evclust: Python library for evidential clustering

Armel Soubeiga; Violaine Antoine

REVIEW 3 major objections 6 minor 1 cited by

evclust: Python library for evidential clustering

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The Python library evclust implements 13 evidential clustering algorithms in one package.

desk verdict A useful Python packaging of evidential clustering algorithms, but the paper itself does not verify the implementations; should go to review with requests for validation. read the letter →

arxiv 2502.06587 v1 pith:AN724UVP submitted 2025-02-10 cs.SE cs.CVcs.LG

Armel Soubeiga , Violaine Antoine This is my paper · ORCID

classification cs.SEcs.CVcs.LG

keywords evidentialclusteringcredalpartitionDempster-ShafertheorybelieffunctionsPythonlibrarysoftuncertaintyquantificationsoftware

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

evclust is a Python library that implements published evidential clustering algorithms, a family of methods in which each object's cluster membership is represented by a Dempster–Shafer mass function on subsets of the cluster set rather than by a single label or fuzzy degree. The paper claims that this fills a concrete gap: evidential clustering tooling previously existed mainly in R, and Python had no comprehensive library for producing or analyzing credal partitions. The package provides 13 algorithms spanning attribute, proximity, categorical, and multi-view data, together with utilities for plotting, summarizing, transforming, and evaluating credal partitions. A reader should care because evidential clustering captures partial ignorance—an object can carry belief in a whole set of clusters without any preference among them—and this library makes that machinery usable inside the standard Python data-science stack.

What carries the argument

The central object is the mass function from Dempster–Shafer theory: a function $m:2^{\Omega}\to[0,1]$ with $\sum_{A\subseteq\Omega} m(A)=1$ that distributes belief over subsets of the cluster set $\Omega$. A credal partition is the collection of such mass functions $m_i$ for the $n$ objects, generalizing hard, fuzzy, and rough partitions. The library's algorithmic core alternates between updating these masses and updating cluster prototypes (or medoids) to minimize ECM-style objective functions, while its utility layer derives belief/plausibility functions, lower and upper approximations, and relational matrices from the masses.

What would settle it

Run each of the 13 functions in Table 2 on its worked example and compare the returned credal partition, objective value, and nonspecificity with the values printed in the paper (for Iris, criterion 38.82 and nonspecificity 0.22) and with the outputs of the reference R evclust package on the same data; any function that crashes, produces non-normalized masses, or disagrees materially would falsify the claim of a comprehensive, faithful implementation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that evclust is a working, comprehensive Python implementation of evidential clustering. Each object is assigned a mass function over subsets of the cluster labels, and the library's functions optimize objective functions in the spirit of evidential c-means (ECM) to determine these masses and cluster prototypes. The paper describes how ecm() handles attribute data, recm() and kevclus() handle dissimilarity matrices, catecm() handles categorical attributes, egmm() provides a model-based variant, and wmvec() and mecmdd() handle multi-view data. It then demonstrates the workflow on the Iris dataset and on the multi-view Prop dataset, showing that outputs include focal sets, prototypes, nonspecificity, and view weights. If the claim is right, Python users can reproduce and extend the published evidential-clustering methods without reimplementing them from the literature.

Load-bearing premise

The load-bearing premise is that the code in the linked repository, version v0.2, really does implement all 13 algorithms listed in Table 2 and behaves as the documentation and examples claim; the paper supplies illustrative outputs but no test suite, benchmark, or comparison to the reference R implementation to verify that.

Editorial extensions

If this is right

Python users can compute credal partitions from attribute, proximity, categorical, and multi-view data within one library that interoperates with numpy, pandas, scipy, and scikit-learn.
Analysts can convert credal partitions into hard, fuzzy, or rough partitions, summarize their focal sets, and visualize them in two-dimensional, PCA, or time-series plots using the utility functions.
The metric functions supply a nonspecificity measure for imprecision and a credal Rand index for comparing two credal partitions, giving concrete quantitative handles on uncertainty.
Researchers can run published algorithm variants (ECM, RECM, k-EVCLUS, CatECM, EGMM, BPEC, ECMdd, MECM, WMVEC, MECMdd, CCM) in Python instead of reimplementing them from the literature.
Bundled datasets such as Iris, Decathlon, Protein, fourclass, and Prop give users ready material for learning and testing the methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct check the paper leaves implicit is numerical fidelity: running each evclust function on the same data as the reference R evclust package and comparing masses, prototypes, and objective values would verify that the Python port reproduces the published methods; the paper itself shows only illustrative outputs.
The roadmap of planned algorithms (belief shift clustering, dynamic evidential c-means, deep evidential clustering, decision-tree evidential clustering, transfer-learning evidential c-means) suggests evclust is intended to grow into the standard Python entry point for the whole evidential-clustering family.
Because the package is MIT-licensed and built on scikit-learn-compatible data structures, it could plausibly be wrapped as a standard scikit-learn cluster estimator, letting evidential clustering participate in general pipelines and model-selection workflows.
The nonspecificity metric, combined with the credal Rand index, could support automatic selection of the number of clusters in evidential clustering, an application the paper mentions only in passing.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The one thing to know: this is a software paper for evclust, a Python library bundling 13 evidential clustering algorithms with utilities and metrics. The algorithms are all previously published; the new content is the packaged implementation. That alone is a legitimate contribution — R has had evclust for years, and Python users didn't have a one-stop library. The code is public, MIT licensed, and the authors document metadata properly. The example outputs look plausible.

What the paper does well: it gives a compact introduction to evidential clustering and a clear table of the methods with complexity. The library design seems sensible — algorithms, utils, metrics in separate modules. The included datasets are a nice touch.

Now the soft spots. The most load-bearing: nothing in the paper verifies that the 13 implementations actually match the published methods. No test suite is mentioned, no comparison against the R evclust reference, no benchmarks. Section 4 shows two examples. A silent implementation bug would invalidate the central claim for that algorithm. That's a real weakness for a paper whose contribution is code. The stress-test note flags a function-name mismatch: §3.1.9 says WMVEC-FP is implemented as wmvec_pf(), while Table 2 lists wmvec_fp(). That one is real and suggests the manuscript wasn't checked against the repo. The other "mismatches" in that note don't hold up — mecm() and mecmdd_rwl() are different algorithms, and ev_pcaplot appears consistently. The RECM sentence 'Although based on the assumption that the input dissimilarities are squared Euclidean distances' is indeed a dangling fragment.

There's also a small inconsistency in the multi-view example: the text says type='simple' means we are only interested in the focal set equal to 2, which is confusingly phrased. Minor.

Is the paper honest? Yes. It doesn't claim new algorithms. The citation pattern is fine — two algorithms are from the authors' own prior work and cited. The central claim is independent of whether those algorithms are correct.

Who gets value: researchers and practitioners using evidential clustering who want a Python-native alternative to R. With a requirements table and clear API signatures, it's a useful starting point.

Recommendation: send to peer review, but the reviewers should ask for a reproducibility check — either a test suite in the repo or a numerical comparison against the R implementations for at least a couple of algorithms, plus fixing the function-name inconsistency and the dangling sentence. The software itself likely works, but the paper needs to demonstrate that.

Referee Report

3 major / 6 minor

Summary. The paper presents evclust, a Python library for evidential clustering. It reviews the basics of belief functions and credal partitions, lists 13 implemented algorithms in Table 2 with complexity and function names, describes utility and metric functions, and gives two usage examples (ECM on Iris, WMVEC on ProPt). The central claim is that evclust fills the gap of a comprehensive Python library for state-of-the-art evidential clustering algorithms, offering visualization, evaluation, and analysis tools integrated with the Python data science ecosystem.

Significance. If the implementation is faithful to the cited algorithms, evclust would be a useful and timely contribution to the Python ecosystem, complementing the existing R package and integrating with numpy, pandas, scikit-learn, and matplotlib. The paper's strengths are its clear conceptual introduction, the breadth of the algorithm list, and the provision of open-source code and documentation. However, the manuscript currently provides no verification that the 13 implementations are correct, no comparison to reference implementations, and no benchmarks, so the central claim of a 'comprehensive' and working library is not yet established by the paper itself.

major comments (3)

[Section 4, Table 2] The central claim that evclust implements state-of-the-art evidential clustering algorithms is not supported by any validation. Section 4 shows only two illustrative runs (ecm on Iris and wmvec on ProPt) with no expected outputs, no unit tests, and no numerical comparison against the R evclust package or the original papers. For a software paper, a silent implementation error in any one of the 13 functions would invalidate the claim for that algorithm and undermine the advertised completeness. Please add a test suite that cross-checks results against reference implementations on small datasets and report the outcomes.
[Section 3.1.9 vs. Table 2] The function name for WMVEC-FP is given as wmvec_pf() in Section 3.1.9 but as wmvec_fp() in Table 2. This is a concrete API inconsistency that suggests the manuscript has not been checked against the repository. Please reconcile the name and systematically verify every function name in the paper against the actual v0.2 code.
[Table 2 vs. Section 3.1.8] Table 2 lists MECM as taking 'Attribute data', but Section 3.1.8 describes MECM as 'a extend version of ECM for partitioning relational data' with prototypes constrained to the original data set, i.e., medoid-like behavior. These statements contradict each other and directly affect a user's choice of algorithm by input type. Please correct the table or the text and make the input requirements explicit.

minor comments (6)

[Section 3.1.2] The sentence 'Although based on the assumption that the input dissimilarities are squared Euclidean distances' is incomplete, and the relationship between τ_ij in Eq. (7) and the squared-distance assumption is never clarified.
[Section 4.2] The explanation 'type="simple" means that we are only interested in the focal set equal to 2' is incorrect or garbled; type='simple' should correspond to focal sets consisting of the empty set, singletons, and Ω, not to a single focal set of size 2.
[Section 4.1] In the ev_pcaplot call, the argument 'splite' appears to be a typo (likely 'split'); please correct it in the displayed code.
[Section 3.1.5] The EGMM description is too vague: 'combines Gaussian mixture using and covariance matrices' is incomplete/garbled, and no function signature or usage example is provided for egmm(), unlike other algorithms.
[Section 1] The claim that 'Python lacked a comprehensive library' for evidential clustering is stated without a survey or citation; please either support it with a broader comparison or soften it to avoid an unsupported gap claim.
[Table 2] The column header 'F unction' contains an unintended space; also, the notation for complexity (e.g., O(np2c) vs O(nf p2c)) should define p and f consistently in the caption or text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports a software implementation of externally published algorithms, with no derivation claim that reduces to its inputs.

full rationale

The paper is a software library description, not a derivation or prediction. Its central claim is that the evclust Python library implements a set of evidential clustering algorithms originally published in separate, non-self-referential papers (ECM, RECM, k-EVCLUS, etc.), plus utilities and metrics. There is no fitted parameter renamed as a prediction, no ansatz smuggled in via citation, and no uniqueness theorem invoked to force a choice. Two algorithms (CatECM and MECMdd) cite the authors' own prior work, but this self-citation is not load-bearing: the algorithms are independently published and the library's implementation claim does not reduce to those citations. The absence of numerical validation against the R reference implementations is a correctness or completeness concern, not a circularity concern. The paper makes no claim whose truth is equivalent to its own assumptions by construction; therefore the circularity score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The paper introduces no fitted parameters and no new theoretical entities. The central claim rests on the correctness of published algorithms as reimplemented in Python, and on the unverified assertion that no comparable Python library existed. Those are domain assumptions rather than ad hoc inventions.

assumptions (2)

domain assumption The 13 published algorithms cited in Table 2 are correct and the Python code reproduces them faithfully.
The paper provides no verification against reference implementations or ground truth; correctness is inherited from the cited literature and the unshown code.
domain assumption Python lacked a comprehensive evidential clustering library before evclust, so the stated gap is real.
The Introduction asserts this gap without a systematic survey of existing Python belief-function or clustering packages.

how reviews work

0 comments

Cite this review

Pith. "Pith review of evclust: Python library for evidential clustering." pith.science (2026). https://pith.science/paper/AN724UVP

@misc{pith2026250206587,
  author       = {Pith},
  title        = {Pith review of: evclust: Python library for evidential clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AN724UVP}},
  note         = {Machine review of arXiv:2502.06587}
}

read the original abstract

A recent developing trend in clustering is the advancement of algorithms that not only identify clusters within data, but also express and capture the uncertainty of cluster membership. Evidential clustering addresses this by using the Dempster-Shafer theory of belief functions, a framework designed to manage and represent uncertainty. This approach results in a credal partition, a structured set of mass functions that quantify the uncertain assignment of each object to potential groups. The Python framework evclust, presented in this paper, offers a suite of efficient evidence clustering algorithms as well as tools for visualizing, evaluating and analyzing credal partitions.

Figures

Figures reproduced from arXiv: 2502.06587 by the authors.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Explainable Evidential Clustering
cs.LG 2025-07 conditional novelty 5.0 of 10

This paper proposes IEMM, a decision-tree explainer for evidential clustering that minimizes a utility-weighted 'evidential mistakeness' cost.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages · cited by 1 Pith paper

[1]

John Wiley & Sons, 2009

Leonard Kaufman and Peter J Rousseeuw.Finding groups in data: an introduction to cluster analysis. John Wiley & Sons, 2009

2009
[2]

Scikit-fuzzy: Fuzzy logic toolbox for python

Joshua Warner et al. Scikit-fuzzy: Fuzzy logic toolbox for python. https://github.com/scikit-fuzzy/ scikit-fuzzy, 2018

work page 2018
[3]

Unsupervised possibilistic clustering

Miin-Shen Yang and Kuo-Lung Wu. Unsupervised possibilistic clustering. Pattern Recognition, 39(1):5–21, 2006

work page 2006
[4]

Rough clustering utilizing the principle of indifference.Information Sciences, 277:358–374, 2014

Georg Peters. Rough clustering utilizing the principle of indifference.Information Sciences, 277:358–374, 2014

work page 2014
[5]

A survey of evidential clustering: Definitions, methods, and applications.Information Fusion, 115:102736, 2025

Zuowei Zhang, Yiru Zhang, Hongpeng Tian, Arnaud Martin, Zhunga Liu, and Weiping Ding. A survey of evidential clustering: Definitions, methods, and applications.Information Fusion, 115:102736, 2025

work page 2025
[6]

evclust: An R Package for Evidential Clustering

Thierry Denoeux. evclust: An R Package for Evidential Clustering. The Comprehensive R Archive Network (CRAN), 2023. R package version 2.0.3

work page 2023
[7]

Representing uncertainty and imprecision in machine learning: A survey on belief functions.Journal of King Saud University - Computer and Information Sciences, 36(1):101904, 2024

Zhe Liu and Sukumar Letchmunan. Representing uncertainty and imprecision in machine learning: A survey on belief functions.Journal of King Saud University - Computer and Information Sciences, 36(1):101904, 2024

work page 2024
[8]

Marie-Hélène Masson and T. Denœux. Ecm: An evidential version of the fuzzy c-means algorithm.Pattern Recognition, 41(4):1384–1397, 2008

work page 2008

Show all 21 references

[9]

Recm: Relational evidential c-means algorithm

Marie-Hélène Masson and Thierry Denœux. Recm: Relational evidential c-means algorithm. Pattern Recognition Letters, 30(11):1015–1026, 2009

2009
[10]

Evclus: Evidential clustering of proximity data.IEEE Transactions on Systems, Man, and Cybernetics-Part B: Cybernetics, 34(1):95–109, 2004

Thierry Denoeux and Marine H Masson. Evclus: Evidential clustering of proximity data.IEEE Transactions on Systems, Man, and Cybernetics-Part B: Cybernetics, 34(1):95–109, 2004

2004
[11]

Evidential clustering of large dissimi- larity data

Thierry Denœux, Songsak Sriboonchitta, and Orakanya Kanjanatarakul. Evidential clustering of large dissimi- larity data. Knowledge-Based Systems, 106:179–195, 2016

2016
[12]

Christie, and Sylvain Moreno

Abdoul Jalil Djiberou Mahamadou, Violaine Antoine, Gregory J. Christie, and Sylvain Moreno. Evidential clustering for categorical data. InProceedings of the International Conference on FuzzySystems (FUZZ-IEEE), pages 1–6, 2019

2019
[13]

Egmm: An evidential version of the gaussian mixture model for clustering.Applied Soft Computing, 129:109619, 2022

Lianmeng Jiao, Thierry Denœux, Zhun ga Liu, and Quan Pan. Egmm: An evidential version of the gaussian mixture model for clustering.Applied Soft Computing, 129:109619, 2022

2022
[14]

Bpec: Belief-peaks evidential clustering

Zhi-gang Su and Thierry Denoeux. Bpec: Belief-peaks evidential clustering. IEEE Transactions on Fuzzy Systems, 27(1):111–123, 2018

2018
[15]

Ecmdd: Evidential c-medoids clustering with multiple prototypes

Kuang Zhou, Arnaud Martin, Quan Pan, and Zhun ga Liu. Ecmdd: Evidential c-medoids clustering with multiple prototypes. Pattern Recognition, 60:239–257, 2016

2016
[16]

Median evidential c-means algorithm and its application to community detection.Knowledge-Based Systems, 74:69–88, 2015

Kuang Zhou, Arnaud Martin, Quan Pan, and Zhun ga Liu. Median evidential c-means algorithm and its application to community detection.Knowledge-Based Systems, 74:69–88, 2015. 12

2015
[17]

Adaptive weighted multi-view eviden- tial clustering with feature preference.Knowledge-Based Systems, 294:111770, 2024

Zhe Liu, Haojian Huang, Sukumar Letchmunan, and Muhammet Deveci. Adaptive weighted multi-view eviden- tial clustering with feature preference.Knowledge-Based Systems, 294:111770, 2024

2024
[18]

Multi-view relational evidential c-medoid clustering with adaptive weighted

Armel Soubeiga, Violaine Antoine, and Sylvain Moreno. Multi-view relational evidential c-medoid clustering with adaptive weighted. In2024 IEEE 11th International Conference on Data Science and AdvancedAnalytics (DSAA), pages 1–10, 2024

2024
[19]

Credal c-means clustering method based on belief functions

Zhun ga Liu, Quan Pan, Jean Dezert, and Grégoire Mercier. Credal c-means clustering method based on belief functions. Knowledge-Based Systems, 74:119–132, 2015

2015
[20]

Thierry Denœux and Orakanya Kanjanatarakul.Beyond Fuzzy,Possibilistic and Rough: An Investigation of Belief Functions in Clustering:, pages 157–164. 01 2017

2017
[21]

Klir and Mark J

George J. Klir and Mark J. Wierman.Uncertainty-Based Information: Elements of Generalized Information Theory, volume 15 ofStudies in Fuzziness and Soft Computing. Springer-Verlag, New York, 1999. 13

1999

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

John Wiley & Sons, 2009

Leonard Kaufman and Peter J Rousseeuw.Finding groups in data: an introduction to cluster analysis. John Wiley & Sons, 2009

2009

[2] [2]

Scikit-fuzzy: Fuzzy logic toolbox for python

Joshua Warner et al. Scikit-fuzzy: Fuzzy logic toolbox for python. https://github.com/scikit-fuzzy/ scikit-fuzzy, 2018

work page 2018

[3] [3]

Unsupervised possibilistic clustering

Miin-Shen Yang and Kuo-Lung Wu. Unsupervised possibilistic clustering. Pattern Recognition, 39(1):5–21, 2006

work page 2006

[4] [4]

Rough clustering utilizing the principle of indifference.Information Sciences, 277:358–374, 2014

Georg Peters. Rough clustering utilizing the principle of indifference.Information Sciences, 277:358–374, 2014

work page 2014

[5] [5]

A survey of evidential clustering: Definitions, methods, and applications.Information Fusion, 115:102736, 2025

Zuowei Zhang, Yiru Zhang, Hongpeng Tian, Arnaud Martin, Zhunga Liu, and Weiping Ding. A survey of evidential clustering: Definitions, methods, and applications.Information Fusion, 115:102736, 2025

work page 2025

[6] [6]

evclust: An R Package for Evidential Clustering

Thierry Denoeux. evclust: An R Package for Evidential Clustering. The Comprehensive R Archive Network (CRAN), 2023. R package version 2.0.3

work page 2023

[7] [7]

Representing uncertainty and imprecision in machine learning: A survey on belief functions.Journal of King Saud University - Computer and Information Sciences, 36(1):101904, 2024

Zhe Liu and Sukumar Letchmunan. Representing uncertainty and imprecision in machine learning: A survey on belief functions.Journal of King Saud University - Computer and Information Sciences, 36(1):101904, 2024

work page 2024

[8] [8]

Marie-Hélène Masson and T. Denœux. Ecm: An evidential version of the fuzzy c-means algorithm.Pattern Recognition, 41(4):1384–1397, 2008

work page 2008