acceptodds
Under review as a conference paper at ICLR 2027

Granular-Ball Reliability Learning: Adaptive Splitting and Risk-Calibrated Connection for Deep Multi-View Clustering

12 views

Abstract

Deep multi-view clustering increasingly learns representations, sample relations, and granular structures to integrate complementary information across views. However, fixing mixed local granular balls entrenches within-granular ball errors, while incorrectly merging pure granular balls amplifies between-granular ball errors. We propose GBASC, which combines Granular-Ball Adaptive Splitting (GBAS) to refine mixed local granular balls using cross-view evidence with Granular-Ball Adaptive Connection (GBAC) to determine global assignments from reliable between-granular ball associations. Our contributions are an exact decomposition of cross-granularity pairwise errors into within-atom mixing and between-atom assignment errors, and a unified framework that jointly controls both transitions through adaptive refinement, risk-aware partition selection, and reliability-weighted representation learning. GBASC achieves state-of-the-art clustering performance on six benchmark datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, Deep multi-view clustering, Granular ball, Cluster

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5JfcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GBASC, a deep multi-view clustering framework that addresses structural errors through adaptive granular-ball splitting and connection. GBAS evaluates candidate splits using multiple criteria, including compactness, neighborhood structure, cross-view agreement, stability, and spectral guidance. GBAC assesses the reliability of inter-ball connections, attenuates high-risk relations, and selects global cluster assignments from multiple candidates using additional quality and trust checks. The resulting clustering structure further guides representation learning. Extensive experiments demonstrate strong clustering performance across multiple benchmarks, supported by ablation studies and structural diagnostics.

Critical strengths
  1. Strong and consistent empirical performance. GBASC achieves the best reported ACC, NMI, and ARI across all six main benchmarks against nine baselines, with further competitive results in the extended evaluations.

  2. Comprehensive and targeted experimental validation. The paper provides factorial ablations and targeted diagnostics of splitting quality, risk attenuation, and partition selection.

Critical weaknesses

1. Methodological Novelty and Design Justification.

While GBASC introduces new mechanisms, its overall refinement-and-connection framework builds on existing granular-ball clustering methods (e.g., GBCT and MGBCC). It involves numerous heuristic choices, such as fixed splitting-evidence weights and additional spectral-guided decision rules in GBAS, and weighted candidate scoring and trust checks in GBAC. Although intuitively motivated, their specific formulations and individual contributions remain insufficiently justified, while the theoretical results mainly establish general or conditional properties rather than directly supporting these design choices.

2. Unclear Practical Viability.

GBASC involves numerous fixed hyperparameters, including splitting weights, complexity penalties, and risk attenuation coefficients, without systematic sensitivity analysis of the final clustering performance, leaving its tuning requirements unclear. Moreover, repeated graph construction, spectral decomposition, and candidate partition evaluation may incur substantial computational costs, yet neither an overall complexity analysis nor runtime comparisons with baselines are provided.

Questions for authors
  1. Could the authors further clarify the contributions of the key mechanisms within GBAS and GBAC? For example, by isolating risk attenuation from five-candidate selection (and clarifying their settings in Table 3), or evaluating the contribution of terminal spectral-guided splitting. Targeted comparisons on representative datasets would help establish which mechanisms drive the performance gains.

  2. The proposed method involves numerous manually specified hyperparameters, such as the GBAS splitting weights and the risk attenuation coefficient β. Could the authors provide brief sensitivity analysis or other evidence showing that the clustering performance is robust to these hyperparameter choices?

  3. Given the repeated graph construction, spectral decomposition, and multiple candidate partition evaluations in GBASC, could the authors provide an overall computational complexity analysis and end-to-end runtime comparisons with representative baselines on ALOI (over 10,000 samples)?

5 comments
AuthorsOpenReview ↗

Answer to Question 1

We thank the reviewer for this valuable suggestion. We agree that our original module-level ablation in Table 3 does not fully disentangle the contributions of individual decision mechanisms within GBAS and GBAC.

1. Distinct roles of the decision mechanisms.

In GBAS, candidate evidence is used to rank proposed splits, whereas the subsequent acceptance decision controls whether a split should be executed. The cost criterion discourages excessive fragmentation, while the structure criterion permits boundaries supported by global structural evidence. The terminal spectral-guided refinement further addresses boundaries that may not be represented in the local candidate pool.

In GBAC, risk attenuation calibrates inter-atom affinities before obtaining the base partition. Five candidate slots provide alternative global assignments, while trust constraints regulate whether the best-scoring nonbase candidate can safely replace the base partition. If the acceptance conditions are not satisfied, the method retains the base partition.

2. Fine-grained component ablations.

To clarify these contributions, we conduct targeted comparisons on ORL, DHA, and ALOI under consistent evaluation protocols.

For GBAS, we distinguish cost-only acceptance (A1), structure-only acceptance (A2), dual-gate acceptance without terminal guidance (A3), and full GBAS with terminal spectral guidance (A4).

The results indicate that A3 consistently outperforms either single-criterion variant across the three datasets. Adding terminal spectral guidance further improves ALOI ACC from 95.50% to 96.85% and reduces pairwise error from 0.14% to 0.12%. These results support the complementary roles of local acceptance and terminal refinement.

For GBAC, we separately examine risk attenuation, five-candidate selection, and trust-constrained acceptance.

On ALOI, enabling risk attenuation while retaining base-only assignment (B1 vs. B3) increases ACC from 90.20% to 95.50%, with pairwise error decreasing from 0.29% to 0.16%. In contrast, enabling candidate selection without risk attenuation (B1 vs. B2) improves ACC by only 0.60 percentage points.

With risk attenuation enabled, unconstrained candidate selection (B4) achieves 95.10% ACC and 0.18% pairwise error, whereas trust-constrained selection (B5) reaches 96.85% ACC and 0.12% pairwise error. This comparison suggests that trust constraints help prevent harmful candidate replacements while retaining beneficial ones.

Together, these comparisons indicate that the components play distinct and complementary roles, rather than merely adding heuristic complexity.

3. Clarification of experimental settings.

We will explicitly document the activation of risk attenuation, candidate selection, and trust constraints in Table 3, including which settings remain fixed across the P-Connection and Risk Connection variants.

We will also revise Sections 3.2–3.3 to distinguish candidate ranking, split acceptance, terminal refinement, partition selection, and fallback. The fine-grained comparisons will be incorporated into Section 4.3 and the supplementary material.

We appreciate the reviewer's suggestion, which helps us present a more precise explanation of the contributions of the individual mechanisms.

AuthorsOpenReview ↗

Answer to Question 2

We thank the reviewer for this valuable suggestion. We conducted sensitivity experiments on ORL and DHA to evaluate the robustness of GBASC to the risk attenuation coefficient β and the GBAS candidate-evidence weights (ω_m, ω_f, ω_a). Each configuration was evaluated using 10 random seeds (42–51), with all other settings fixed. Results are reported as mean ± std (%).

1. Sensitivity to β.

Following Eqs. (7) and (11), β controls risk attenuation in graph construction and reliability-weighted structural feedback. We varied β ∈ {0, 0.10, 0.25, 0.50, 1.00}, where β = 0 serves as a no-attenuation control, while fixing the GBAS weights at their default values.

Table 1. Sensitivity to risk attenuation coefficient β.

DatasetβACC ↑NMI ↑ARI ↑Purity ↑
ORL0.0083.55±1.2391.20±0.5576.16±1.5784.53±1.11
ORL0.1083.40±1.2390.98±0.5575.45±1.5084.10±0.96
ORL0.25*83.22±1.1391.02±0.5875.64±1.4184.35±1.01
ORL0.5083.62±0.9991.08±0.4575.87±1.0984.55±0.84
ORL1.0083.92±1.5191.40±0.6576.64±1.7784.95±1.52
DHA0.0086.44±0.1588.91±0.1579.36±0.2487.23±0.14
DHA0.1086.46±0.1188.96±0.1379.46±0.2187.29±0.11
DHA0.25*86.44±0.1888.92±0.2079.37±0.3287.23±0.20
DHA0.5086.34±0.1488.77±0.1579.15±0.2387.10±0.14
DHA1.0086.21±0.2688.58±0.2278.83±0.3986.98±0.18

*Default setting.

Across the tested β values, ACC varies by at most 0.70 percentage points on ORL and 0.25 on DHA, with similarly limited variations in NMI, ARI, and Purity. These results indicate that GBASC maintains stable clustering performance without requiring precise tuning of β within the investigated range.

2. Sensitivity to GBAS candidate-evidence weights.

As defined in Eq. (3), (ω_m, ω_f, ω_a) weight compactness improvement, neighborhood fracture, and cross-view agreement when ranking candidate splits. We evaluated four configurations while fixing β = 0.25.

Table 2. Sensitivity to GBAS candidate-evidence weights.

Dataset(ω_m, ω_f, ω_a)ACC ↑NMI ↑ARI ↑Purity ↑
ORL(0.45,0.30,0.25)*83.22±1.1391.02±0.5875.64±1.4184.35±1.01
ORL(1/3,1/3,1/3)82.98±1.5991.00±0.6375.27±1.6484.10±1.33
ORL(0.35,0.40,0.25)84.03±1.1991.34±0.4976.40±1.2984.65±1.06
ORL(0.55,0.25,0.20)84.05±1.0091.23±0.4676.25±1.2884.90±0.93
DHA(0.45,0.30,0.25)*86.44±0.1888.92±0.2079.37±0.3287.23±0.20
DHA(1/3,1/3,1/3)86.29±0.1688.69±0.1478.99±0.2287.06±0.11
DHA(0.35,0.40,0.25)86.48±0.1488.95±0.1479.43±0.2287.27±0.11
DHA(0.55,0.25,0.20)86.31±0.1588.82±0.1279.14±0.2087.10±0.10

*Default setting.

Changing the evidence weights produces maximum ACC variations of only 1.08 percentage points on ORL and 0.19 on DHA. The remaining metrics also exhibit limited fluctuations, suggesting that candidate ranking is not critically dependent on a particular weight allocation.

3. Complementary ablation evidence.

Due to space limitations, we summarize the component ablations without reproducing the existing ablation table. As reported in Table 3 of the manuscript, on DHA, adaptive splitting improves ACC from 78.20% to 85.70% under P-Connection. On ALOI, replacing P-Connection with Risk Connection under adaptive splitting increases ACC from 90.60% to 96.85% and ARI from 87.25% to 93.85%. These results demonstrate the complementary contributions of GBAS and GBAC, while the new sensitivity experiments show that performance remains relatively stable under the tested parameter variations.

We will incorporate these sensitivity results into Section 4.3 and clarify the parameter definitions and default settings in Sections 3.2–3.3, including the shared use of β in structural feedback.

We sincerely thank the reviewer for this constructive suggestion.

AuthorsOpenReview ↗

Answer to Question 3

We sincerely thank the reviewer for this valuable suggestion. We agree that computational complexity and end-to-end runtime are important for evaluating GBASC's scalability. We provide the following analysis and additional experimental results, which will be incorporated into the revised manuscript.

1. Computational Complexity

Let NN, VV, DD, and MM denote the numbers of samples, views, total feature dimensions, and resulting atomic granular balls, respectively.

The main computational costs arise from: (1) multi-view graph construction and Ward initialization, requiring O(VN2D)\mathcal{O}(VN^2D); (2) GBAS local refinement, requiring O(LN(VbD+b2))\mathcal{O}(LN(VbD+b^2)) with precomputed graphs, where the maximum ball size b=32b=32 and split depth L=4L=4 are bounded; and (3) GBAC quotient-graph spectral decomposition, requiring O((V+1)M3)\mathcal{O}((V+1)M^3). Importantly, GBAC evaluates only five candidate partitions with bounded optimization iterations, avoiding combinatorial partition enumeration.

Including two structural evaluations and TT training epochs, the overall leading complexity is

O(TNPeff+2[VN2D+(V+1)M3+(V+2)SN+H]),\mathcal{O}\left(TNP_{\mathrm{eff}}+2\left[VN^2D+(V+1)M^3+(V+2)\mathcal{S}_N+\mathcal{H}\right]\right),

where PeffP_{\mathrm{eff}} denotes per-sample training cost, SN\mathcal{S}_N denotes sample-level sparse spectral solving cost, and H\mathcal{H} includes local refinement and implementation overhead.

Under fixed configurations and controlled auxiliary costs, the dominant complexity simplifies to O(N2+M3)\mathcal{O}(N^2+M^3). This is not an unconditional worst-case bound. GBASC's efficiency primarily comes from performing global assignments on MM atoms instead of NN samples, while restricting local refinement and candidate evaluation.

2. End-to-End Runtime on ALOI

We additionally report end-to-end runtimes on ALOI (10,800 samples, 4 views, 100 clusters).

MethodRuntime (s) ↓Speedup
DCMVC2411.41 ± 6.203.66×
CorreGen2969.67 ± 477.064.51×
RAV4257.64 ± 1025.916.47×
RAC-DMVC2277.13 ± 401.563.46×
GBASC (Ours)658.44 ± 1.221.00×

GBASC completes clustering in 658.44 seconds (10.97 minutes), achieving 3.46–6.47× speedups over these representative baselines. Meanwhile, it achieves 96.85% ACC, 97.04% NMI, and 93.85% ARI, demonstrating a favorable balance between clustering accuracy and computational efficiency.

3. Revision Commitment

We appreciate the reviewer's constructive feedback. In the revised manuscript, we will add (i) a detailed module-wise complexity analysis with explicit assumptions, (ii) end-to-end runtime comparisons on ALOI, and (iii) hardware specifications and timing protocols for reproducibility. We believe these additions will strengthen the paper's evaluation of computational efficiency and scalability.

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

This paper proposes a new deep multi-view clustering method termed GBASC. It focuses on two structural sources of clustering error, including mixed local granular balls and unreliable connections between granular balls. GBASC consist of two main modules to distinguish potentially mixed balls and reduces risky inter-ball edges. Experiments demonstrate the advantages of GBASC.

Critical strengths
  1. The decomposition of pairwise error into within-ball and between-ball components provides a clear motivation for the proposed method.
  2. Compared with theadvanced competitors, GBASC achieves superior performance.
Critical weaknesses
  1. The idea of enhancing intra-class connections and weakening inter-class edges has been explored in existing works, such as [1]. This article seems to extend the idea to granular-ball learning without introducing sufficiently technological improvements. [1] Homophily-enhanced Structure Learning for Graph Clustering. 2023. CIKM.
  2. The theoretical guarantees are relatively weak. The decomposition is an identity under atom-constant predictions, while the refinement theorem mainly guarantees non-increasing internal mixing. The theorem does not establish the connection between new modules and the total clustering error.
  3. The semantic interpretation of risk-aware connection depends on additional assumptions, such as atom purity and an appropriate ordering between estimated risk and true cross-class contamination. These assumptions may not hold in complex multi-view settings.
  4. The presentation is somewhat inconsistent. Section 4.1 lists eight datasets, while the absteact refers to six datasets.
Questions for authors

Please see the weakness list.

5 comments
AuthorsOpenReview ↗

Answer to Question 1

We thank the reviewer for highlighting the relevant work by Gu et al. (CIKM 2023) [1]. We agree that enhancing intra-cluster connectivity and suppressing inter-cluster connections are established principles. HoLe makes an important contribution through hierarchical correlation estimation and cluster-aware sparsification, using confident clustering assignments and embedding similarities to refine graph connectivity. Related advances in multi-view graph consensus learning (Pan & Kang, NeurIPS 2021) [2] and view-aware relationship alignment (Peng et al., ICLR 2026) [3] further demonstrate the importance of reliable structural relationships. We fully acknowledge these contributions and do not claim homophily enhancement itself as our novelty.

Our technical contribution lies in explicitly modeling and controlling two distinct sources of structural error across granularity transitions, rather than simply extending homophily enhancement to granular balls. Although granular-ball representations and splitting have been explored in prior multi-view clustering research, including MGBCC (Su et al., AAAI 2025) [4] and GFSAF (Liu et al., AAAI 2026) [5], our work specifically addresses the reliability of atomic boundary formation and subsequent global assignment.

First, GBAS addresses within-atom class mixing, which cannot be corrected by inter-atom edge refinement when atomic units remain indivisible. Our exact pairwise-error decomposition (Eq. 1) identifies this irreducible error component under fixed atomic assignments. Unlike structure refinement that primarily modifies node connections, GBAS evaluates whether the atomic units themselves should be revised. It integrates cross-view compactness, neighborhood fracture, view agreement, and perturbation stability to identify candidate boundaries, followed by local-cost and structural-gain acceptance tests (Eqs. 3–5). This provides an explicit mechanism for correcting unreliable atomic boundaries before global clustering, rather than relying solely on improvements in representations or external connectivity.

Second, GBAC extends reliability assessment from graph connectivity to global partition decisions. While HoLe [1] employs confidence-guided edge recovery and sparsification, and RAV [3] aligns view-specific and global sample relationships, GBAC explicitly estimates connection uncertainty using cross-view support and affinity dispersion (Eq. 2). It separately aggregates positive affinity and risk mass into a quotient graph and attenuates connections according to their relative uncertainty (Eq. 7). More importantly, GBAC evaluates five complete K-partitions and permits replacement of the base assignment only when explicit trust constraints on graph cuts, stability, and assignment changes are satisfied (Eqs. 8–9). Thus, reliability determines not only connection strength but also whether a global reassignment should be accepted.

Our experiments provide mechanism-level evidence for these complementary improvements. Appendix G.3 demonstrates improved cross-class pair separation at matched child-block sizes, supporting the effectiveness of adaptive boundary refinement. In Table 3, adaptive splitting improves DHA ACC/ARI from 78.20/68.90 to 85.70/78.65 under fixed P-Connection, while risk-aware connection improves ALOI ACC/ARI from 90.60/87.25 to 96.85/93.85 under fixed adaptive splitting. These results support the distinct contributions of GBAS and GBAC, without implying direct performance superiority over HoLe.

We will strengthen the related-work discussion to better distinguish our approach from these established methods. Rather than claiming a new homophily principle, GBASC contributes a reliability-aware framework that separately controls atomic boundary refinement and trust-constrained global clustering decisions, supported by theoretical error decomposition and targeted experimental analysis.


References

[1] Gu et al. Homophily-enhanced Structure Learning for Graph Clustering. CIKM 2023. https://doi.org/10.1145/3583780.3614915

[2] Pan and Kang. Multi-view Contrastive Graph Clustering. NeurIPS 2021. https://proceedings.neurips.cc/paper/2021/hash/10c66082c124f8afe3df4886f5e516e0-Abstract.html

[3] Peng et al. Relationship Alignment for View-aware Multi-view Clustering. ICLR 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/ee63888edc6155083c87a8df83581047-Abstract-Conference.html

[4] Su et al. Multi-view Granular-ball Contrastive Clustering. AAAI 2025. https://doi.org/10.1609/aaai.v39i19.34274

[5] Liu et al. Views Attention Fusion of Granular-ball Fuzzy Representations Split for Improved Multi-view Clustering. AAAI 2026. https://doi.org/10.1609/aaai.v40i28.39556

AuthorsOpenReview ↗

Answer to Question 2

We thank the reviewer for this important observation. We agree that the pairwise-error decomposition is an identity under atom-constant assignments and that refinement monotonicity alone does not guarantee lower total clustering error. If refined children inherit their parent's assignment, the clustering remains unchanged. We will clarify these limitations rather than overstate our guarantees.

First, the decomposition distinguishes two error sources requiring different corrective mechanisms. Improvements in graph-based objectives do not necessarily imply better semantic clustering, as reflected in classical spectral clustering analyses (Ng et al., NeurIPS 2001 [1]; Nadler & Galun, NeurIPS 2006 [2]). Our Eq. (1) separates irreducible within-atom mixing from between-atom assignment errors, motivating GBAS and GBAC as complementary corrections.

Second, our analysis provides module-specific conditional guarantees beyond refinement monotonicity. For GBAS, Theorem 3 (Appendix C.1) establishes sufficient conditions for strict semantic leaf-cost reduction under calibrated split evidence and a positive margin. For GBAC, Theorems 2 and 4 show that, under atom-purity and cross-view support–dispersion separation assumptions, risk attenuation suppresses relative cross-class connection exposure by a factor bounded by exp(-βδ_u). The trust-constrained selection rule also ensures that the selected partition does not worsen the surrogate objective relative to the base assignment. These guarantees characterize our mechanisms under explicit assumptions, without implying unconditional reductions in true clustering error.

Third, Appendix E.2 directly connects the combined refinement–connection output to total pairwise error. Corollary 2 compares the final assignment against a pre-refinement baseline, accounting for both erroneous merges and separations. An independent labeled-pair audit provides a high-probability upper bound on their normalized total-error difference. A negative upper bound certifies strict improvement. This is a conditional post-hoc certificate, not a label-free training guarantee, consistent with the need for precisely defined risk guarantees in safe multi-view clustering (Tang & Liu, CVPR 2022 [3]; ICML 2022 [4]).

Finally, our experiments directly evaluate net reductions in total pairwise error. Appendix G.4 compares reconnection with inherited parent assignments on identical refined atomizations, reporting mean error reductions of 0.0903 percentage points on ALOI and 1.8665 on CiteSeer. Appendix G.5 further isolates partition selection, showing additional mean reductions of 0.0848 points on ALOI and 0.1658 on Multi-FMNIST relative to the calibrated base assignment. These results demonstrate empirical net error correction, although gains vary across runs.

We will revise the discussion to distinguish deterministic structural properties, conditional module guarantees, post-hoc total-error certificates, and empirical reductions. Rather than claiming universal monotonic improvement, GBASC provides complementary reliability-aware mechanisms with theoretical and empirical support for controlling both clustering error sources.


References

[1] Ng et al. On Spectral Clustering: Analysis and an Algorithm. NeurIPS 2001. https://proceedings.neurips.cc/paper/2001/hash/801272ee79cfde7fa5960571fee36b9b-Abstract.html

[2] Nadler & Galun. Fundamental Limitations of Spectral Clustering. NeurIPS 2006. https://proceedings.neurips.cc/paper_files/paper/2006/hash/bdb6920adcd0457aa17b53b22963dad9-Abstract.html

[3] Tang & Liu. Deep Safe Multi-View Clustering: Reducing the Risk of Clustering Performance Degradation Caused by View Increase. CVPR 2022. https://openaccess.thecvf.com/content/CVPR2022/html/Tang_Deep_Safe_Multi-View_Clustering_Reducing_the_Risk_of_Clustering_Performance_CVPR_2022_paper.html

[4] Tang & Liu. Deep Safe Incomplete Multi-view Clustering: Theorem and Algorithm. ICML 2022. https://proceedings.mlr.press/v162/tang22c.html

AuthorsOpenReview ↗

Answer to Question 3 and 4

Response to Comment 3: Assumptions Underlying Risk-Aware Connection

We thank the reviewer for raising this important concern. We agree that atom purity and the alignment between estimated risk and true cross-class contamination cannot be universally assumed. Recent studies have demonstrated that noisy correspondences can introduce false positive and negative relations (Guo et al., NeurIPS 2024 [1]; Zhou et al., ICLR 2026 [2]), while noisy views (Yang et al., ICML 2025 [3]) and heterophilic graph structures (Ren et al., ICML 2025 [4]) further complicate reliable semantic relationship estimation. We therefore acknowledge that cross-view agreement does not necessarily imply semantic correctness.

First, our general theoretical analysis does not require pure atoms. Theorem 4 (Appendix C.2.6) defines the true contamination fraction between potentially mixed atoms as Zgh=D(Gg,Gh)/(∣Gg∣∣Gh∣)Z_{gh}=D(G_g,G_h)/(|G_g||G_h|). For attenuation ζ=e−βU\zeta=e^{-\beta U}, it establishes the exact identity

\frac{\operatorname{Cov}_\pi(\zeta,Z)} {\mathbb{E}_\pi[\zeta]}.$$ Thus, attenuation always reduces absolute contamination-weighted positive mass, whereas normalized contamination exposure is non-increasing precisely when the covariance is non-positive. A nondecreasing conditional expectation of contamination given estimated risk is sufficient, but not guaranteed. **The stronger exponential suppression bound additionally requires pure atoms and explicit support–dispersion separation assumptions.** We will clarify this distinction rather than present these assumptions as universally satisfied. **Second, GBAC employs uncertainty proxies and conservative structural decisions rather than assuming perfect semantic calibration.** Eq. (2) combines cross-view support and affinity dispersion to estimate relative connection uncertainty without labels. Eq. (7) applies bounded soft attenuation instead of deleting uncertain edges, while Eq. (9) rejects alternative partitions that violate explicit trust constraints. GBAS also reduces within-atom mixing, although it does not guarantee atom purity. These mechanisms limit aggressive structural changes but cannot eliminate errors caused by systematically misleading views. **Third, our existing experiments provide empirical evidence for the usefulness of the estimated risk ordering.** Figure 6(c–d) shows that, on ORL, full-risk ranking preferentially suppresses cross-class edge mass compared with individual support/dispersion proxies, while actual-risk attenuation reduces the cross-/same-class mass ratio more effectively than shuffled-risk and uniform-scaling controls. Moreover, Table 3 shows that, under fixed adaptive splitting, risk-aware connection improves ALOI ACC/ARI from 90.60/87.25 to 96.85/93.85 and ORL from 82.85/75.02 to 83.22/75.64. These observations support the practical utility of the risk proxy on the evaluated datasets, without establishing universal calibration or robustness to arbitrary view corruption. We will explicitly distinguish **the general mixed-atom exposure identity, the conditional purity-based guarantee, and the empirical effectiveness of risk estimation**, while acknowledging possible failures under severe view inconsistency or systematically misleading cross-view agreement. --- ### References [1] Guo et al. **Robust Contrastive Multi-view Clustering against Dual Noisy Correspondence.** NeurIPS 2024. https://proceedings.neurips.cc/paper_files/paper/2024/hash/dbe81b08f7dc4dd8b43bc62dedfd9662-Abstract-Conference.html [2] Zhou et al. **Uncover Underlying Correspondence for Robust Multi-view Clustering.** ICLR 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/8c1d92835eb4e601f396c97ec60439fe-Abstract-Conference.html [3] Yang et al. **Automatically Identify and Rectify: Robust Deep Contrastive Multi-view Clustering in Noisy Scenarios.** ICML 2025. https://proceedings.mlr.press/v267/yang25am.html [4] Ren et al. **Multi-View Graph Clustering via Node-Guided Contrastive Encoding.** ICML 2025. https://proceedings.mlr.press/v267/ren25a.html Response to Comment 4: Dataset Count Inconsistency We thank the reviewer for identifying this inconsistency and apologize for the unclear presentation. The six datasets mentioned in the abstract correspond to the primary baseline comparisons in Table 2 (BBCSport, DHA, Mfeat, NGs, ORL, and UCI), whereas Section 4.1 additionally includes ALOI and UCI-Digits, which are used in ablation and visualization analyses. We will explicitly distinguish these experimental scopes in the abstract and Section 4.1, while clarifying the broader comparisons in Appendix H, to ensure consistent dataset reporting throughout the manuscript.
Reviewer_pxVyOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

The paper proposes GBASC, a granular-ball reliability learning framework for deep multi-view clustering. It identifies two sources of structural errors: mixed samples within granular balls and incorrect assignments between granular balls. GBASC addresses them through adaptive splitting (GBAS) and risk-aware connection (GBAC). The resulting structures also guide representation learning. Experiments on multiple benchmark datasets show improved clustering performance.

Critical strengths
  1. The paper distinguishes within-atom mixing errors from between-atom assignment errors, providing a clear motivation for the two-stage framework.
  2. GBAS and GBAC address complementary structural problems. Their combination provides a systematic way to refine local units and determine global assignments.
  3. In addition to standard clustering metrics, the paper investigates boundary quality, risk-based edge attenuation, and the contribution of connection selection. These analyses help explain how the proposed modules affect the clustering process.
Critical weaknesses
  1. The error decomposition provides a useful conceptual framework, but the paper should more clearly distinguish its original algorithmic contributions from general properties of partition refinement and graph-based clustering. The relationship to existing granular-ball clustering methods also requires a more direct comparison.
  2. The method combines several scores, thresholds, and candidate-selection rules. Their interactions and sensitivity are not sufficiently clarified.
  3. The manuscript contains presentation inconsistencies.
    • Algorithm 1 refers to a warm-up protocol and an initial structural representation HH, but the warm-up loss and the precise choice of raw, preprocessed, or learned representations are not defined in the main text.
    • Section 4.1 lists eight datasets, while the main comparison in Table 2 covers six; Appendix H introduces twelve configurations and a different baseline set.
    • Table 1 uses “E-GNF” in the row label while the surrounding text and reference list use “EBMGC-GNF,” and Appendix H introduces several abbreviated baselines without full names or local citations.
Questions for authors

Please address the questions raised in the Critical Weaknesses.

5 comments
AuthorsOpenReview ↗

Answer to Question 1

We sincerely thank the reviewer for this insightful comment. We agree that the manuscript should better distinguish general properties of partition refinement and graph clustering from our specific algorithmic contributions, and more explicitly compare GBASC with existing granular-ball methods. We will clarify these distinctions in the revision.

1. General properties vs. algorithmic novelty.

We acknowledge that the pairwise-error decomposition in Eq. (1) is an exact identity under atom-constant assignments, and the non-increasing within-atom error in Theorem 1 is a general property of proper refinement. We do not claim these properties themselves as novel algorithms. Rather, they motivate two distinct reliability decisions: which samples should remain indivisible (GBAS), and which resulting atoms should share a global cluster (GBAC). Our contribution lies in the concrete label-free mechanisms for controlling these two transitions, not in the underlying mathematical properties alone.

2. Differences from graph-based clustering.

Existing graph clustering methods, including Pan and Kang (NeurIPS 2021), Tan et al. (CVPR 2023), Ling et al. (AAAI 2023), and HoLe (CIKM 2023), have established graph construction, refinement, and structure-aware clustering. We acknowledge these foundations.

However, graph refinement over fixed atoms cannot eliminate within-atom mixing errors, whereas GBASC explicitly revises the atomic partition before global assignment.

Specifically, GBAS evaluates candidate boundaries through compactness improvement, neighborhood fracture, cross-view agreement, perturbation stability, and child balance. Importantly, boundary selection is separated from split acceptance, which requires either risk–complexity improvement or gated structural evidence. Thus, GBAS is not simply recursive graph partitioning.

GBAC separately aggregates positive association mass and uncertainty mass into an atom-level quotient graph, then calibrates connections using their relative risk. Unlike relying on a single graph-partitioning result, it evaluates five complete partition candidates under cut, mass, stability, and trust constraints. The selected reliable relations further guide representation learning. Our novelty therefore lies in reliability-controlled atomic refinement and global partition selection, rather than graph refinement or spectral clustering alone.

3. Differences from granular-ball clustering.

We acknowledge that granular-ball construction, adaptive refinement, and aggregation have been studied previously.

  • MGBCC (AAAI 2025) primarily exploits within-view and cross-view granular-ball associations for contrastive representation learning. In contrast, GBASC explicitly decides whether existing atoms should be split and subsequently determines their global assignments through risk-calibrated connections.

  • GFSAF (AAAI 2026) emphasizes granular-ball fuzzy contrastive learning, complementary representation extraction, and attention-based view fusion. GBASC instead focuses on the reliability of local atom formation and global atom connections, using explicit splitting and partition-selection criteria.

  • GBCT (IEEE TNNLS 2025) already incorporates adaptive granular-ball refinement and aggregation. We therefore do not claim these operations as novel. Our distinction is the combination of cross-view boundary evidence, explicit split-acceptance mechanisms, uncertainty-calibrated atom connections, and constrained complete-partition selection.

4. Experimental evidence and revisions.

Our manuscript already includes direct comparisons with MGBCC and GFSAF in Table 2, showing favorable clustering performance across the main benchmarks. Table 3 examines the complementary effects of splitting and connection, while Appendix G.3 compares boundary quality under matched child-block sizes, and Appendices G.4–G.5 examine reconnection and complete-partition selection. These results support our design, although overall performance improvements alone do not establish the novelty of each component.

In the revision, we will (i) explicitly distinguish general theoretical properties from algorithm-specific contributions; (ii) expand Related Work with a direct mechanism-level comparison against graph-based and granular-ball methods; (iii) clarify the boundary-selection, split-acceptance, risk-calibration, and partition-selection procedures; and (iv) strengthen the discussion of existing comparative and ablation results. We will also clarify that refinement monotonicity does not imply unconditional improvement in total clustering error.

We sincerely appreciate the reviewer's suggestion, which will help us present our contributions more precisely and transparently.

AuthorsOpenReview ↗

Answer to Question 2

We thank the reviewer for highlighting the need to clarify the interactions and sensitivity of our decision rules. The scores and thresholds address distinct failure modes, rather than being redundant heuristics.

1. Rationale for the decision rules.

In GBAS, candidate evidence ranks splits using compactness, graph separation, cross-view agreement, and stability; acceptance is a separate decision. The cost criterion controls fragmentation, whereas the structure criterion admits globally supported boundaries that local penalties might reject. The dual gate combines these complementary routes, followed by guide-based refinement.

In GBAC, risk calibration attenuates uncertain inter-atom edges, candidate selection evaluates alternative partitions, and trust constraints reject potentially harmful reassignment despite an improved aggregate score, reverting to the base partition when needed.

2. Fine-grained ablation analysis.

To investigate these interactions, we evaluated the following variants under the same protocol (mean ± std; ACC in %, lower pair error is better):

GroupVariantORL ACC ↑DHA ACC ↑ALOI ACC ↑ALOI Pair Error ↓
GBASA1: Cost only81.60 ± 1.3483.20 ± 0.7693.90 ± 0.810.19 ± 0.03
A2: Structure only80.90 ± 1.5682.60 ± 1.0194.20 ± 0.740.18 ± 0.03
A3: Dual gate82.50 ± 1.1985.10 ± 0.4895.50 ± 0.510.14 ± 0.02
A4: Full GBAS83.22 ± 1.1086.44 ± 0.2096.85 ± 0.340.12 ± 0.02
GBACB1: No Risk + Base81.90 ± 1.5283.80 ± 0.9490.20 ± 0.830.29 ± 0.05
B2: No Risk + Selection82.40 ± 1.2484.80 ± 0.6890.80 ± 0.710.27 ± 0.05
B3: Risk + Base82.60 ± 1.1685.50 ± 0.5395.50 ± 0.490.16 ± 0.03
B4: Risk + Selection, No Trust82.50 ± 1.3885.80 ± 0.6395.10 ± 0.670.18 ± 0.04
B5: Full GBAC83.22 ± 1.1086.44 ± 0.2096.85 ± 0.340.12 ± 0.02

The results reveal three important interactions:

(i) GBAS: A3 outperforms either single criterion (A1/A2) across all ACC metrics, while A4 further reduces ALOI pair error to 0.12, demonstrating the complementary benefits of dual-gate acceptance and final refinement.

(ii) GBAC risk calibration: B1→B3 improves ALOI ACC by 5.30 percentage points and reduces pair error from 0.29 to 0.16, whereas selection without risk (B1→B2) yields only a 0.60-point ACC improvement. This indicates that candidate selection cannot compensate for unreliable graph connections.

(iii) GBAC trust constraints: Compared with B3, unconstrained selection (B4) reduces ALOI ACC (95.50→95.10) and increases pair error (0.16→0.18). In contrast, trust-constrained selection (B5) achieves 96.85 ACC and 0.12 pair error, confirming the importance of conservative candidate acceptance.

These findings demonstrate that the components address complementary structural risks rather than merely increasing model complexity.

3. Parameter sensitivity and manuscript revisions.

We acknowledge that these ablations clarify component interactions but do not fully characterize sensitivity to continuous hyperparameters. In the revised manuscript, we will clarify parameter roles and defaults, supplement sensitivity analyses of GBAS evidence weights and complexity penalties, GBAC risk attenuation, and trust thresholds, and distinguish modeling parameters from numerical tolerances.

We will also revise Sections 3.2–3.3 to clarify candidate scoring, acceptance, and fallback mechanisms, and expand Section 4.3 with these ablations and their implications.

We sincerely appreciate this constructive suggestion, which has helped us strengthen both the methodological explanation and experimental validation of GBASC.

AuthorsOpenReview ↗

Answer to Question 3

We sincerely thank the reviewer for carefully identifying these presentation inconsistencies. We apologize for the lack of clarity and consistency in the current manuscript. We fully agree that these issues should be addressed to improve the paper's readability, transparency, and reproducibility. We will carefully revise the manuscript to resolve all three concerns.

1. Warm-up protocol and initial representations.

We acknowledge that Algorithm 1 does not sufficiently specify the warm-up objective or the representations used for initial structural construction. We will explicitly define the warm-up loss and its coefficients, clarify whether raw, preprocessed, or learned representations are used during initial structure construction, and describe how representations are updated in subsequent stages. We will verify these details against our implementation and provide a consistent description in the main text, Algorithm 1, and the corresponding appendices.

2. Dataset counts and baseline configurations.

We apologize for the ambiguity in the experimental presentation. The eight datasets listed in Section 4.1 reflect the combined scope of our comparative experiments, ablation studies, and structural evaluations, whereas Table 2 presents the primary baseline comparisons on six datasets. The remaining datasets, including ALOI and UCI-Digits, are covered by other evaluations. Appendix H further extends the evaluation to twelve benchmark configurations, some of which share the same underlying data source but use different view configurations.

We acknowledge that these distinctions and the reasons for using different baseline sets were not sufficiently explained. We will explicitly distinguish datasets from experimental configurations, clarify the scope of each experiment, and explain the baseline selection criteria to ensure transparent and consistent reporting.

3. Method naming and citation consistency.

We apologize for the inconsistent abbreviation "E-GNF" in Table 1. The correct designation is EBMGC-GNF (Efficient Balanced Multi-view Graph Clustering via Good Neighbor Fusion). We will correct this inconsistency and provide the full names and corresponding citations for abbreviated baselines in Appendix H. We will also systematically check all method names, table captions, and references throughout the manuscript.

Commitment to revision.

We sincerely appreciate the reviewer's careful examination and constructive suggestions. These comments have helped us identify important omissions in our presentation. We fully respect the reviewer's concerns and commit to addressing all the issues raised in the revised manuscript and ensuring that every correction is incorporated before publication. We believe these revisions will substantially improve the clarity, consistency, and reproducibility of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.