sensitivity to isolated perturbations. Building on this insight, we propose Cross-View Granular-Ball Structure Transfer (CGBST), a region-mediated framework that transfers granular structure rather than missing features or individual relations. CGBST partitions observed embeddings into adaptive regions and constructs their local topology. Co-observed samples establish soft cross-view region correspondences, allowing source-view affiliations to infer target-region distributions for missing samples. The inferred affiliations and target topology are projected back to the sample level to recover missing relations. Experiments on four benchmarks and three missing rates show that CGBST achieves state-of-the-art performance in 42 of 48 evaluation settings, validating region-level structure transfer for incomplete multi-view clustering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Incomplete multi-view clustering, Granular ball, Cross-view Structure Transfer, Missing Relation Recovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jqwcOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CGBST, which models the local region structure through Granular Ball and establishes the corresponding relationship of regions by sharing samples across views, thereby restoring the sample relationships of missing views. The experiments demonstrate a certain improvement in clustering performance, but the paper still has significant deficiencies in terms of research motivation, method innovation, and experimental persuasiveness.

Critical strengths
  1. The overall structure of the paper is clear, and the description of the methods is relatively complete.
  2. Good clustering results were achieved on multiple datasets, and certain ablation analyses were provided.
Critical weaknesses
  1. The research significance is limited. The core motivation of the paper is to reduce the sensitivity of instance-level relationships to noise through regional-level structure transmission. However, it does not adequately explain how serious this issue is in the existing IMVC methods, nor does it provide direct evidence to prove that Granular Ball is the necessary choice to solve this problem.
  2. The method lacks innovation. The overall framework mainly combines regional division, cross-view correspondence based on shared samples, graph relationship restoration, as well as conventional contrastive learning and graph regularization. The combination of these modules is somewhat reasonable, but it lacks sufficient prominent new ideas or theoretical insights.
Questions for authors
  1. Can the author provide more direct empirical evidence to demonstrate that the existing instance-level structure transfer methods do indeed have significant flaws in noisy or cross-view overlapping sparse scenarios? It is currently unclear how important the problems solved by introducing Granular Ball are.
  2. Table 2 shows that the Instance-level Transfer outperforms the full method in some metrics on the CUB dataset. Can the author further explain the core advantages of Granular Ball compared to instance-level transfer and prove the necessity of this design by comparing it with other region aggregation or prototype methods?
  3. The missing rate defined in the paper is η/V, and the actual missing rate is relatively limited. Can the author provide experimental results with a higher actual missing rate, especially in cases where cross-view shared samples are very few, to verify the effectiveness of the method in the claimed difficult scenarios?
4 comments

R1. Empirical Evidence for Granular-Ball Transfer under Noise and Sparse Cross-View Overlap (W1 & Q1)

We appreciate the reviewer's suggestion to further evaluate CGBST under sparse cross-view overlap and noisy conditions. We have added a dedicated evaluation comparing CGBST with pure instance-level transfer under noise perturbations and partial cross-view overlap (Page 9, Figure 3; Page 18, Appendix F.2, Table 8). The experiments vary the common observation rate ρ∈{0.3,0.4,0.5}\rho\in\{0.3,0.4,0.5\} at a fixed noise level σ=0.25\sigma=0.25, with the actual pairwise co-observation rates explicitly reported.

CGBST consistently outperforms pure instance-level transfer across all four metrics on MSRCV1, achieving improvements of 7.15, 6.16, 8.88, and 5.16 percentage points in ACC, NMI, ARI, and F-score at ρ=0.3\rho=0.3, respectively. It also improves all four metrics on OutdoorScene at ρ=0.3\rho=0.3 and ρ=0.4\rho=0.4. These results demonstrate the advantages of region-mediated structure transfer under sparse and noisy cross-view observations.

Table R1. Partial cross-view overlap and noise robustness (σ=0.25\sigma=0.25; Mean, %).

Datasetρ\rhoTransfer variantACCNMIARIF-score
MSRCV10.3Instance-pure64.5256.2543.0856.48
CGBST (ours)71.6762.4151.9661.64
0.4Instance-pure74.2966.1558.3165.81
CGBST (ours)79.1069.3361.3968.41
0.5Instance-pure80.6272.0960.6670.74
CGBST (ours)85.8675.2269.8075.83
OutdoorScene0.3Instance-pure63.6346.5337.1648.51
CGBST (ours)64.5648.6238.7449.68
0.4Instance-pure59.5349.2839.1749.83
CGBST (ours)66.2851.3042.1851.33
0.5Instance-pure68.5752.6442.2753.69
CGBST (ours)68.1751.9241.5652.93

Pairwise co-observation rates for ρ=0.3/0.4/0.5\rho=0.3/0.4/0.5: MSRCV1 28.10–34.29% / 28.57–42.86% / 45.71–52.86%; OutdoorScene 28.61–32.03% / 38.06–40.36% / 49.40–51.53%.

R2. Methodological Novelty and Advantages over Instance-Level and Prototype-Based Transfer (W2 & Q2)

The central contribution of CGBST is a cross-view structural transfer mechanism that uses granular balls as intermediate units for missing-relation recovery. Existing granular-ball-based multi-view clustering methods primarily employ regional structures for representation learning or multi-granularity alignment. CGBST extends granular-ball modeling to incomplete-view structure recovery through three connected mechanisms: adaptive view-specific granular regions, co-observed-anchor-driven soft region correspondences, and target-topology-guided sample relation recovery. This establishes a complete transfer pathway from source-view regional organization to missing sample relations in the target view.

To further evaluate the contribution of granular-ball modeling, we have added two complementary controls: Instance-level comparison (Page 8, Table 2). The original Instance-level Transfer retains granular-ball construction and replaces region-level correspondence with instance-anchor transfer. We additionally introduce Instance-pure Transfer, which removes granular-ball construction and directly transfers instance-level graph relations. Overall, CGBST outperforms Instance-pure Transfer in 13 out of 16 evaluation metrics across all four datasets, demonstrating the effectiveness of granular-ball structure modeling for cross-view relation recovery.

Prototype replacement (Page 19, Appendix F.3, Table 9). We replace adaptive granular-ball construction with fixed-K K-means prototypes K=20K=20, keeping the remaining structure-transfer and training pipeline unchanged. CGBST outperforms the K-means prototype control across all four metrics at every evaluated missing rate on MSRCV1. At η=0.5\eta=0.5, the improvements reach 6.11, 5.74, 9.73, and 7.14 percentage points in ACC, NMI, ARI, and F-score, respectively.

Together, these comparisons demonstrate the contribution of adaptive granular regions and region-mediated correspondences to recovering missing neighborhood structure.

Table R2(a). Instance-level transfer ablation at η=0.3\eta=0.3 (Mean, %; extracted from Page 8, Table 2).

DatasetVariantACCNMIARIF-score
CUBFull CGBST77.3776.6465.5173.01
Instance-pure77.9475.6465.0670.87
MSRCV1Full CGBST93.4085.9384.4087.32
Instance-pure92.1184.5182.0585.85
OutdoorSceneFull CGBST74.1360.7952.4060.50
Instance-pure71.1458.4250.0658.29
NUS-WIDEFull CGBST55.0453.0834.5352.29
Instance-pure54.5753.9534.3952.32

Table R2(b). K-means prototype replacement on MSRCV1 (Mean, %; Page 19, Appendix F.3, Table 9).

η\etaVariantACCNMIARIF-score
0.1K-means prototype (K=20)92.1084.6981.8785.87
CGBST (ours)94.6888.1186.3089.51
0.3K-means prototype (K=20)90.0582.9478.6983.52
CGBST (ours)93.4085.9384.4087.32
0.5K-means prototype (K=20)85.4977.0771.1977.46
CGBST (ours)91.6082.8180.9284.60
0.7K-means prototype (K=20)82.7675.6168.5475.14
CGBST (ours)86.0578.2373.8778.88

R3. Evaluation under Higher Missing Rates and Sparse Co-Observed Samples (Q3)

We have extended the missing-rate evaluation to η=0.7\eta=0.7 (Page 18, Appendix F.1, Table 7) and additionally evaluated sparse cross-view overlap through directly controlled common observation rates (Page 9, Figure 3; Page 18, Appendix F.2, Table 8). At η=0.7\eta=0.7, CGBST achieves the best ACC, NMI, and ARI on both MSRCV1 and OutdoorScene among the compared methods, obtaining 86.05% ACC, 78.23% NMI, and 73.87% ARI on MSRCV1, and 72.03% ACC, 57.71% NMI, and 48.99% ARI on OutdoorScene.

The partial-overlap experiments further evaluate scenarios with limited co-observed anchors. At the lowest target common observation rate ρ=0.3\rho=0.3, the realized pairwise co-observation rates are 28.10–34.29% on MSRCV1 and 28.61–32.03% on OutdoorScene. CGBST achieves higher performance than pure instance-level transfer across all four metrics on both datasets.

These complementary experiments demonstrate the advantages of CGBST over instance-level transfer in recovering neighborhood structure under high missing rates and sparse cross-view observations.

Table R3. Supplementary comparison at η=0.7\eta=0.7 (Mean, %; Page 18, Appendix F.1, Table 7).

MethodMSRCV1 ACCNMIARIF-scoreOutdoorScene ACCNMIARIF-score
Mask-IMvC83.0675.2666.8575.0255.8045.9932.0645.14
DSIMVC++35.2426.2915.8030.8755.5845.5532.9255.38
GIGA39.5230.7512.7638.7420.228.550.9813.04
BURG69.2161.0551.6665.2941.8336.6821.7440.20
CPSPAN70.0058.7450.4167.5362.6950.5140.3662.70
CPSPAN-mask74.6065.5357.2472.6564.1751.9042.1364.19
DCCA35.7122.9110.3037.5621.597.262.6822.22
DCG50.7931.0622.5350.1766.6756.5747.4766.50
GFSAF20.634.61-0.3819.8415.300.460.0414.38
SMHGC27.6212.325.1124.9121.077.593.7919.59
DIMVC-HIA19.526.112.309.0560.7348.9937.4660.76
SCM78.7370.8160.6379.1566.8254.8944.8166.64
CGBST (ours)86.0578.2373.8778.8872.0357.7148.9957.88

We would like to follow up on our responses and the additional analyses presented in the rebuttal. We hope that the new experimental evidence and clarifications have addressed your questions regarding the methodological contributions and empirical evaluation of CGBST. We welcome any further feedback and would be happy to clarify any remaining concerns.

Reviewer_KBq4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the issue that traditional sample-level relation recovery is susceptible to local noise. It introduces granular balls as intermediate structural units and infers structural information for missing samples through cross-view region correspondences. The method is evaluated on four benchmark datasets under different missing rates. Overall, the experiments are relatively comprehensive and provide evidence of the effectiveness of the proposed method.

Critical strengths

I appreciate the authors’ idea of transferring granular-ball structures to missing views, as it provides an alternative to traditional sample-level matching for recovering sample relationships in missing views. The overall method is well designed, combining cross-view structure transfer with graph-guided contrastive learning. The authors also conduct experiments on multiple datasets and use ablation studies to analyze the contributions of different modules.

Critical weaknesses
  1. The paper does not sufficiently consider the varying confidence levels of different granular-ball correspondences. When only a small number of samples are observed across views, or when granular-ball partitions differ substantially between views, incorrect correspondences may be propagated, affecting subsequent structure recovery and clustering results.

  2. The paper lacks an in-depth analysis of the complementary relationship between graph-guided contrastive learning and Laplacian regularization, and does not sufficiently explain the necessity of using both components together.

  3. The experiments lack validation under high missing-rate scenarios. It is recommended to report the results of CGBST and the main baseline methods separately under multiple missing rates.

  4. The parameter settings in Section 4.1 show substantial differences across datasets. For example, the parameters for MSRCV1 are set to (0,0,−1)(0,0,-1), while those for NUS-WIDE are set to (2,−3,3)(2,-3,3). Although the authors conduct parameter analysis, they do not explain the rationale behind these parameter choices.

  5. The paper lacks sufficient experiments to demonstrate the actual runtime, GPU memory consumption, and scalability of the proposed method on larger-scale datasets.

Questions for authors
  1. The paper claims that granular-ball structure transfer reduces sensitivity to local noise. However, the robustness of the model to noise perturbations is not directly evaluated, and the claimed advantage in noise resistance has not been sufficiently validated experimentally.

  2. The authors do not report standard deviations across multiple runs or conduct statistical significance tests. For example, on the OutdoorScene dataset under the 0.1 missing-rate setting, the NMI is 61.70, while Mask-IMvC achieves 61.62. Without variance information, it is difficult to determine whether the reported performance difference is statistically significant.

  3. The paper lacks a unified notation table. The methods section introduces numerous mathematical symbols and variables, making it difficult for readers to understand the overall method. It is recommended that the authors include a notation table summarizing the main variables, matrices, and hyperparameters.

7 comments

R1. Reliability of Granular-Ball Correspondences under Sparse Overlap and Noise (W1 & Q1)

We appreciate the reviewer's suggestion to further evaluate CGBST under sparse cross-view overlap and noisy conditions. We have added a dedicated evaluation under partial cross-view overlap and noise perturbations (Page 9, Figure 3; Page 18, Appendix F.2, Table 8). The experiments vary the common observation rate ρ∈{0.3,0.4,0.5}\rho\in\{0.3,0.4,0.5\} at a fixed noise level σ=0.25\sigma=0.25, comparing CGBST with pure instance-level transfer and reporting the realized pairwise co-observation rates.

CGBST consistently outperforms pure instance-level transfer across all four metrics on MSRCV1, achieving gains of 7.15, 6.16, 8.88, and 5.16 percentage points in ACC, NMI, ARI, and F-score at ρ=0.3\rho=0.3, respectively. It also improves all four metrics on OutdoorScene at ρ=0.3\rho=0.3 and ρ=0.4\rho=0.4. These results highlight the effectiveness of region-mediated structure transfer in recovering neighborhood relationships under sparse cross-view observations and noisy conditions.

Table R1. Partial cross-view overlap and noise robustness (σ=0.25\sigma=0.25; Mean, %).

Datasetρ\rhoTransfer variantACCNMIARIF-score
MSRCV10.3Instance-pure64.5256.2543.0856.48
CGBST (ours)71.6762.4151.9661.64
0.4Instance-pure74.2966.1558.3165.81
CGBST (ours)79.1069.3361.3968.41
0.5Instance-pure80.6272.0960.6670.74
CGBST (ours)85.8675.2269.8075.83
OutdoorScene0.3Instance-pure63.6346.5337.1648.51
CGBST (ours)64.5648.6238.7449.68
0.4Instance-pure59.5349.2839.1749.83
CGBST (ours)66.2851.3042.1851.33
0.5Instance-pure68.5752.6442.2753.69
CGBST (ours)68.1751.9241.5652.93

Pairwise co-observation rates for ρ=0.3/0.4/0.5\rho=0.3/0.4/0.5: MSRCV1 28.10–34.29% / 28.57–42.86% / 45.71–52.86%; OutdoorScene 28.61–32.03% / 38.06–40.36% / 49.40–51.53%.

R2. Complementarity and Necessity of Graph-Guided Contrastive Learning and Laplacian Regularization (W2)

We appreciate the reviewer's suggestion to further examine the complementary roles of graph-guided contrastive learning and Laplacian regularization. The two objectives provide complementary supervision through the recovered graph. Graph-guided contrastive learning incorporates recovered relations into positive and negative pair weighting, promoting neighborhood-aware semantic grouping in the shared clustering output. Laplacian regularization encourages connected samples to remain close in the common embedding space, preserving local geometric continuity along transferred neighborhoods. Together, they guide both clustering discrimination and representation geometry.

We have expanded this analysis and added a joint-removal ablation alongside the individual component ablations (Page 8, Section 4.3, Table 2). Removing both graph fusion and Laplacian regularization reduces all four clustering metrics on CUB, MSRCV1, and OutdoorScene. Specifically, the full CGBST improves ACC over the joint-removal variant by 7.24, 4.04, and 9.30 percentage points, respectively. On CUB, the joint-removal variant also performs below both individual-removal controls across all four metrics.

These results demonstrate the complementary contributions of graph-guided clustering supervision and local geometric regularization to common representation learning.

Table R2. Recovered-graph learning ablation at η=0.3\eta=0.3 (Mean, %; extracted from Page 8, Table 2).

DatasetVariantACCNMIARIF-score
CUBFull CGBST77.3776.6465.5173.01
w/o graph fusion77.9076.1964.8371.31
w/o Laplacian71.3371.5957.4565.76
w/o graph fusion and Laplacian70.1369.4054.8863.64
MSRCV1Full CGBST93.4085.9384.4087.32
w/o graph fusion88.4082.3977.8982.61
w/o Laplacian90.0982.5678.5283.30
w/o graph fusion and Laplacian89.3682.7078.3083.01
OutdoorSceneFull CGBST74.1360.7952.4060.50
w/o graph fusion72.9959.9751.3459.39
w/o Laplacian64.8654.5245.5254.86
w/o graph fusion and Laplacian64.8354.4045.8155.03
NUS-WIDEFull CGBST55.0453.0834.5352.29
w/o graph fusion55.5853.1733.1552.87
w/o Laplacian54.2245.1438.8548.04
w/o graph fusion and Laplacian55.8846.4640.3949.20

R3. Performance under Higher Missing Rates (W3)

We have extended the missing-rate evaluation to η=0.7\eta=0.7, supplementing the original comparisons at η∈{0.1,0.3,0.5}\eta\in\{0.1,0.3,0.5\} (Page 18, Appendix F.1, Table 7). At η=0.7\eta=0.7, CGBST achieves the best ACC, NMI, and ARI on both MSRCV1 and OutdoorScene among the compared methods. Specifically, CGBST obtains 86.05% ACC, 78.23% NMI, and 73.87% ARI on MSRCV1, and 72.03% ACC, 57.71% NMI, and 48.99% ARI on OutdoorScene.

These results extend the empirical evidence for CGBST to more challenging missing-view conditions, demonstrating strong clustering performance across multiple missing rates.

Table R3. Supplementary comparison at η=0.7\eta=0.7 (Mean, %; Page 18, Appendix F.1, Table 7).

MethodMSRCV1 ACCNMIARIF-scoreOutdoorScene ACCNMIARIF-score
Mask-IMvC83.0675.2666.8575.0255.8045.9932.0645.14
DSIMVC++35.2426.2915.8030.8755.5845.5532.9255.38
GIGA39.5230.7512.7638.7420.228.550.9813.04
BURG69.2161.0551.6665.2941.8336.6821.7440.20
CPSPAN70.0058.7450.4167.5362.6950.5140.3662.70
CPSPAN-mask74.6065.5357.2472.6564.1751.9042.1364.19
DCCA (all-pairwise)35.7122.9110.3037.5621.597.262.6822.22
DCG (multiview)50.7931.0622.5350.1766.6756.5747.4766.50
GFSAF20.634.61-0.3819.8415.300.460.0414.38
SMHGC27.6212.325.1124.9121.077.593.7919.59
DIMVC-HIA19.526.112.309.0560.7348.9937.4660.76
SCM78.7370.8160.6379.1566.8254.8944.8166.64
CGBST (ours)86.0578.2373.8778.8872.0357.7148.9957.88

R4. Rationale for Dataset-Dependent Hyperparameter Settings (W4)

The contrastive weight α\alpha is selected with reference to dataset characteristics, including sample size, feature dimensionality, number of views, and missingness patterns. We further examine the relative gradient magnitudes of reconstruction, contrastive learning, and Laplacian regularization. Specifically, a label-free gradient-balancing criterion recommends α\alpha by bringing the weighted contrastive gradient to a scale comparable to the other optimization objectives, with the recommendation mapped to {−3,−2,…,3}\{-3,-2,\ldots,3\}. This provides an optimization-based reference for hyperparameter selection without relying on ground-truth labels. These dataset-specific characteristics and gradient magnitudes account for the substantial differences in parameter settings across datasets.

R5. Computational Efficiency and Scalability (W5)

We have added runtime and GPU memory profiling for all four benchmarks (Pages 14–15, Appendix B.5, Table 6). CGBST completes training in 2.97–20.80 minutes, with peak allocated GPU memory ranging from 52.29 MB to 1,984.29 MB on an NVIDIA RTX 3050 GPU (4 GB). Regarding scalability, we have discussed this limitation in the Conclusion (Page 10, Section 5). The current formulation involves quadratic sample-level relation graphs, which constrain its applicability to larger datasets. Future work will investigate sparse relation recovery and uncertainty-aware region transfer to extend CGBST to larger-scale and more severely incomplete datasets.

R6. Unified Notation for Mathematical Formulation (Q3)

We appreciate the reviewer's suggestion to improve notation clarity. We have added a unified notation table (Page 13, Appendix A, Table 3), summarizing the main variables, matrices, and hyperparameters used throughout the methodology.

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

The paper proposes CGBST for incomplete multi-view clustering. Observed embeddings in each view are partitioned into adaptive granular-ball regions, and co-observed samples establish soft correspondences between the resulting region systems. These mappings infer region affiliations for samples missing a target view. Membership overlap and target-region topology then reconstruct sample relations, which guide contrastive learning and Laplacian smoothing of a shared embedding. Experiments compare clustering performance on four datasets under three missingness settings.

Critical strengths
  • Soft correspondences between granular regions accommodate unequal region counts and ambiguous cross-view matches. This provides a concrete way to transfer neighborhood structure without synthesizing missing target-view features.

  • The recovery mask prevents completed relations from introducing new edges between already observed samples. This constrains reconstruction to the missing-information problem instead of freely replacing observed structure.

  • Combining membership overlap with inter-region topology allows recovered relations to reflect broader neighborhood organization. Contrastive learning and Laplacian smoothing then connect this transferred structure to the shared embedding.

Critical weaknesses
  • The transfer mapping depends on co-observed samples to connect granular regions. Its stability when that support becomes sparse is not directly established, leaving the mechanism under the setting that most strongly motivates regional aggregation insufficiently validated.

  • The instance-level transfer control is not defined precisely enough to isolate regional aggregation. Mixed comparative results therefore leave unclear whether the benefit comes from the region representation, its correspondence mapping, or other retained components.

  • The balance between contrastive learning and structural smoothing is configured separately across conditions without a clear unsupervised selection rule. This limits practical use of the joint objective without access to evaluation labels.

Questions for authors
  • How was $ \alpha$ selected for each reported setting? Please specify the criterion and any validation data, including whether evaluation labels were used.

  • Can the authors provide a dedicated low-overlap evaluation, reporting actual pairwise co-observation rates, to test transfer when anchor evidence becomes sparse?

  • How is the instance-level transfer control constructed, particularly what replaces the region mapping and which other components remain fixed?

  • What epochs do the points in Figure 4 represent relative to the epoch-100 activation of granular recovery and the stated 1,500-epoch training schedule?

4 comments

R1. Robustness of Cross-View Structure Transfer under Sparse Overlap (W1 & Q2)

We appreciate the reviewer's suggestion to further evaluate CGBST under sparse cross-view overlap and noisy conditions. We have added a dedicated partial cross-view overlap evaluation to directly examine the effectiveness of granular-ball structure transfer with limited co-observed anchors (Page 9, Figure 3; Page 18, Appendix F.2, Table 8).

The experiments vary the target common observation rate ρ∈{0.3,0.4,0.5}\rho\in\{0.3,0.4,0.5\} under a fixed noise level σ=0.25\sigma=0.25, comparing CGBST with a pure instance-level transfer control. We also report the realized pairwise co-observation rates in Appendix F.2, providing a direct characterization of the available cross-view anchor evidence.

CGBST consistently outperforms pure instance-level transfer across all four metrics on MSRCV1, achieving gains of 7.15, 6.16, 8.88, and 5.16 percentage points in ACC, NMI, ARI, and F-score at ρ=0.3\rho=0.3, respectively. It also improves all four metrics on OutdoorScene at ρ=0.3\rho=0.3 and ρ=0.4\rho=0.4. These results highlight the effectiveness of region-mediated transfer in preserving neighborhood structure under sparse cross-view observations.

Table R1. Partial cross-view overlap and noise robustness (σ=0.25\sigma=0.25; Mean, %).

Datasetρ\rhoTransfer variantACCNMIARIF-score
MSRCV10.3Instance-pure64.5256.2543.0856.48
CGBST (ours)71.6762.4151.9661.64
0.4Instance-pure74.2966.1558.3165.81
CGBST (ours)79.1069.3361.3968.41
0.5Instance-pure80.6272.0960.6670.74
CGBST (ours)85.8675.2269.8075.83
OutdoorScene0.3Instance-pure63.6346.5337.1648.51
CGBST (ours)64.5648.6238.7449.68
0.4Instance-pure59.5349.2839.1749.83
CGBST (ours)66.2851.3042.1851.33
0.5Instance-pure68.5752.6442.2753.69
CGBST (ours)68.1751.9241.5652.93

Pairwise co-observation rates for ρ=0.3/0.4/0.5\rho=0.3/0.4/0.5: MSRCV1 28.10–34.29% / 28.57–42.86% / 45.71–52.86%; OutdoorScene 28.61–32.03% / 38.06–40.36% / 49.40–51.53%.

R2. Clarification and Evaluation of Instance-Level Transfer Controls (W2 & Q3)

We appreciate the reviewer's careful observation. We acknowledge that our original description of the Instance-level Transfer control was not sufficiently clear, which may have caused confusion regarding the contribution of granular-ball structure modeling.

The original Instance-level Transfer retains granular-ball region construction but replaces the region-level correspondence mapping with instance-anchor-based transfer. The remaining components, including relation recovery, graph fusion, contrastive learning, and Laplacian regularization, are preserved.

To further isolate the contribution of granular-ball structure modeling, we have added an Instance-pure Transfer control that removes granular-ball construction and directly transfers instance-level graph relations while retaining the same downstream learning framework (Page 8, Table 2).

The results show that CGBST outperforms Instance-pure Transfer across all four metrics on MSRCV1 and OutdoorScene. At η=0.3\eta=0.3, the improvements reach 2.35 percentage points in ARI on MSRCV1 and 2.99 percentage points in ACC on OutdoorScene. Together with the partial-overlap evaluation (Page 9, Figure 3; Page 18, Appendix F.2, Table 8), these results highlight the value of adaptive regional aggregation and cross-view region correspondences for recovering missing neighborhood structure.

Table R2. Instance-level transfer ablation at η=0.3\eta=0.3 (Mean, %; extracted from Page 8, Table 2).

DatasetVariantACCNMIARIF-score
CUBFull CGBST77.3776.6465.5173.01
Instance-pure77.9475.6465.0670.87
MSRCV1Full CGBST93.4085.9384.4087.32
Instance-pure92.1184.5182.0585.85
OutdoorSceneFull CGBST74.1360.7952.4060.50
Instance-pure71.1458.4250.0658.29
NUS-WIDEFull CGBST55.0453.0834.5352.29
Instance-pure54.5753.9534.3952.32

R3. Label-Free Hyperparameter Selection (W3 & Q1)

We distinguish between shared training hyperparameters and dataset-dependent hyperparameters. Most training and structural hyperparameters, including the training schedule, granular-ball structure transfer coefficients, and Laplacian regularization weight, are kept consistent across experimental settings to ensure fair comparisons.

The contrastive weight α\alpha is selected with reference to dataset characteristics, including sample size, feature dimensionality, number of views, and missingness patterns. We further examine the relative gradient magnitudes of reconstruction, contrastive learning, and Laplacian regularization. Specifically, a label-free gradient-balancing criterion recommends α\alpha by bringing the weighted contrastive gradient to a scale comparable to the other optimization objectives, with the recommendation mapped to the predefined range {−3,−2,…,3}\{-3,-2,\ldots,3\}. This provides an optimization-based criterion for hyperparameter selection without relying on ground-truth labels.

R4. Clarification of Training Epochs and Granular Structure Activation (Q4)

We appreciate the reviewer's careful observation. The original Figure 4 displayed the first 100 epochs, corresponding to the warm-up stage before granular structure recovery was activated. Specifically, CGBST follows a 1,500-epoch training schedule, with granular structure recovery activated at epoch 100.

To illustrate the training dynamics after activation, we have redrawn the curves to cover the first 500 epochs (Page 10, Figure 5). The updated figure includes both the warm-up stage and the subsequent optimization with granular structure recovery, showing the evolution of the training objective and clustering performance (ACC, NMI, ARI, and F-score) across all four datasets.

Reviewer_jdeQOpenReview ↗

My problems have been resolved.

Thank you for the author's reply. My concerns had been resolved. I suggest accepting this paper.

Thank you for your positive feedback and for reconsidering our work. We are pleased that our responses have addressed your concerns. We sincerely appreciate your constructive suggestions and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.