acceptodds
Under review as a conference paper at ICLR 2027

Bridged Clustering: Semi-Supervised Sparse Bridging

Abstract

We introduce Bridged Clustering, a semi-supervised framework to learn predictors from any unpaired input and output dataset. Our method first clusters and independently, then learns a sparse, interpretable bridge between the and clusters using only a few paired examples. At inference, a new input is assigned to its nearest input cluster, and the centroid of the linked output cluster is returned as the prediction . Unlike traditional semi-supervised learning, Bridged Clustering explicitly leverages output-only data, and unlike dense transport-based methods, it maintains a sparse and interpretable alignment. Through theoretical analysis, we show that with bounded mis-clustering and mis-bridging rates, our algorithm becomes an effective and efficient predictor. Empirically, our method is competitive with SOTA methods while remaining simple, model-agnostic, and label-efficient in low-supervision settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
clustering, semi-supervised 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 3 reviews, above 27% of the others.
Reviewer_ESTHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Bridged Clustering, a semi-supervised method that independently clusters unpaired inputs and outputs and uses a small paired set to align the clusters. Predictions are obtained by assigning an input to a cluster and returning the centroid of its corresponding output cluster, with soft-bridging and regression-based variants for less structured settings. Experiments on four multimodal datasets show strong performance on BIOSCAN, COCO, and Flickr30k, but weaker results on WIT. The paper also relates performance to clustering and bridging errors.

Critical strengths
  1. Bridged Clustering addresses a meaningful limitation of conventional semi-supervised learning by using independently available outputs. Its combination of separate clustering, sparse cluster alignment, and centroid prediction is computationally lightweight, model-agnostic, and easy to inspect. The soft-bridging and cluster-wise regression variants also provide natural extensions for imperfect clustering and within-cluster variation.

  2. The paper evaluates four multimodal datasets, both prediction directions, inductive and transductive settings, multiple supervision levels, and several adverse conditions. The use of 30 randomized trials per configuration and the analysis of cluster quality, bridge accuracy, overlapping categories, misspecified cluster counts, and larger cluster numbers provide useful evidence about when the method succeeds or fails.

Critical weaknesses
  1. The experimental construction strongly favors the method’s assumptions and uses unavailable latent information. The experiments provide one to four supervised examples per ground-truth cluster, guaranteeing that every cluster can be bridged. The datasets are also curated into balanced and relatively separable categories, while constrained k-means and the voting procedure were selected using AMI and ground-truth latent associations. WIT and Flickr30k categories are partly constructed from text features related to the evaluated output representations. These choices make it unclear whether the method would work with randomly sampled pairs and genuinely independent, imbalanced, or partially mismatched datasets.

  2. The theory and evaluation do not fully support the paper’s broad prediction claims. The theoretical section provides an existing misclustering bound and an informal majority-vote argument, but no finite-sample bridge-error or end-to-end prediction-risk guarantee. Empirically, the method predicts pretrained embedding centroids rather than actual images, captions, or sequences, and the multimodal datasets are made “unpaired” by hiding correspondences from originally paired data. The results therefore demonstrate latent-space cluster imputation under favorable conditions more convincingly than general prediction between independently collected modalities.

Questions for authors
  1. How would Bridged Clustering handle an input cluster for which the randomly sampled paired set contains no examples? Can you report results when the paired samples are drawn globally, without guaranteeing one or more examples per latent cluster?

  2. How much of the performance comes from the pretrained representation models? What happens with weaker, domain-shifted, self-supervised, or randomly initialized representations?

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

The paper introduces Bridged Clustering, a method designed for prediction tasks where there are abundant output-only samples, but few paired examples. In the proposed approach, the input and output data are clustered separately, then a mapping between the clusters is established using majority voting on the observed pairs. At inference, predictions are performed by associating each input cluster to its linked output centroid. Additionally, soft and regression-based extensions are proposed to address the limitations of hard cluster correspondence and centroid-based predictions.

The theoretical framework is built on the assumption of a finite shared latent grouping, with the input and output data being conditionally independent given the group. Empirical studies are conducted on four multimodal datasets, withholding most pairings, and evaluating mean squared error in the embedding space for both input-to-output and output-to-input prediction, considering both transductive and inductive scenarios. The proposed methods perform best on BIOSCAN, COCO, and Flickr30k, while Gromov--Wasserstein yields stronger results on WIT. However, further analysis is needed to clarify the source of performance gains and to

Critical strengths
  • The method provides a simple, interpretable way to incorporate input-only data and output-only data when paired samples are limited.
  • The decoupling of clustering and supervised alignment ensures that the framework remains modular and cost-effective as long as representations are precomputed.
  • The soft and regressed variants directly address hard correspondence and centroid-only prediction limitations.
  • Evaluation across four datasets, both directions, and two inference settings provides a useful empirical starting point.
Critical weaknesses
  • The experiments do not directly isolate the contribution of output-only data from clustering and representation quality.
  • Balanced, curated groups and knowledge of the true cluster count favor the chosen clustering pipeline and limit generality.
  • Marginal clustering accuracy does not justify the stated majority-vote bridge argument, and a prediction-risk guarantee is missing.
  • The cited clustering bound is used without its full assumptions and changes an empirical error guarantee into a population one.
Questions for authors
  1. Isolate the benefit of output-only data (lines 85--100, 297--315, 372--377).

    The empirical setup does not isolate the effect of output-only data. Please fix input clustering, paired set, representations, and splits, then vary output-only sample quantity (including none). As a baseline, predict the mean output of the pairs in each input cluster (with shrinkage if few pairs). This clarifies the value of output-only clustering. Report paired MSE differences and uncertainty, plus win rates, to show the impact.

  2. Correct the scope of the quoted clustering guarantee (lines 197--203, 445--455).

    The displayed Lloyd bound is attributed to Theorem 3.2 in Lu and Zhou (2016) https://arxiv.org/abs/1612.02099. That result controls empirical mismatch ratio after iterations and requires conditions on sample size, minimum cluster mass, signal-to-noise ratio, and initialization. Here ϵX\epsilon_X is defined as a fresh-point population probability, and only good initialization is mentioned. Please adapt the source assumptions, distinguish training and population errors, and supply the argument needed for the latter. The cited theorem also concerns Lloyd's algorithm, whereas the selected empirical procedure uses balanced constrained clustering; its applicability should be stated accurately.

  3. Account for the favorable curation and cluster information (lines 357--363; 683--695, lines 809--822).

    The experiments assume balanced groups, use the true cluster count, and tune clustering via Adjusted Mutual Information (AMI), which gives extra supervision. Please clarify whether AMI tuning used separate data and separate these idealized settings from less supervised scenarios. Tests with cluster-count variation and overlaps are included, but real-world issues like unbalanced groups and unsupervised cluster selection are not addressed. In Table 2, baseline comparisons use fewer components than the number of groups in some datasets; please match component counts and specify if and when validation pairs were used.

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

This paper introduces Bridged Clustering, a simple framework for learning from unpaired inputs and outputs with a few paired examples. The method clusters the two spaces independently, learns cluster correspondences through majority voting, and predicts the centroid of the corresponding output cluster. It also introduces soft bridging and clusterwise regression variants. The method shows promising results on several datasets. However, the experiments rely on carefully constructed groups with known cluster structure, and I am not convinced that they support the broader claims about learning from independently collected data.

Critical strengths
  • The method is very simple and easy to understand. Separating clustering from cross-space alignment makes the learned correspondences easy to inspect.
  • The authors compare against a broad set of baselines and evaluate both inductive and transductive settings. The additional experiments with overlapping groups and incorrect cluster counts are useful.
  • The results suggest that output-only data can be useful when the two spaces share a clear cluster structure. Soft bridging is a reasonable extension for imperfect cluster correspondences.
Critical weaknesses
  • The experimental setup is particularly favorable to the proposed method. The main experiments sample only 3-7 equally sized groups, use the known number of groups, and provide 1-4 paired examples per group. Appendix B also describes choosing balanced clustering based on agreement with latent labels. These choices use considerable knowledge of the underlying structure. The robustness experiments are helpful, but do not establish performance on naturally imbalanced data with an unknown cluster count and paired examples sampled without group labels.
  • The benefit of output-only data is not sufficiently isolated. A useful comparison would cluster inputs and predict the average paired target within each input cluster. An ablation varying the amount of output-only data while holding the other data fixed would also help. Without these controls, it is difficult to separate the value of modeling the output distribution from the benefit of the cluster-based prediction rule.
  • The theoretical support is weaker than the claims suggest. Section 8 gives an existing clustering bound and an informal majority-vote argument, but does not derive a prediction-error bound for the method. Low overall mis-clustering rates do not by themselves guarantee reliable voting within every input cluster, particularly for rare groups. The dependence on group frequency, paired-sample coverage, and centroid estimation should be made explicit.
  • Evaluation is limited to MSE in the embedding space. Since the basic method returns a cluster centroid, this does not establish useful instance-specific prediction. A downstream retrieval or biological prediction experiment would help clarify the practical value beyond recovering broad categories.
Questions for authors
  • How does the method perform with naturally imbalanced groups, a cluster count selected without ground-truth labels, and paired examples sampled uniformly from the full dataset?
  • Can the authors add the output-data ablation and input-clustering-only baseline described above?
  • Can the theoretical analysis quantify how many paired examples are needed to cover and correctly bridge the clusters, and connect this to prediction error?

Comments on the paper

2 comments

Dear reviewers, We value your time and thoughtful comments, which we addressed with the utmost respect. We have invested significant effort to answer your questions and check concerns raised by you, and we would appreciate any further feedback. If any additional analyses or experiments are in order, we would gladely add them. If not, and all of your concerns were resolved, we would appreciate your response, and if you could consider raising your scores accordingly. Thank you again for helping us improve the paper.

AuthorsOpenReview ↗

A summary of the rebuttal

Dear AC,

In direct response to the reviewers’ concerns, we substantially revised the manuscript and conducted five new experimental studies (Section 7.2 and Appendix A):

  • Output-only contribution: Introduced the reviewer-requested input-only clustering baseline and controlled ablations, demonstrating consistent benefits from output-only data.

  • Benchmark robustness: Tested global random pairing, imbalanced groups, unconstrained clustering, and independently selected cluster counts. BC variants outperform the strongest non-BC baselines in 11/15 comparisons.

  • Fairer baselines: Matched EM component counts and evaluated Ridge refinement, with BC outperforming matched-count EM across all nine dataset/count averages.

  • Downstream utility: Added output-gallery retrieval experiments, showing improved Top-1 group retrieval over the input-only baseline on all three datasets.

  • Representation sensitivity: Evaluated robustness to Gaussian noise and dimensionality reduction.

We also substantially revised the theoretical analysis to clarify assumptions and provide explicit paired-coverage and bridge-error bounds.

We believe these additions directly and substantively address the reviewers’ central concerns. Unfortunately, despite these extensive efforts, we have received minimal to no engagement with the new evidence. Reviewer ZXEL acknowledged our response but did not refer to anything specific we added, nor did they indicate whether it has addressed their original concerns. The other reviewers have not yet provided follow-up feedback. We respectfully ask the AC to consider the revised manuscript and new evidence when assessing whether the original concerns remain valid, particularly where the added experiments and analysis directly address them.

Thank you, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.