pith. sign in

arxiv: 2508.03310 · v3 · submitted 2025-08-05 · 📊 stat.ME

Robust fuzzy clustering with cellwise outliers

Giorgia Zaccaria , Lorenzo Benzakour , Luis A. Garc\'ia-Escudero , Francesca Greselin , Agust\'in Mayo-\'Iscar This is my paper

Pith reviewed 2026-05-19 00:46 UTC · model grok-4.3

classification 📊 stat.ME
keywords fuzzy clusteringcellwise outliersrobust statisticscluster-specific relationshipsdata contaminationmembership degreeshigh-dimensional data
0
0 comments X

The pith

Fuzzy clustering detects and corrects individual outlying cells by using relationships among variables that are specific to each cluster.

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

The paper proposes a fuzzy clustering method designed to handle contamination that affects single entries in a data matrix rather than whole rows. It assigns units to clusters with adjustable fuzziness while using those assignments to spot which cells are anomalous and to replace them based on the variable patterns found inside each cluster. Traditional robust approaches often discard entire cases when only a few cells are bad, losing the information in the remaining reliable entries. By contrast, this method keeps partial information from contaminated cases and lets the detected cluster relationships guide the cleaning step. A reader would care because many modern datasets have scattered anomalies that grow with the number of variables, making casewise deletion wasteful.

Core claim

The central claim is that a robust fuzzy clustering procedure can simultaneously control the fuzziness of unit assignments and identify outlying cells by exploiting the cluster-specific relationships among variables that the fuzzy approach itself uncovers, thereby correcting those cells without discarding the rest of the information in a contaminated row.

What carries the argument

The joint procedure that couples fuzzy membership degrees with cellwise outlier detection and imputation, letting the memberships highlight reliable cells and the within-cluster variable relationships serve as the basis for correction.

If this is right

  • Reliable cells inside contaminated cases remain available for cluster assignment instead of being lost.
  • Cluster-specific variable relationships improve the accuracy of cellwise outlier identification.
  • Tuning parameters allow users to balance fuzziness against robustness to contamination.
  • Simulation studies and real-data examples show the method works under cellwise contamination scenarios.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same idea of using fuzzy memberships to guide cell correction might transfer to other partitioning methods that produce soft assignments.
  • In very high dimensions the approach could reduce the need for separate imputation steps before clustering.
  • Users could test whether the corrected data matrix yields downstream predictions or visualizations that are more stable than those from casewise robust alternatives.

Load-bearing premise

The fuzzy clustering step can still recover accurate cluster-specific variable relationships even when some individual cells are contaminated.

What would settle it

Apply the method to a simulated data set with known cellwise outliers and check whether the detected and corrected cells match the planted anomalies more closely than a standard fuzzy clustering run followed by separate outlier screening.

Figures

Figures reproduced from arXiv: 2508.03310 by Agust\'in Mayo-\'Iscar, Francesca Greselin, Giorgia Zaccaria, Lorenzo Benzakour, Luis A. Garc\'ia-Escudero.

Figure 1
Figure 1. Figure 1: Artificial data with 5% of contamination per variable. Stars denote observations [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Artificial data: objective function curves [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Artificial data: difference between the knee point of ∆ [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Artificial data: ∆ plots for the first two variables, where units are sorted accord [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Artificial data: clustering structure with [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Artificial data: effect of the fuzzifier parameter [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Artificial data: effect of the fuzzifier parameter [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Artificial data: clustering results with [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Body fat data: analysis of the eleven variables selected [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Body fat data: objective function curves [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Body fat data: difference between the knee point of ∆ [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Body fat data: clustering results (Cluster 1: black, Cluster 2: blue, Cluster 3: [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: OECD data: proportion of hard assignments depending on [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: OECD data: objective function curves – it ranges from 14% to 91% depending on the choices of K and α, as shown in Figure 13c. The fuzzifier parameter m is set to 1.8 as it provides desirable levels of WA, with 29 weakly assigned regions out of 447 (18 for m = 1.6 and 55 for m = 2, as shown in the Supplementary Material). Once defined c, S, and m, it is possible to compute the objective function curves ( … view at source ↗
Figure 15
Figure 15. Figure 15: OECD data: difference between the knee point of ∆ [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: OECD data: clustering results. The color gradient indicates fuzzy assignments. [PITH_FULL_IMAGE:figures/full_fig_p030_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: OECD data: outliers of selected regions (yellow: reliable cells; blue: contam [PITH_FULL_IMAGE:figures/full_fig_p032_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Artificial data: ∆ plot for the last three variables of the data set, where units [PITH_FULL_IMAGE:figures/full_fig_p037_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Artificial data: additional example on the effect of the constant [PITH_FULL_IMAGE:figures/full_fig_p038_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Body fat data: tetrahedron plot showing cluster assignments (Cluster 1: black, [PITH_FULL_IMAGE:figures/full_fig_p039_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Body fat data: ∆ plot for each variable, where units are sorted according to [PITH_FULL_IMAGE:figures/full_fig_p040_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: OECD data: proportion of weak assignments depending on a subset of [PITH_FULL_IMAGE:figures/full_fig_p042_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: OECD data: ∆ plot for each variable, where units are sorted according to their [PITH_FULL_IMAGE:figures/full_fig_p044_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: OECD data: outlying cells per cluster 45 [PITH_FULL_IMAGE:figures/full_fig_p045_24.png] view at source ↗
read the original abstract

In a data matrix, we may distinguish between cases, each represented by a row vector for a statistical unit, and cells, which correspond to single entries of the data matrix. Recent developments in Robust Statistics have introduced the cellwise contamination paradigm, which assumes contamination on cells rather than on entire cases. This approach becomes particularly relevant as the number of variables increases. Indeed, discarding or downweighting entire cases because of a few anomalous cells in them, as done by traditional (casewise) robust methods, can result in substantial information loss, since the non-contaminated (or reliable) cells can still be highly informative. This philosophy can also be considered in fuzzy clustering, by assuming that reliable cells within a case may still provide useful information for determining fuzzy memberships. A robust fuzzy clustering proposal is thus introduced in this work, combining the advantages of dealing with outlying cells and simultaneously controlling the degree of fuzziness of unit assignments. The cluster-specific relationships among variables, detected by the fuzzy clustering approach, are also key to better identifying outlying cells and correct them. The strengths of the proposed methodology are illustrated through a simulation study and two real-world applications. The effects of the model's tuning parameters are explored, and some guidance for users on how to set them suitably is provided.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a robust fuzzy clustering method for data matrices subject to cellwise contamination. It combines fuzzy membership assignments with mechanisms to detect and correct outlying cells by exploiting cluster-specific variable relationships, thereby avoiding the information loss of traditional casewise robust approaches. The proposal is evaluated via a simulation study and two real-world applications, with exploration of tuning parameters for fuzziness and outlier detection.

Significance. If the iterative procedure reliably recovers cluster structures and relationships under cellwise noise, the work would meaningfully extend robust statistics and fuzzy clustering to high-dimensional settings where partial contamination is common. The dual focus on controlling assignment fuzziness while using detected relationships for cell correction represents a potentially useful synthesis, provided the fixed-point behavior is stable.

major comments (2)
  1. [Abstract / Algorithm] Abstract and method description: The central claim that 'the cluster-specific relationships among variables detected by the fuzzy clustering approach are key to better identifying outlying cells and correct them' depends on the fuzzy procedure reliably estimating those relationships even when cellwise contamination is present. No breakdown-point analysis or consistency result for the alternation between membership updates and cell corrections is supplied, leaving open the risk that early-iteration distance metrics biased by contamination produce self-reinforcing errors rather than accurate imputations.
  2. [Simulation study] Simulation study: While the effects of tuning parameters are explored, the description supplies no concrete performance metrics (e.g., adjusted Rand index, cellwise false-positive rates) or comparison against existing cellwise-robust or fuzzy methods under controlled contamination levels. This weakens the empirical support for the claim that the approach outperforms casewise alternatives without substantial information loss.
minor comments (2)
  1. [Abstract] The abstract is clear but would benefit from a one-sentence statement of the objective function or key update rules to allow readers to gauge technical novelty immediately.
  2. [Tuning parameters] Guidance on setting the tuning parameters for fuzziness and outlier detection is provided, yet explicit default values or a data-driven selection procedure would improve usability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the revisions we plan to incorporate.

read point-by-point responses

  1. Referee: [Abstract / Algorithm] Abstract and method description: The central claim that 'the cluster-specific relationships among variables detected by the fuzzy clustering approach are key to better identifying outlying cells and correct them' depends on the fuzzy procedure reliably estimating those relationships even when cellwise contamination is present. No breakdown-point analysis or consistency result for the alternation between membership updates and cell corrections is supplied, leaving open the risk that early-iteration distance metrics biased by contamination produce self-reinforcing errors rather than accurate imputations.

    Authors: We acknowledge the absence of a formal breakdown-point analysis or consistency result for the iterative alternation. The proposed method relies on an alternating optimization scheme in which fuzzy memberships and cell corrections are updated sequentially, with the simulations demonstrating stable recovery of cluster structures and variable relationships across contamination levels. To address the concern, we will add a discussion subsection on the iterative procedure, its initialization, and the empirical safeguards against self-reinforcing errors, while clarifying that the contribution is primarily methodological and simulation-supported rather than theoretical. revision: yes

  2. Referee: [Simulation study] Simulation study: While the effects of tuning parameters are explored, the description supplies no concrete performance metrics (e.g., adjusted Rand index, cellwise false-positive rates) or comparison against existing cellwise-robust or fuzzy methods under controlled contamination levels. This weakens the empirical support for the claim that the approach outperforms casewise alternatives without substantial information loss.

    Authors: We agree that explicit quantitative metrics and direct comparisons would strengthen the empirical evidence. The current simulation section explores tuning-parameter effects and illustrates performance, but we will revise it to report adjusted Rand index values for clustering accuracy, cellwise false-positive and false-negative rates for outlier detection, and comparisons against representative cellwise-robust and fuzzy clustering baselines under controlled contamination scenarios. These additions will provide clearer support for the advantages over casewise approaches. revision: yes

Circularity Check

0 steps flagged

No significant circularity; proposal integrates existing robust cellwise and fuzzy clustering concepts without reducing claims to inputs by construction.

full rationale

The paper proposes a new algorithm for robust fuzzy clustering under cellwise contamination, alternating between membership estimation and cell correction using cluster-specific relationships. No equations or steps in the abstract or described method show a result defined in terms of itself, a fitted parameter renamed as a prediction, or a central claim justified solely by overlapping self-citation. The approach is presented as building on prior robust statistics literature with new integration, validated via simulations and real data, rendering the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; tuning parameters for fuzziness and outlier detection are mentioned but not specified.

free parameters (1)
  • tuning parameters for fuzziness and outlier detection
    Effects explored in the work; specific forms or values not detailed in abstract.

pith-pipeline@v0.9.0 · 5776 in / 1015 out tokens · 46629 ms · 2026-05-19T00:46:21.720697+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Cellwise Outliers

    stat.ME 2026-03 unverdicted novelty 2.0

    Cellwise outliers can contaminate over half the cases even at low proportions, necessitating specialized robust techniques for location, covariance, regression, PCA, and tensor data that differ from casewise approaches.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    J. MacQueen, Some methods for classification and analysis of multivari- ate observations, in: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Statistics, Univer- sity of California Press, Berkeley, Calif., 1967, pp. 281–297

    work page 1967

  2. [2]

    G. H. Ball, D. J. Hall, A clustering technique for summarizing multi- variate data, Syst. Res. 12 (1967) 153–155

    work page 1967

  3. [3]

    G. J. McLachlan, D. Peel, Finite mixture models, Wiley, New York, 2000

    work page 2000

  4. [4]

    Bezdek, Pattern recognition with fuzzy objective function algorithms, Plenum Press, New York, 1981

    J. Bezdek, Pattern recognition with fuzzy objective function algorithms, Plenum Press, New York, 1981

    work page 1981

  5. [5]

    J. C. Dunn, A fuzzy relative of the ISODATA process and its use in detecting compact well-separated clusters, J. Cybernet. 3 (3) (1973) 32–57. 34

    work page 1973

  6. [6]

    D. E. Gustafson, W. C. Kessel, Fuzzy clustering with a fuzzy covariance matrix, in: Proceedings of the IEEE lnternational Conference on Fuzzy Systems, San Diego, 1979, p. 761–766

    work page 1979

  7. [7]

    Trauwaert, L

    E. Trauwaert, L. Kaufman, P. Rousseeuw, Fuzzy clustering algorithms based on the maximum likelihood priciple, Fuzzy Sets Syst. 42 (2) (1991) 213–227

    work page 1991

  8. [8]

    P. J. Rousseeuw, E. Trauwaert, L. Kaufman, Fuzzy clustering using scatter matrices, Comput. Stat. Data Anal. 23 (1) (1996) 135–151

    work page 1996

  9. [9]

    P. J. Rousseeuw, E. Trauwaert, L. Kaufman, Fuzzy clustering with high contrast, J. Comput. Appl. Math. 64 (1) (1995) 81–90

    work page 1995

  10. [10]

    P. J. Huber, Robust estimation of a location parameter, Ann. Math. Stat. 35 (1) (1964) 73–101

    work page 1964

  11. [11]

    L. A. Garc´ ıa-Escudero, A. Gordaliza, C. Matr´ an, A. Mayo-´Iscar, A gen- eral trimming approach to robust cluster analysis, Ann. Stat. 36 (3) (2008) 1324–1345

    work page 2008

  12. [12]

    Fritz, L

    H. Fritz, L. A. Garc´ ıa-Escudero, A. Mayo- ´Iscar, Robust constrained fuzzy clustering, Inf. Sci. 245 (2013) 38–52

    work page 2013

  13. [13]

    P. J. Rousseeuw, Least median of squares regression, J. Am. Stat. Assoc. 79 (388) (1984) 871–880

    work page 1984

  14. [14]

    P. J. Rousseeuw, Multivariate estimation with high breakdown point, in: W. Grossmann, G. Pflug, I. Vincze, W. Wertz (Eds.), Mathematical Statistics and Applications, 1985, pp. 283–297

    work page 1985

  15. [15]

    R. N. Dave, Characterization and detection of noise in clustering, Pat- tern Recognit. 12 (11) (1991) 657–664

    work page 1991

  16. [16]

    Alqallaf, S

    F. Alqallaf, S. Van Aelst, V. J. Yohai, R. H. Zamar, Propagation of outliers in multivariate data, Ann. Stat. 37 (1) (2009) 311–331

    work page 2009

  17. [17]

    Raymaekers, P

    J. Raymaekers, P. J. Rousseeuw, The cellwise minimum covariance de- terminant estimator, J. Am. Stat. Assoc. 119 (548) (2023) 2610–2621

    work page 2023

  18. [18]

    D. B. Rubin, Inference and missing data, Biometrika 63 (3) (1976) 581– 592. 35

    work page 1976

  19. [19]

    A. P. Dempster, N. M. Laird, D. B. Rubin, Maximum likelihood from incomplete data via the EM algorithm, J. R. Stat. Soc., Series B (Sta- tistical Methodology) 39 (1) (1977) 1–38

    work page 1977

  20. [20]

    Zaccaria, L

    G. Zaccaria, L. Garc´ ıa-Escudero, F. Greselin, A. Mayo- ´Iscar, Cellwise outlier detection in heterogeneous populations, Technometrics (2025) 1–16doi:10.1080/00401706.2025.2497822

    work page doi:10.1080/00401706.2025.2497822 2025

  21. [21]

    Puchhammer, I

    P. Puchhammer, I. Wilms, P. Filzmoser, A smooth multi-group Gaussian Mixture Model for cellwise robust covariance estimation, arXiv (2025) https://doi.org/10.48550/arXiv.2504.02547

    work page doi:10.48550/arxiv.2504.02547 2025

  22. [22]

    Raymaekers, P

    J. Raymaekers, P. J. Rousseeuw, Challenges of cell- wise outliers, Econometrics and Statistics (2024) https://doi.org/10.1016/j.ecosta.2024.02.002

    work page doi:10.1016/j.ecosta.2024.02.002 2024

  23. [23]

    Ghahramani, M

    Z. Ghahramani, M. Jordan, Learning from incomplete data, Tech. Rep. AI Lab Memo No. 1509, CBCL Paper No. 108, MIT AI Lab (1995)

    work page 1995

  24. [24]

    Fritz, L

    H. Fritz, L. A. Garc´ ıa-Escudero, A. Mayo- ´Iscar, A fast algorithm for robust constrained clustering, Comput. Stat. Data Anal. 61 (2013) 124– 136

    work page 2013

  25. [25]

    Hampel, Beyond location parameters: robust concepts and methods, Bull

    F. Hampel, Beyond location parameters: robust concepts and methods, Bull. Int. Stat. Inst. 46 (1) (1975) 375–382

    work page 1975

  26. [26]

    Hennig, T

    C. Hennig, T. Liao, How to find an appropriate clustering for mixed- type variables with application to socio-economic stratification, J. R. Stat. Soc., C: Appl. Stat. 62 (3) (2013) 309–369

    work page 2013

  27. [27]

    Robust fuzzy clustering with cellwise outliers

    L. Garc´ ıa-Escudero, A. Mayo-Iscar, Robust clustering based on trim- ming, Wiley Interdiscip. Rev. Comput. Stat. 16 (4) (2024) e1658. 36 Supplementary Material to “Robust fuzzy clustering with cellwise outliers” This document includes the supplementary material to the main article “Robust fuzzy clustering with cellwise outliers”. Specifically, it contain...

    work page 2024

  28. [28]

    As also for the first two variables, the choice of α = 0.05, which corresponds to the true level of contamination in the data, is confirmed by the ∆ plots

    Additional results on the effects of the tuning parameters In Figure 18, we display the ∆ ij values with α = 0.05 for the last three variables of the artificial data set, whose generation is detailed in the main article (first example). As also for the first two variables, the choice of α = 0.05, which corresponds to the true level of contamination in the...

    work page 2000

  29. [29]

    Body fat data set The preliminary analysis described in Section 4.1 of the main article on the body fat data set allows us to choose the fuzzifier parameter m

    Additional results for the real data analyses 7.1. Body fat data set The preliminary analysis described in Section 4.1 of the main article on the body fat data set allows us to choose the fuzzifier parameter m. Specif- ically, we select m by examining the fuzzification obtained by cellFCLUST. Recalling that we select c = 2 to avoid obtaining only one elon...

    work page