acceptodds
Under review as a conference paper at ICLR 2027

A General Domain-Adaptive Clustering Approach via Cross-Dataset Prototype Alignment

Abstract

Deep clustering typically learns cluster structures independently for each dataset, limiting its ability to exploit shared patterns across heterogeneous domains. We propose CPC, a general domain-adaptive clustering approach based on cross-dataset prototype alignment, which jointly learns a shared prototype space across multiple domains. By allowing different domains to activate shared prototypes, CPC establishes cross-dataset structural correspondence and discovers representative prototypes that serve as compact cluster anchors. Experiments on 11 heterogeneous real-world datasets show that CPC achieves the best or tied-best accuracy on 9 datasets and substantially outperforms existing clustering methods, with improvements of over 40 percentage points on challenging datasets such as Primary Tumor. These results demonstrate the effectiveness of shared prototype spaces for domain-adaptive clustering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep Clustering, Domain-Adaptive, Cross-Dataset Prototype Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_j8SfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a clustering framework that establishes shared prototypes for heterogeneous datasets, named CPC. Specifically, Cross-Dataset Prototype Alignment first projects heterogeneous datasets into a common space and then learns a shared prototype codebook. After that, Representative Prototype Discovery constructs interactions among activated prototypes and selects representative ones as cluster centers. Experiments on various clustering datasets demonstrate the effectiveness of the proposed CPC.

Critical strengths
  1. The paper addresses an important problem in clustering, namely learning common prototypes for heterogeneous datasets.

  2. Experiments on various datasets demonstrate the effectiveness of the proposed CPC.

  3. The paper is well structured and clearly written.

Critical weaknesses
  1. My main concern is the motivation of this work. To be honest, in traditional classification tasks, I think the weight vectors of the linear classification layer could be considered cluster centers, which are first learned from large-scale data and then applied to various downstream datasets. Therefore, I do not think that explicitly constructing shared prototypes for heterogeneous datasets is an entirely new setting. The authors should discuss this and address my concern.

  2. As discussed in Weakness 1, shared prototypes could potentially be obtained through the linear layer. I wonder why the proposed method requires more complex components to establish shared prototypes, such as dataset-specific adapters and a shared Transformer encoder. The authors should provide theoretical or empirical evidence to verify the effectiveness of these components.

  3. The proposed method appears to be incremental. Specifically, the authors should clarify the differences between the four objectives in the first stage and existing representation learning objectives, such as the commitment loss in VQ-VAE and BYOL, the contrastive loss in SimCLR, and the entropy regularization in Contrastive Clustering. Given that the novelty of the setting is questionable, it is difficult for me to recognize the contribution of this work unless the authors could clarify the novelty of the proposed methods.

  4. Following Weakness 3, the proposed CPC integrates seven loss terms, which remains unclear that whether each loss could work with a distinct and necessary purpose. In particular, the paper does not provide any ablation study to clarify the individual contribution of each term.

  5. I also have concerns about the experiments. It is recommended to evaluate the proposed CPC on widely used benchmarks such as ImageNet or CIFAR-100. To be honest, I am quiet unfamiliar with the datasets in this paper, thus I think that the authors should clarify the sources of these datasets and justify whether they could searve as benchmarks for evaluating deep clustering methods.

Questions for authors

See the weaknesses.

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

This paper proposes CPC, a two-stage framework for jointly learning clustering representations across heterogeneous datasets. In the first stage, Cross-Dataset Prototype Alignment (CPA) uses dataset-specific adapters, a shared Transformer encoder, and a vector-quantized codebook to learn a common prototype space. In the second stage, Representative Prototype Discovery (RPD) uses attention-based interactions and representative selection to derive dataset-specific cluster assignments from activated prototypes. The paper reports experiments on 11 UCI datasets against seven clustering baselines, together with ablation studies and codebook-size sensitivity analysis.

Critical strengths

Leveraging shared information across heterogeneous datasets for clustering is a meaningful research direction.

Critical weaknesses
  1. Serious Concerns About Experimental Reliability. The striking discrepancies between ACC and NMI for some baselines on SF (Table 2) prompted me to examine the released data and evaluation code. While severe class imbalance helps explain this pattern (which the paper does not discuss), the inspection revealed two additional issues:
  • Inconsistent dataset statistics. Table 1 reports six clusters for SF, whereas the released SF dataset contains only three ground-truth classes. This raises questions about the dataset and ground-truth labels actually used in the experiments.
  • Incorrect evaluation protocol. The released evaluation code uses ground-truth labels to assign classes to prototypes and propagates these labels to generate predictions. It then computes ACC, NMI, ARI, and F1 at the prototype level rather than the original sample level. This differs fundamentally from the standard unsupervised clustering evaluation described in the paper, where ground-truth labels should only be used after clustering to assess the resulting sample assignments.

These issues seriously undermine the reliability of the reported results and the claimed performance advantages over the baselines.

  1. Insufficient Justification and Validation of the Core Methodology. The paper assumes that heterogeneous datasets contain reusable structural patterns, yet neither clearly defines these patterns nor provides direct evidence that such structures are shared. The proposed CPA mainly relies on a shared encoder and prototype codebook, without demonstrating how meaningful cross-dataset structural correspondences are established. Improvements in conventional clustering metrics alone do not validate this central hypothesis. Furthermore, the methodology largely combines existing techniques without sufficiently establishing their individual contributions or advantages over related multi-task clustering approaches. In particular, removing RPD improves all 4 metrics on 6 of the 11 datasets (Tables 2 and 5), casting doubt on its effectiveness. The lack of direct structural diagnostics, evaluation on unseen domains, and sufficient sensitivity analysis for key representative-selection parameters further limits the evidence supporting the proposed framework.
Questions for authors
  1. Could the authors explain why Table 1 reports 6 classes for SF while the released dataset contains only 3? More importantly, why does the evaluation code use ground-truth-based prototype labeling and prototype-level metrics instead of the standard sample-level clustering evaluation described in the paper? Could the authors identify the exact evaluation procedure used to generate Table 2 and provide reproducible sample-level clustering results consistent with the protocol described in the paper?

  2. What specific structural patterns are expected to be shared across heterogeneous datasets, and how does CPA actually capture them? Could the authors provide direct evidence of cross-dataset prototype alignment and clarify the methodological contribution beyond existing multi-task clustering approaches?

  3. Why does removing RPD improve all 4 metrics on 6 datasets? Under what conditions does RPD provide benefits, and how sensitive are its results to the representative-selection ratio, threshold, and other key hyperparameters?

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

This manuscript addresses the problem that deep clustering typically learns cluster structures independently for each dataset and cannot exploit patterns shared across heterogeneous datasets. It proposes a "domain-adaptive clustering" setting in which multiple heterogeneous datasets jointly learn a shared prototype space without assuming a source–target relationship. The proposed CPC method follows a two-stage Learn–Select framework. In the first stage, Cross-Dataset Prototype Alignment (CPA) maps heterogeneous features into a common latent space via dataset-specific adapters and a shared Transformer encoder, and learns a shared prototype codebook through vector quantization, so that different datasets activating the same prototype constitutes an implicit prototype-level alignment. In the second stage, Representative Prototype Discovery (RPD) operates on the fixed prototype space, models interactions among each dataset's activated prototypes with self-attention, learns a "representative probability" for each prototype, selects a small set of representative prototypes as cluster anchors and merges the remaining ones, and is trained with three self-supervised losses.

Critical strengths
  1. The division of labor between the two stages is internally coherent. CPA learns a shared prototype vocabulary and RPD selects cluster anchors from it; Table 3 shows that the number of activated prototypes is generally much larger than the number of ground-truth classes, which motivates the RPD step, and each of the three self-supervised losses has a clear purpose.

  2. The evaluation protocol is reasonably standard, and the model is lightweight with reported costs. Four metrics, 20 independent runs with mean and standard deviation, and baselines using official implementations tuned per dataset; the full model has about 171.2K parameters, RPD needs only 25.3K per dataset, and Appendix B gives the parameter breakdown.

Critical weaknesses
  1. The credibility of the core results is questionable; the numbers exceed what is plausible for this task. For example, Primary Tumor is a classic hard dataset with 339 samples and 21 classes, on which supervised classification accuracy is typically 40–50%. This paper's unsupervised clustering reaches 82.76% ACC and 74.68% NMI after Hungarian matching, more than 30 points above the supervised ceiling.

  2. The number of representative prototypes closely matches the number of ground-truth classes, while the hyperparameters of the selection process are not disclosed.

  3. The ablation results contradict the paper's own conclusion: removing RPD is actually better on most datasets.

  4. The "cross-dataset alignment" in the title contributes little to the results, and there is no direct evidence for it.

  5. Hyperparameter sensitivity is studied only for K. There is no sensitivity analysis for ρ, γ, α or the six λ's, even though point 2 shows that γ and α directly determine the number of clusters.

  6. The manuscript quality is poor. All figures are raster images rather than vector graphics; the method section contains no architecture figure,.

Questions for authors

Please refer to Weaknesses

Comments on the paper

1 comment

Dear reviewers and authors,

As the discussion period is limited, please follow up on the latest responses and comments in a timely manner to address any remaining concerns. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.