acceptodds
Under review as a conference paper at ICLR 2027

Discrete Co-Evolution of Multi-Resolution Partition Hypergraphs for Multi-View Clustering

Abstract

Multi-view clustering (MVC) aims to exploit complementary information across different views to uncover a reliable consensus partition. However, existing methods often treat view-specific structures as fixed inputs to consensus learning, resulting in one-way information flow in which local structures guide the consensus but cannot be refined by it. Moreover, intrinsically categorical assignments are frequently optimized through continuous relaxation followed by post-hoc discretization, creating a mismatch between optimization and the final solution. To address these limitations, we propose DCoE, a discrete co-evolution framework for multi-view clustering. DCoE treats resolution-specific partitions as learnable discrete latent structures and jointly refines them with a common discrete consensus. These partitions induce multi-resolution hypergraphs that capture complementary group-wise structures, with adaptive weighting integrating their contributions. All partition variables are optimized directly in the discrete domain through exact reassignment gains, avoiding relaxation and post-hoc discretization while guaranteeing monotonic objective improvement. We further analyze a faithful row-simplex relaxation and show that, under a verifiable structural–spectral condition, every local maximizer is one-hot. A resolution–balance guarantee further connects this categorical geometry to the learned multi-resolution structures. Sparse anchor relations and group-level sufficient statistics enable matrix-free optimization with linear storage in the number of samples. Extensive experiments demonstrate the effectiveness and efficiency of DCoE.

Primary area:
optimization
Keywords:
multi-view clustering, optimization, multi-resolution clustering, partition hypergraph, anchor-based clustering, consensus learning

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 4 reviews, above 27% of the others.
Reviewer_p2ezOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a discrete co-evolution framework for multi-view clustering (DCoE). The method jointly refines multi-resolution view partitions and a discrete consensus, with theoretical conditions for one-hot local maximizers of the relaxed consensus problem. Experiments show competitive results against baselines.

Critical strengths
  1. Discrete updates use exact reassignment gains, while group statistics avoid constructing large similarity matrices. This improves computational efficiency.
  2. The paper derives sufficient conditions for one-hot local maximizers of the relaxed consensus problem and verifies these conditions experimentally.
  3. DCoE achieves the best ACC on eleven datasets against thirteen baselines. Ablation studies and runtime comparisons further support its effectiveness and efficiency.
Critical weaknesses
  • The mutual refinement of view partitions and consensus in DCoE has also been explored in CMVC [1]. What is the main difference between the two methods?
  • The partition hypergraph consistency term in Eq. (3) is closely related to category utility, with an additional term depending on consensus cluster sizes. What is the benefit of using the hypergraph formulation?
  • The relaxed assignments in Table 5 are already close to one-hot, while DCoE and Relaxed-F use different optimization methods. It is unclear whether the improvement in Table 3 comes from direct discrete optimization or the difference between the solvers.
  • Table 3 shows improvements over Fixed-G, but it is unclear whether consensus feedback can correct noisy or inaccurate view partitions. Could the authors provide more evidence of this process?
  • The ablation study does not separately examine the contributions of multiple resolutions and adaptive weights. Comparisons with a single resolution and fixed weights would be useful.
  • The running times of DCoE in Table 4 are identical to its optimization times in Appendix A.6 (Table 6). Does the reported time include initialization and preprocessing?

ref: [1] Hongfu Liu and Yun Fu. 2018. Consensus Guided Multi-View Clustering. ACM Trans. Knowl. Discov. Data 12, 4, Article 42 (August 2018), 21 pages. https://doi.org/10.1145/3182384

Questions for authors

See the weakness.

Reviewer_b5ugOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a discrete co-evolution framework (DCoE) for multi-view clustering, which jointly optimizes multi-resolution view-specific partitions and a shared consensus partition. The method employs adaptively weighted partition hypergraphs to capture multi-granularity structures and directly optimizes discrete assignments through exact reassignment gains, avoiding continuous relaxation and post-hoc discretization. Theoretical analysis and extensive experiments demonstrate its effectiveness and computational efficiency.

Critical strengths

(1)The integration of multi-resolution partition hypergraphs and discrete co-evolution enables view-specific structures to be dynamically refined by the global consensus. (2)The paper establishes monotonic convergence of the discrete optimization procedure and derives sufficient conditions under which the continuous relaxation admits one-hot local maximizers. (3)The method achieves strong clustering performance on 11 benchmark datasets. Ablation studies, parameter sensitivity analyses, and runtime comparisons further demonstrate its effectiveness.

Critical weaknesses

(1)Although Table 3 demonstrates the benefits of partition hypergraphs and co-evolution, direct comparisons between single-resolution and multi-resolution structures are missing. The contribution of adaptive resolution weighting is also not independently evaluated. Therefore, the individual benefits of these two designs remain unclear. (2)Appendix A.5 states that all hyperparameters are selected through grid search but does not specify the selection criterion. Since multi-view clustering is an unsupervised task, it is important to clarify whether ground-truth labels were used for hyperparameter tuning and whether competing methods were tuned under comparable conditions. (3)The DCoE runtimes in Table 4 are identical to the optimization-only times reported in Appendix Table 6. It is therefore unclear whether preprocessing, anchor graph construction, and initialization costs are included. Additionally, although the results are averaged over five independent runs, standard deviations are not reported, making it difficult to assess the stability of smaller performance improvements. (5) The learned partitions Gl(v)G_l^{(v)} have no explicit constraints preventing severe group-size imbalance or redundant partitions across resolutions. The authors should report group-size distributions, inter-resolution similarity, and learned resolution weights. (6) The exact objectives and update rules of variants such as “w/o Partition Hypergraph,” Fixed-G, and the relaxed models are unclear. The authors should specify which terms and variables are removed, fixed, or relaxed in each experiment.

Questions for authors

1)Could the authors provide ablation studies using a single resolution and fixed resolution weights to further validate the necessity of these components? (2)Multi-granularity partition learning and discrete optimization have been explored in previous multi-view clustering methods. Could the authors further clarify the key methodological differences between DCoE and closely related approaches, particularly the unique contribution of the proposed co-evolution mechanism? (3)Table 5 shows that the relaxed variants already produce nearly one-hot assignments, yet their clustering performance remains inferior to DCoE in Table 3. Could the authors clarify whether this improvement primarily comes from discrete optimization itself or from differences in optimization strategies and convergence behavior?

Reviewer_jzHmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author proposed a discrete co-evolutionary approach for multi-resolution partitioned hypergraphs in multi-view clustering, which utilizes the number of clusters at different scales to capture different local structures and then explores global attributes. The clustering is completed by optimizing mutually through the resolution matrix and consensus matrix. The results are demonstrated in certain experiments.

Critical strengths

(1) The author proposed a relatively effective method and extracted some theories. The validity of this method was also verified through experiments. (2) The author incorporates the concept of ensemble learning into it, integrating the dual advantages of multi-view clustering and ensemble clustering.

Critical weaknesses

(1) Since the hyperparameters control the sparsity of F, for the objective function, the hyperparameters are extremely sensitive. However, the author merely made one parameter very small, resulting in the components not functioning effectively. The parameter range of another hyperparameter remained almost unchanged. (2) For such ideas, they are quite common in ensemble clustering or multi-view clustering. However, the author did not provide a good application scenario and did not fundamentally consider the issue.

Questions for authors

(1) Could you provide a more comprehensive parameter analysis? (2) The author should have expressed more clearly the contribution of this method to multi-view clustering, rather than merely presenting a common but completely different approach and providing some theoretical explanations.

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

This paper proposes a multi-view clustering framework called DCoE to jointly learn a discrete consensus partition and view-specific partitions at multiple resolutions. It treats multi‑resolution discrete partitions as latent learnable variables that co‑evolve bidirectionally with a global discrete consensus partition. Then, these partitions form partition‑induced hypergraphs, whose contributions are adaptively weighted. It uses sparse anchor relations and group-level statistics to avoid constructing dense sample-wise matrices. The whole optimization runs directly over discrete assignments using exact reassignment gains, avoiding post‑processing. Besides, it analyzes a row-simplex relaxation, deriving conditions under which its local maximizers are one-hot.

Critical strengths
  1. The designed framework enables bidirectional refinement of view-specific structures and consensus.

  2. The empirical evaluation is comprehensive.

Critical weaknesses
  1. The benefit of using multiple resolutions is not fully isolated in the ablations.

  2. The hyperparameter-selection protocol is not sufficiently specified.

  3. The objective sums the view contributions without an explicit mechanism for estimating the reliability of each view.

  4. The benchmarks contain mostly well‑balanced informative views. There are no experiments with synthetically injected low‑quality views, such as noisy views.

Questions for authors
  1. Are the reported results averaged over five independent runs for every method and dataset? Could the authors report standard deviations?

  2. Is the one-hot guarantee intended to apply only to the relaxed consensus subproblem with the other variables fixed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.