acceptodds
Under review as a conference paper at ICLR 2027

Mapping-based Adaptively Weighted Ensemble Clustering

Abstract

Ensemble clustering fuses multiple base clusterings into a partition that is more accurate and more robust than any single one. Existing methods, however, fix the weights before the fusion and apply them at a single granularity, split representation learning and partition solving into two stages, are prone to trivial solutions, and rely on the expensive co-association matrix. We therefore propose Mapping-based Adaptively Weighted Ensemble Clustering (MAWEC). A continuous weight rates each base clustering and a mapping matrix merges each base cluster into a target cluster, so the two granularities are optimized jointly. Consistency learning, structured learning and weight learning are coupled through one soft label matrix, whose row-wise maximum gives the labels directly. A structural regularizer promotes row sparsity and column balance, and a weight regularizer keeps the weights from collapsing, so three trivial solutions are excluded. Since the consensus is carried by the association matrix, the fusion is linear in the number of samples. Experiments on eight real-world datasets show that MAWEC attains the highest ACC on all of them and the shortest running time.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
ensemble clustering, mapping matrix, weight learning, soft label matrix, trivial solutions, co-association matrix

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

This paper proposes a way to merge several base clusterings (Eq. (5)). The main new perspective is to introduce a learnable weight vector α\alpha. Then the authors developed an optimization procedure to solve the problem (5).

Critical strengths

None

Critical weaknesses

Overall, the contribution appears primarily incremental. The method combines cluster-label mapping, adaptive partitiion weights, and regularization within a joint framework. There is no substantial conceptual advance from combining these components.

The paper claims "two granularity weighting", but the cluster level mapping Z merely assigns each base cluster to a target cluster. It does not weight or suppress unreliable clusters. The claim that both unreliable partitiion and individual clusters are downweighted is therefore not suppored.

The relationship in Eq (10) is straightforward and elementray, no need to have a theorem. Moreover, the regularizer does not establish the claimed exclusion of trivial solutions.

Each base cluster is mapped as a whole to one target cluter. If it contains samples from several true classes, that partition's contribution cannot separate them. The abstract says "a partition that is more accurate and more robust than any single one", which is quite non-rigious.

Monotonic decrease and a lower bound establish convergence of the objective sequence. They do not establish convergence to a useful clustering, a global optimum, or even convergence of all optimization variables.

Questions for authors

For hard memberships, the structural term explicitly favors equally sized clusters. The paper needs controlled experiments with unequal cluster sizes and rare groups to establish when this prefrerence helps or harms performance.

Ablations, generator comparisons, and parameter sensitivity are reported only on Pendigits. There are no controlled tests that progressively introduce unreliable partitions, correlated error, or class imbalance.

Initialization uses base-cluster centers computed in the original feature space. Consequently, the implemented procedure requires more than the base clustering lables, contrary to the motivation that consensus uses only labels.

Reviewer_Txf8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Mapping-based Adaptively Weighted Ensemble Clustering (MAWEC). It jointly optimizes partition-level weights, hard mappings from base clusters to consensus clusters, and a soft consensus assignment matrix. A centered-variance regularizer encourages discrete and balanced assignments, while an L2 penalty discourages concentrated partition weights. Optimization uses majorization–minimization, coordinate descent, and projected gradient updates. Experiments on eight benchmarks report the highest mean ACC and shortest consensus-stage runtime among the evaluated methods. The appendices provide proofs, algorithmic details, dataset statistics, and ablation objectives.

Critical strengths
  • The objective provides a compact formulation connecting consensus assignments, partition weights, and cluster mappings.

  • Sparse association structures and cached coordinate updates offer a plausible route to efficient consensus computation.

  • The structural upper bound and the objective-value convergence argument are largely sound within their stated scope.

  • The reported evaluation includes eight datasets, four clustering metrics, and twenty repetitions. The authors also acknowledge cases where competing methods outperform MAWEC.

Critical weaknesses
  1. Novelty over PFEC is insufficiently established: the closest cited method is neither adequately distinguished nor directly compared.

  2. The cluster mapping changes destinations but provides no independent magnitude for downweighting unreliable clusters within a partition.

  3. MAWEC-w/oB starts at a global optimum under the prescribed initialization, so it cannot demonstrate the claimed optimization-induced degeneration.

  4. The regularization analysis does not guarantee avoidance of severe imbalance; a reachable 99/1 fixed point remains possible.

Questions for authors
  1. What is new relative to PFEC? PFEC already dynamically weights both base clusterings and clusters and avoids additional post-processing. Please identify MAWEC’s substantive advance over this precedent, distinguishing hard mapping and any efficiency gains from the shared claims of weighting at two levels and joint optimization. Why is PFEC absent from the experiments? Please provide a direct comparison using matched ensembles, initialization information, tuning budgets, and runtime protocols.

  2. How are unreliable base clusters independently downweighted? Each row of ZkZ_k is one-hot, so ∑ℓαk(Zk)jℓ=αk\sum_{\ell} \alpha_k(Z_k)_{j\ell} = \alpha_k. Every base cluster in partition kk therefore retains total contribution αk\alpha_k. How can a low-quality cluster receive a smaller reliability magnitude independently of other clusters in that partition? Please distinguish changing a contribution's destination from reducing its magnitude, and identify the parameter or mechanism that implements the claimed suppression. A controlled corruption experiment and direct diagnostics of per-cluster contributions would help establish this mechanism.

  3. How can MAWEC-w/oB demonstrate optimization-induced degeneration? With λ=0\lambda=0, the objective is ∥Y−AZ∥F2+γ∥a∥22\|Y-AZ\|_F^2+\gamma\|a\|_2^2. The prescribed initialization Y0=A0Z0Y^0=A^0Z^0 and a0=1/Ta^0=\mathbf{1}/T already gives zero fidelity error and the minimum weight penalty γ/T\gamma/T. Under the stated updates and tie-preserving rule, the variant therefore remains at a global optimum. Do you agree, or which update changes this state? Please report initialization performance and objective and assignment trajectories, and reconcile this with the claim that the variant "slides toward trivial solutions."

  4. What excludes severely imbalanced fixed points in the actual algorithm? Consider 99 coincident samples at one location and one sample at another, with every base clustering using two clusters, an allowed outcome of the stated generation rule. The centroid initialization yields the same 99/1 hard partition in every mapped base clustering. Let BB denote this partition and pp its cluster proportions. At Y=AZ=BY=AZ=B, the MM input for row II is Bi⋅+λ(Bi⋅−p)B_{i\cdot}+\lambda(B_{i\cdot}-p), whose simplex projection remains BiB_i for every λ>0\lambda>0. The mapping has zero fidelity error, and the uniform weights remain optimal. How is this reachable fixed point with two nonempty clusters excluded? Please distinguish Theorem 1's extremal property of the regularizer from guarantees for the alternating algorithm, and clarify the scope of the exclusion claim.

Reviewer_fZPpOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MAWEC, an ensemble clustering method that jointly learns base-clustering weights, cluster mappings, and consensus labels. It introduces regularization terms to encourage discriminative and balanced assignments and prevent weight collapse. An alternating optimization algorithm is developed, with theoretical analyses of convergence and computational complexity. Experiments on eight datasets demonstrate competitive clustering performance and efficiency.

Critical strengths
  1. The paper formulates ensemble clustering as a unified optimization problem that jointly learns base-clustering weights, cluster mappings, and consensus labels, avoiding a strictly decoupled fusion pipeline.

  2. The mapping-based representation avoids explicitly constructing an (n\times n) co-association matrix, offering potential computational and memory advantages for scalable ensemble clustering.

  3. The method provides theoretical analyses of the regularization properties and monotonic objective descent, together with an efficient alternating optimization algorithm.

  4. Experiments on eight datasets, including ablation studies and parameter sensitivity analyses, provide empirical evidence for the effectiveness and efficiency of the proposed method.

Critical weaknesses
  1. I appreciate the authors' efforts in developing this method. However, I have significant concerns about its novelty. Many existing works explore closely related ideas, even if their objective functions are not exactly the same as the one proposed in this paper. Examples include [1] and [2]. Therefore, I am not convinced that this work provides a sufficiently novel insight or a substantial methodological contribution to the ensemble clustering literature.
  2. The authors state that the weight vector alpha and the mapping matrices Z_k constitute the two granularities of the proposed method. However, these two components serve fundamentally different purposes: alpha controls the relative contributions of different base clusterings, whereas Z_k maps base clusters to the target clusters to obtain the final clustering result. The authors appear to conflate adaptive weighting with cluster mapping, and the claimed dual-granularity perspective is therefore insufficiently justified. A clearer conceptual distinction and stronger motivation are needed.
  3. The balance regularizer encourages clusters to have similar sizes. However, real-world data often exhibit highly imbalanced cluster distributions. Although the authors introduce this regularizer to avoid degenerate solutions, it remains unclear whether imposing such a balance prior is appropriate in general. The authors should justify this modeling assumption and evaluate its impact on datasets with different degrees of cluster-size imbalance.
  4. In Eqs. (14)–(15), the sign in the squared norm is inconsistent with the stated update of Y. In addition, Appendix B appears to confuse the complexity of a single row update, O(m+c), with that of a full coordinate-descent epoch, O(m(m+c)). These issues should be corrected or clarified.
  5. Adaptive weighting has been extensively studied in ensemble clustering, and I do not find the proposed modeling perspective particularly novel. In contrast, computational efficiency may represent a more compelling contribution. In my own experiments, I have found LWGP to be an efficient algorithm that also achieves strong clustering performance. The results reported in this paper suggest that the proposed method can outperform LWGP in both runtime and clustering quality. To substantiate this advantage and better demonstrate the practical value of the work, I strongly encourage the authors to evaluate their method on substantially larger datasets, potentially containing one million or even ten million samples, and compare it against other efficient ensemble clustering methods. Such experiments would provide more convincing evidence of the method's scalability and practical significance.
  6. The authors have not released their code, making it difficult to independently verify the reported results. Furthermore, although the paper describes the hyperparameter selection procedure for the proposed method, it is unclear whether the authors tuned its hyperparameters separately for each dataset and, if so, whether comparable tuning was performed for the baseline methods. A fair comparison requires transparent reporting of the hyperparameter configurations and selection protocols for all methods. The authors should clarify these details and provide code to facilitate reproducibility.
  7. The authors should make a greater effort to keep up with the latest research developments and incorporate relevant recent studies into both the related work and experimental comparisons. For example, the authors should incorporate and discuss more relevant studies published in 2026.
Questions for authors

See the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.