acceptodds
Under review as a conference paper at ICLR 2027

GRANITE: Granular-Ball Autoencoding with Signed Triplet Evidence for Uncertainty-Aware Deep Clustering

Abstract

Clustering is a fundamental task in data mining and plays an important role in complex data analysis. Traditional single-granularity instance-level clustering methods treat samples as isolated points, limiting their ability to learn multi-scale representations. Granular-ball(GB) clustering obtains multi-granularity representations through spherical topology evolution, making it more suitable for handling complex data. However, existing GB clustering methods often rely on fixed feature and overlook uncertainty in ball–cluster assignments. To address them, we propose the Granular-Ball Autoencoding with Intuitionistic Triplet Evidence (GRANITE) algorithm for uncertainty-aware deep clustering. A GB autoencoder is designed to learn dynamic deep feature representations in the latent space. In addition, to explicitly characterize the uncertainty in ball–cluster relationships, we introduce intuitionistic fuzzy triplet evidence comprising membership, non-membership, and hesitation degrees. Furthermore, GRANITE employs a new hesitation evidence and introduces the GB distribution measure to regularize the clustering process, thereby improving robustness and efficiency. Experiments on multiple benchmark datasets demonstrate that GRANITE achieves superior clustering performance compared with existing methods.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Autoencoder, Fuzzy Clustering, Intuitionistic Fuzzy Set, Granular-Ball Computing

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_8gcQOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRANITE, an uncertainty-aware deep clustering framework based on granular-ball representations. The method first constructs granular balls from samples and represents each ball using geometric descriptors including center, radius, and an intrinsic hesitation descriptor. A GB-level autoencoder is introduced to learn latent representations, followed by an intuitionistic fuzzy triplet formulation to model ball-cluster relationships using membership, non-membership, and hesitation evidence. The paper further introduces a signed evidence-based residual objective, covariance-aware robust clustering loss, and a dynamic GB reconstruction procedure to update sample-to-ball assignments during training.

The authors validate GRANITE on eight UCI clustering datasets and two image segmentation benchmarks, reporting improvements over traditional clustering, deep clustering, and granular-ball-based baselines. Additional ablation studies, visualization analyses, and convergence analysis are provided.

Critical strengths
  1. The paper presents a technically comprehensive framework that integrates granular-ball representations with deep clustering.
  2. The method achieves competitive performance on the evaluated benchmarks.
Critical weaknesses
  1. GRANITE combines several existing components, including granular-ball representations, autoencoder-based clustering, intuitionistic fuzzy modeling, robust metric learning, and alternating optimization. However, the authors introduce many additional variables and optimization mechanisms, but the experiments do not establish that these complexities are necessary.
  2. The clustering objective mainly depends on the signed coefficient wij=μij−νijw_{ij}=\mu_{ij}-\nu_{ij}, while the additional hesitation variable only affects the entropy regularization term. Therefore, it is unclear whether the full intuitionistic fuzzy triplet provides meaningful uncertainty modeling beyond a simpler signed assignment formulation.
  3. The learned radius weighting may allow the model to reduce the contribution of difficult granular balls by increasing latent scale variables, but the paper does not analyze whether these learned scales preserve their intended geometric meaning.
  4. The experimental validation does not adequately support the claimed advantages. The clustering evaluation is mainly conducted on small-scale UCI datasets, without standard large-scale deep clustering benchmarks.
Questions for authors
  1. How does the intuitionistic fuzzy triplet provide additional benefits compared with directly optimizing the signed coefficient wijw_{ij} with an appropriate regularization term?
  2. Can the authors provide an analysis showing that the learned latent scale variables correspond to meaningful granular-ball geometric properties rather than merely reducing the contribution of difficult samples in the clustering objective?
  3. How much performance improvement is contributed by dynamic granular-ball reconstruction and the learned weighting mechanism? Have the authors compared against variants with fixed initial granular balls and without learned scale weighting?
  4. Could the authors clarify whether the method has been evaluated on larger-scale deep clustering benchmarks (e.g., MNIST/CIFAR-10) and provide runtime comparisons with existing approaches?
Reviewer_D9fNOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRANITE, an uncertainty-aware Granular Ball (GB) deep clustering method. Addressing the issues of existing GB clustering methods relying on static features and lacking explicit modeling of ball-cluster assignment uncertainty, the authors designed a GB-level autoencoder to encode the center, radius, and intrinsic hesitancy descriptor of each initial GB into a latent space and reconstruct these structured descriptors. In the latent space, GRANITE introduces intuitionistic fuzzy ternary evidence (membership degree, non-membership degree, hesitancy degree) for each ball-cluster pair, defines a symbolic hesitancy evidence coefficient, and jointly optimizes the reconstruction loss, negative entropy regularization, covariance-aware GB distribution loss, and prototype regularization. The authors provide an optimization derivation and convergence analysis using block coordinate descent, and compare GRANITE with various traditional, deep, and GB clustering methods on eight UCI datasets and image segmentation tasks (BSD500, DUTS), reporting metrics such as ACC, NMI, ARI, and mIoU.

Critical strengths
  1. The problem motivation is reasonable, combining two valuable directions. Granular sphere computing excels at multi-scale and efficient representation of data, deep autoencoders excel at learning nonlinear representations, and intuitionistic fuzzy sets excel at modeling uncertainty. Combining the three for GB clustering is a natural and innovative idea.
  2. The method design is relatively complete. GRANITE not only proposed a loss, but also constructed a GB level auto-encoder, intuitionistic fuzzy triplet, covariance aware distribution loss, and joint optimization objective, and provided block update formulas for prototype, allocation, covariance, and network parameters (Eq. 10-13, Eq. 38-52). The derivation in Appendix C is more detailed. Theoretical analysis has a certain depth. The author proved the descent property of the surrogate function (Theorem 1, 2) and provided monotonic descent and finite lower bounds for the objective value (Eq. 14-15, Eq. 69-72). Although based on block coordinates and proxy functions, it provides theoretical support for method optimization.
  3. The experiment covers a wide range. ACC, NMI, ARI were reported on eight UCI datasets, and ablation, parameter sensitivity, and t-SNE visualization were performed. The image segmentation experiments (Table 3, Fig. 6) demonstrate the generalization attempts of the method.
  4. Good reproducibility. Provides anonymous code links, pseudo-code, dataset statistics, hyper-parameter settings, and convergence analysis, with a relatively complete appendix.
Critical weaknesses
  1. The number of granular balls is fixed, and as far as I know, almost all granular balls are adaptive. How to achieve adaptive iteration with a fixed number, in other words, how to construct a multi granularity representation that conforms to the true distribution with a fixed number? Meanwhile, the impact of the fixed quantity parameter on the results was not discussed in the experiment.

  2. This article proposes four contributions, but they do not correspond to the problems to be solved. Also, are the four contributions too many? Similarly, there has been a lot of work on introducing fuzzy granular computing, and this article has limited research and understanding of related work.

  3. There are already methods using fixed feature GB, such as AWGBGAE, W-GBC, etc. Especially the former also uses autoencoders to iteratively optimize features and jointly optimize clustering objectives end-to-end. I would like to ask the author what specific problem this work aims to solve? (Unclear problems and scenarios to be solved in the job)

  4. The basic logic of the algorithm cannot be clearly expressed in flowchart 2.

  5. The experimental section selected 8 datasets, which were too simple. The highest dimension was only 36 dimensions, with a maximum sample size of around 10000. However, the work in this paper is based on autoencoders, and the feature dimension and sample size cannot fully reflect the effectiveness and role of the feature work in this paper.

  6. At the same time, the work in this article is based on deep clustering, so it should be compared with the current cutting-edge work related to deep clustering. The selected baseline methods for deep clustering are too few.

Questions for authors
  1. On adaptive iteration with a fixed number of granular balls and the construction of a multi-granularity representation: The paper states that the number of granular balls is fixed while also claiming that the granular balls are adaptively iterated. Please clarify what exactly is “adaptive” when the number is fixed—i.e., which variables or mechanisms are updated adaptively. How can a fixed number of granular balls be used to construct a multi-granularity representation that conforms to the true data distribution? In addition, how is the fixed-number parameter determined, and why was its impact on clustering results, representation quality, and iteration stability not discussed in the experiments? Please provide a parameter sensitivity analysis or theoretical/empirical justification.

  2. On the research problems, contributions, and distinction from existing fixed-feature GB methods: The paper lists four contributions, but their correspondence to the problems to be solved is unclear, and it is also questionable whether four contributions are too many or too scattered. Please explicitly state the core problem(s) and scenario(s) this work aims to address, and map each contribution to these problems. Moreover, methods using fixed feature GB already exist, such as AWGBGAE and W-GBC. In particular, AWGBGAE also uses an autoencoder to iteratively optimize features and jointly optimize the clustering objective end-to-end. Please clarify the essential differences between the proposed method and these works, and specify which concrete, previously unsolved problems this work actually solves.

  3. On the basic logic of the algorithm and the clarity of Flowchart 2: Flowchart 2 does not clearly convey the basic logic and key steps of the algorithm. Please explain the inputs, outputs, update order, and termination conditions of each module in Flowchart 2—especially the interaction among the fixed number of granular balls, feature representation learning, multi-granularity construction, and clustering objective optimization—and revise the flowchart or add pseudocode accordingly so that the algorithm can be reproduced.

  4. On the experimental datasets and validation of the autoencoder-based method: The experiments use only eight datasets, with a maximum dimensionality of only 36 and a maximum sample size of around 10,000, which are relatively simple. Since the proposed method is based on an autoencoder, are such data scale and feature dimensionality sufficient to demonstrate the role of deep feature learning? Please justify the dataset selection and supplement experiments on higher-dimensional, larger-scale, or more complex data to verify the effectiveness and scalability of the method under different feature dimensionalities and sample sizes.

  5. On the insufficient comparison with deep clustering baselines and related frontier work: Since this work is based on deep clustering, the selected deep clustering baselines are too few to demonstrate its advantages relative to current frontier work. Please add systematic comparisons with representative recent deep clustering methods and explain the baseline selection criteria. In addition, please further clarify the relationship between this work and related work on fuzzy granular computing, fixed-feature GB, and deep clustering, so as to justify the sufficiency of the experimental comparisons and the accuracy of the research positioning.

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

The paper introduces GRANITE, a deep clustering method based on granular balls. First, k-Means is used to create 'granular balls', then, these balls are combined with 'intuitionistic fuzzy triplets' for handling uncertainty of membership, an autoencoder, and a covariance-aware distribution loss (based on the Mahalanobis distance).

The granular balls are a concept introduced by Xia et al. and in this paper, they are used within the reconstruction loss. The "intuitionistic triplet evidence" describes for each pair of a ball and a cluster the probabilities for membership μ\mu, non-membership ν\nu, and leaves 1−μ−ν1-\mu-\nu as "cluster specific hesitation degree". The Mahalanobis distance is then used for a distribution loss that is part of the clustering objective. All aspects together are passed to an autoencoder.

Critical strengths

S1 Interesting Idea

Performing deep clustering on different granularity levels is a promising idea to find complex structures in large, high-dimensional datasets.

Critical weaknesses

W1 Experiments on benchmark data are not useful.

a) The datasets are very low-dimensional and small.

Deep Clustering makes sense if data is large and/or high-dimensional. Thus, such datasets should be used for testing and comparing to other methods. Dimensionality of tested data ranges from 4 to 36 and most datasets are smaller than 1000 data points (the biggest three datasets are 2k - 11 k datapoints). Instead, use datasets like CIFAR-19, CIFAR-20, COIL20, COIL100, HAR, USPS, MNIST, FMNIST, KMNIST, ImageNet, etc.

b) Competitors are not adequate.

You need to compare your method with methods developed for similar goals. The chosen competitors rather seem like a random bunch of methods that are never discussed (see W2), nor described briefly. Especially SpEx, that you classify as deep method in the upper part of Tab. 1, and as traditional method in the lower part of the same table is problematic: a) it is a explainable clustering method aimed at making existing clusterings more explainable (with a trade-off regarding quality). On page 25 you state "In particular, SpEx was originally proposed as a spectral explainable clustering method; in this work, we enhance it with deep learning and therefore categorize the enhanced variant as a deep clustering method". This is highly problematic for several reasons. 1) You should compare to the original method. 2) If you make changes, they need to be disclosed before using the method and discussed in the context of the experiment results. 3) Such changes need to be explained in detail. Only stating you enhance it with deep learning is not reproducible. Showing the results as is, it is misleading readers to believe GRANITE is better than any method available, which is not true.

c) Experimental details are left out.

You need to state how you optimized parameters. Only saying they are fixed throughout is not enough. Especially for the competitors, it is important to know how they were tuned and that this was done in a fair manner (i.e., state all tested hyperparameters, make sure you test similarly many settings for each method, and state your criterion by which the reported results were chosen). Also for your ablation study, explain which values were tested and why, especially as they have very different scales.

W2 Related work is missing and disregarded as competitors

There is no related work section and deep clustering, as the main field this paper positions itself in, is not discussed sufficiently. The only deep clustering methods included in the comparative experiments are DeepCluster, TableDC, MiniClustering, and these are not even cited properly or referred to in the experiment section (only in the appendix). Deep clustering methods are aimed large, high-dimensional datasets, while the tested datasets here are mostly small, low-dimensional tabular datasets.Thus, it is not too surprising that GRANITE seems to outperform competitors. Please be fair and give details about the hyperparameter optimization and compare at least to some standard deep clustering approaches, e.g. (but not limited to): DEC [1], and IDEC [2], DCN [3], DipDECK [4], SCAN [5], ...

[1] J. Xie, R. B. Girshick, and A. Farhadi, “Unsupervised deep embedding for clustering analysis,” in ICML, ser. JMLR Workshop and Conference Proceedings, vol. 48. JMLR.org, 2016, pp. 478–487.

[2] X. Guo, L. Gao, X. Liu, and J. Yin, “Improved deep embedded clustering with local structure preservation,” in IJCAI. ijcai.org, 2017, pp. 1753–1759.

[3] B. Yang, X. Fu, N. D. Sidiropoulos, and M. Hong, “Towards k-meansfriendly spaces: Simultaneous deep learning and clustering,” in ICML, Proceedings of Machine Learning Research, vol. 70. PMLR, 2017, pp. 3861–3870.

[4] C. Leiber, L. G. M. Bauer, B. Schelling, C. Bohm, and C. Plant, “Dip-based deep embedded clustering with k-estimation,” in KDD. ACM, 2021, pp.903–913.

[5] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European Conference on Computer Vision, pages 268–285. Springer, 2020.

W3 Important details missing or only in appendix.

Many details on how the method works or which hyperparameters were chosen for the experiments are left out or only mentioned in the appendix. Even there, I did not find the hyperparameter settings for the competitors. Large parts of the method would be very hard to reproduce without the appendix, please include all relevant steps of the method in the main paper (e.g., the full section B in the Appendix)

W4 Missing motivation

It is not really clear when this method should be used. Deep clustering is typically for large and/or high-dimensional data, but these advantages are not discussed or shown in the experiment section (see above). Synthetic experiments could help to show what the actual goal of the method is.

Mixed:

  • The accuracy values in Fig. 1 (with two decimals even) imply that these are actual experiments that were performed, but the image and description implies it is rather an illustration. Please clarify. If this is a real experiment, what was the data and the settings?
  • The title suggests that the method can handle uncertainty in data, which is a large research area on its own that is not properly discussed in the paper -- please clarify more explicitly and early what you mean with uncertainty-awareness.
  • There are many hidden parameters, but it does not become clear how many (and which ranges) you tested for the experiments.
  • If you average your results over 5 runs, you should also report the standard deviation.

Minor:

  • Especially on page 1, some whitespaces before citations are missing. Please be more careful in your use of \citep and \citet.
  • Using comic sans (Fig.1, Fig.2) is not very common in scientific papers.
  • Please increase font sizes a lot. The captions and texts in most figures are not readable at all.
  • The diagram in Fig. 4 should be a bar chart to allow better comparison.
  • The code is without a single comment.
Questions for authors

Q1) The main part describes the loss functions and how the embedding of the data is found, but not how the final clustering is extracted. I'd guess it is taken from the triplets?

Q2) How is k chosen for k-Means for the initialization of the granular balls? This is a hidden parameter that should be discussed more.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.