pith. sign in

arxiv: 2604.15106 · v1 · submitted 2026-04-16 · 📊 stat.ME

Cellwise Robust Twoblock Dimension Reduction

Sven Serneels This is my paper

Pith reviewed 2026-05-10 10:34 UTC · model grok-4.3

classification 📊 stat.ME
keywords cellwise robustnessdimension reductiontwoblock SVDrobust statisticsoutlier detectionimputationvariable selectionmultivariate analysis
0
0 comments X

The pith

CRTB provides the first cellwise robust approach to simultaneous dimension reduction for predictor and response blocks by imputing contaminated cells rather than discarding rows.

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

The paper introduces Cellwise Robust Twoblock (CRTB) as a method for dimension reduction that protects against outliers at the cell level instead of the row level. Classical approaches lose efficiency when contamination is spread across individual entries because they remove entire observations. CRTB uses a column-wise pre-filter to detect bad cells and then imputes them using model-based estimates inside an iteratively reweighted loop. This preserves information from partially clean rows and allows the procedure to function even when more than half the observations contain some contamination. Simulations show it maintains efficiency on clean data, accurately identifies outliers, and selects relevant variables in the sparse case.

Core claim

CRTB is the first cellwise robust method for simultaneous dimension reduction of multivariate predictor and response blocks, in both dense and sparse variants. It combines a column-wise pre-filter for cellwise outlier detection with model-based imputation of flagged cells inside an iteratively reweighted M-estimation loop. The algorithm uses the classical twoblock SVD as a warm start and converges quickly, retaining clean cells of partially contaminated rows instead of discarding the observation.

What carries the argument

The iteratively reweighted M-estimation loop that integrates column-wise cellwise outlier pre-filtering and model-based imputation to perform twoblock dimension reduction while retaining usable cells from contaminated rows.

If this is right

  • CRTB can handle contamination affecting more than 50% of rows without breakdown.
  • It recovers the cellwise outlier pattern with high fidelity from the data.
  • In sparse settings, it correctly identifies the informative variables.
  • The method provides interpretable results in domain-specific examples with cellwise outliers present.

Where Pith is reading between the lines

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

  • Similar pre-filter and imputation strategies could be adapted to other dimension reduction techniques like principal component analysis for cellwise robustness.
  • The approach may prove particularly useful in high-dimensional datasets where casewise deletion would remove too much data.
  • Further work could explore extensions to nonlinear or kernel-based twoblock methods.

Load-bearing premise

The column-wise pre-filter must correctly flag contaminated cells without too many errors, and the imputation step must preserve the underlying low-dimensional structure without bias.

What would settle it

A dataset with known cellwise contamination where more than 50 percent of rows are affected and CRTB fails to recover the true dimension reduction directions or misidentifies the outliers would contradict the claims.

Figures

Figures reproduced from arXiv: 2604.15106 by Sven Serneels.

Figure 1
Figure 1. Figure 1: MSE of the estimated coefficient matrix under cellwise contamination ( [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Variable selection F1 score for sparse methods under cellwise contamination. Left: [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cellwise outlier detection quality of CRTB’s column-wise pre-filter. Each panel [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative MSE increase as a function of cell contamination percentage ( [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Out-of-fold weighted MSE on the clean liver toxicity data. TB provides a baseline [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Held-out weighted MSE when 20% of Y cells are corrupted by one-sided 15×MAD shifts. Dashed line: TB trained on the clean Y . TB trained on contaminated Y collapses to wMSE≈0.18; CRTB dense recovers to within ∼0.013 of the clean reference under both prefilter and DDC initialisation. The results ( [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cellwise outlier map of Y . Rows (rats) are ordered by dose group (grouped blocks) and by time point within each dose group. Dark cells indicate CRTB Hampel weight < 0.5. The 20% cellwise contamination is detected with precision 0.800, recall 1.000, F1 = 0.889 at the row level against the high-dose × late-time ground truth. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Loadings of the ten clinical chemistry markers on the first (and only) Y-component [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: First two CRTB X-scores, coloured by dose group and sized by sacrifice time. The [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Top gene loadings on the first CRTB X-component (the liver-injury axis). Sparse [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: 97.5% tolerance ellipse fitted to the bivariate emissions block (CO, NOx) using a minimum covariance determinant (MCD) estimator on n = 5000 gas turbine hours. Crosses mark the nout = 1092 naturally-occurring outliers (21.8% of hours) arising from operational anomalies such as start-ups, shut-downs and off-design transients. 4.2.2 Results on clean and contaminated data Seven estimators are compared under … view at source ↗
Figure 12
Figure 12. Figure 12: Held-out weighted MSE (inverse-variance weights) on the clean gas-turbine test [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Number of cells flagged per case by the dense CRTB [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Per-case cellwise detection map for the two emissions columns, CRTB [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Precision, recall and F1 of CRTB Y-cell flagging against the tolerance-ellipse ground truth, evaluated per emissions column on the rows where per-column truth is defined. Recall is perfect (R = 1.000) for both columns; precision is 0.701 for CO and 0.789 for NOx. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗
read the original abstract

Cellwise Robust Twoblock (CRTB) is introduced, the first cellwise robust method for simultaneous dimension reduction of multivariate predictor and response blocks, in both a dense and a sparse variable-selecting variant. Classical robust methods protect against casewise outliers by downweighting or removing entire observations, a strategy that becomes inefficient -- and eventually breaks down -- when contamination is scattered across individual cells rather than concentrated in whole rows. CRTB combines a column-wise pre-filter for cellwise outlier detection with model-based imputation of flagged cells inside an iteratively reweighted M-estimation loop, retaining the clean cells of partially contaminated rows instead of discarding the observation. An efficient algorithm is provided that uses the classical twoblock SVD as a warm start and converges in a handful of IRLS iterations at a moderate computational cost. The method resists settings where more than $50\%$ of rows contain contaminated cells while retaining comparable efficiency on clean data. A simulation study confirms these properties and shows that CRTB additionally recovers the underlying cellwise outlier pattern with high fidelity and, in the sparse setting, the correct set of informative variables. Two compelling examples illustrate CRTB's practical utility. In each of these, CRTB is shown to be conducive to results that are highly interpretable in the respective domains in the presence of cellwise outliers. As a by-product, the corresponding cells are identified with high fidelity.

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

3 major / 2 minor

Summary. The manuscript introduces the Cellwise Robust Twoblock (CRTB) method for simultaneous dimension reduction of multivariate predictor and response blocks in both dense and sparse variants. It employs a column-wise pre-filter to detect cellwise outliers, followed by model-based imputation within an iteratively reweighted M-estimation framework that uses classical twoblock SVD as a warm start. The central claims are that CRTB achieves a breakdown point exceeding 50% with respect to the proportion of rows containing contaminated cells, maintains efficiency on clean data, recovers the cellwise outlier pattern with high fidelity, and in the sparse case identifies the correct informative variables. These are supported by a simulation study and two real-data examples demonstrating practical utility and interpretability.

Significance. Should the method's robustness properties and recovery performance be rigorously established, this would constitute a significant contribution to the field of robust multivariate statistics. By addressing cellwise rather than casewise contamination, CRTB enables more efficient use of data in settings where outliers are scattered across observations, which is common in modern high-dimensional applications. The efficient algorithm and dual dense/sparse variants enhance its applicability.

major comments (3)
  1. §3 (Method description): The reliance on a column-wise pre-filter for cellwise outlier detection ignores potential correlations within and between the predictor and response blocks. This is a load-bearing assumption for the imputation step and the claimed breakdown point, as misflagged cells could bias the twoblock SVD estimates. The manuscript should either provide a theoretical justification or additional simulations under correlated designs to validate this.
  2. §4 (Algorithm): No convergence analysis or proof of the breakdown point is provided for the full IRLS procedure. The claims appear to rest on the pre-filter's accuracy and the warm-start strategy, but without formal results, it is difficult to assess whether the >50% resistance holds in general.
  3. Simulation study (Section 6): The simulation study reports high-fidelity recovery, but lacks details on how the data generation incorporates the twoblock structure and correlations; this makes it hard to evaluate if the results support the general claims for both dense and sparse settings.
minor comments (2)
  1. Introduction: Some references to related work on cellwise robust methods could be expanded for better context.
  2. Notation section: Clarify the dimensions of the predictor and response matrices early on to aid readability.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive and insightful comments on our manuscript. We address each major comment below, indicating planned revisions where appropriate. Our responses focus on clarifying the methodological choices and strengthening the empirical support.

read point-by-point responses

  1. Referee: §3 (Method description): The reliance on a column-wise pre-filter for cellwise outlier detection ignores potential correlations within and between the predictor and response blocks. This is a load-bearing assumption for the imputation step and the claimed breakdown point, as misflagged cells could bias the twoblock SVD estimates. The manuscript should either provide a theoretical justification or additional simulations under correlated designs to validate this.

    Authors: The column-wise pre-filter is intentionally marginal to enable scalable cellwise detection without requiring full joint modeling at the detection stage. However, the subsequent model-based imputation explicitly uses the twoblock SVD, which incorporates correlations both within and between the predictor and response blocks. This two-stage structure allows the method to leverage joint information after initial flagging. To empirically address concerns about correlated designs, we will add a new set of simulations with varying correlation structures in the revised manuscript. revision: yes

  2. Referee: §4 (Algorithm): No convergence analysis or proof of the breakdown point is provided for the full IRLS procedure. The claims appear to rest on the pre-filter's accuracy and the warm-start strategy, but without formal results, it is difficult to assess whether the >50% resistance holds in general.

    Authors: We agree that a rigorous convergence analysis and breakdown-point proof for the complete IRLS procedure would be desirable. Deriving such formal guarantees for this specific combination of pre-filtering, imputation, and twoblock M-estimation is technically involved and falls outside the primary scope of the present work, which emphasizes algorithmic development and practical performance. The >50% resistance claim is supported by extensive Monte Carlo experiments across diverse contamination levels. We will revise the manuscript to include an explicit discussion of the empirical nature of these robustness results and the role of the warm start. revision: partial

  3. Referee: Simulation study (Section 6): The simulation study reports high-fidelity recovery, but lacks details on how the data generation incorporates the twoblock structure and correlations; this makes it hard to evaluate if the results support the general claims for both dense and sparse settings.

    Authors: We will expand Section 6 with a more detailed description of the data-generating process. This will explicitly document how the twoblock low-rank structure, block-wise correlations, and sparse variable selection are implemented for both the dense and sparse variants, thereby clarifying how the reported recovery performance relates to the general claims. revision: yes

standing simulated objections not resolved
  • No formal convergence analysis or proof of the breakdown point for the full IRLS procedure

Circularity Check

0 steps flagged

No significant circularity; algorithm uses standard IRLS warm-start with independent simulation validation

full rationale

The provided abstract and description present CRTB as an algorithmic combination of a column-wise pre-filter for cellwise outlier detection followed by model-based imputation inside a standard IRLS loop that starts from the classical twoblock SVD. Performance claims (resistance to >50% row contamination, outlier pattern recovery, variable selection in sparse case) are stated to be confirmed by a separate simulation study rather than derived by construction from the fitted parameters themselves. No equations, self-citations, or steps are quoted that reduce a central prediction or uniqueness claim to a fitted input or prior self-result. The derivation chain is therefore self-contained against external benchmarks and does not match any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review yields limited visibility into assumptions; the method implicitly relies on standard M-estimation regularity conditions and on the effectiveness of the column-wise pre-filter plus imputation step.

axioms (2)
  • domain assumption Standard regularity conditions for M-estimators and IRLS convergence hold.
    Required for the iteratively reweighted loop to produce a stable solution.
  • ad hoc to paper The column-wise pre-filter identifies cellwise outliers with sufficient accuracy that imputation does not introduce systematic bias into the dimension reduction.
    Central to the claim that partially contaminated rows can be retained rather than discarded.

pith-pipeline@v0.9.0 · 5536 in / 1432 out tokens · 27972 ms · 2026-05-10T10:34:59.170828+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    J., and Zamar, R

    Alqallaf, F., Van Aelst, S., Yohai, V. J., and Zamar, R. H. (2009). Propagation of outliers in multivariate data.The Annals of Statistics, 37(1):311–331. 25

    work page 2009

  2. [2]

    R., Wolfinger, R

    Bushel, P. R., Wolfinger, R. D., and Gibson, G. (2007). Simultaneous clustering of gene expression data with clinical chemistry and pathological evaluations reveals phenotypic prototypes.BMC Systems Biology, 1:15

    work page 2007

  3. [3]

    Centofanti, F., Hubert, M., and Rousseeuw, P. J. (2026). Robust principal components by casewise and cellwise weighting.Technometrics, (just-accepted):1–25

    work page 2026

  4. [4]

    D., Forzani, L., and Liu, L

    Cook, R. D., Forzani, L., and Liu, L. (2023). Partial least squares for simultaneous reduc- tion of response and predictor vectors in regression.Journal of Multivariate Analysis, 196:105163

    work page 2023

  5. [5]

    Debruyne, M., Höppner, S., Serneels, S., and Verdonck, T. (2019). Outlyingness: Which variables contribute most?Statistics and Computing, 29(4):707–723

    work page 2019

  6. [6]

    Filzmoser, P., Höppner, S., Ortner, I., Serneels, S., and Verdonck, T. (2020). Cellwise robust M regression.Computational Statistics & Data Analysis, 147:106944

    work page 2020

  7. [7]

    R., Ronchetti, E

    Hampel, F. R., Ronchetti, E. M., Rousseeuw, P. J., and Stahel, W. A. (1986).Robust Statistics: The Approach Based on Influence Functions. Wiley Series in Probability and Statistics. John Wiley & Sons

    work page 1986

  8. [8]

    J., and Van den Bossche, W

    Hubert, M., Rousseeuw, P. J., and Van den Bossche, W. (2019). Macropca: An all-in-one pca method allowing for missing values as well as cellwise and rowwise outliers.Technometrics, 61(4):459–473

    work page 2019

  9. [9]

    Kaya, H., Tüfekci, P., and Uzun, E. (2019). Predicting CO and NOx emissions from gas turbines: Novel data and a benchmark PEMS.Turkish Journal of Electrical Engineering and Computer Sciences, 27(6):4783–4796. Dataset available as UCI Machine Learning Repository #551

    work page 2019

  10. [10]

    Leung, A., Zhang, H., and Zamar, R. (2016). Robust regression estimation and inference in the presence of cellwise and casewise contamination.Computational Statistics & Data Analysis, 99:1–11

    work page 2016

  11. [11]

    Maronna, R. A. and Zamar, R. H. (2002). Robust estimates of location and dispersion for high-dimensional datasets.Technometrics, 44(4):307–317

    work page 2002

  12. [12]

    Pfeiffer, P., Vana-Gür, L., and Filzmoser, P. (2025). Cellwise robust and sparse principal component analysis.Advances in Data Analysis and Classification, pages 1–30

    work page 2025

  13. [13]

    and Rousseeuw, P

    Raymaekers, J. and Rousseeuw, P. J. (2024). Challenges of cellwise outliers.Econometrics and Statistics. In press. Preprint available athttps://arxiv.org/abs/2302.02156

    work page arXiv 2024

  14. [14]

    Rousseeuw, P. J. (1984). Least median of squares regression.Journal of the American Statistical Association, 79(388):871–880

    work page 1984

  15. [15]

    Rousseeuw, P. J. and Van den Bossche, W. (2018). Detecting deviating data cells.Techno- metrics, 60(2):135–145. 26

    work page 2018

  16. [16]

    Serneels, S. (2025). Sparse twoblock dimension reduction: A versatile alternative to sparse PLS2 and CCA.Journal of Chemometrics, 39:e70051

    work page 2025

  17. [17]

    Serneels, S. (2026). Robust twoblock dimension reduction. Submitted for publication, preprint available athttps://arxiv.org/abs/2603.24820v1

    work page arXiv 2026

  18. [18]

    Serneels, S., Croux, C., Filzmoser, P., and Van Espen, P. J. (2005). Partial robust M- regression.Chemometrics and Intelligent Laboratory Systems, 79(1–2):55–64

    work page 2005

  19. [19]

    Wold, H. (1966). Nonlinear estimation by iterative least squares procedures. In David, F., editor,Papers in Statistics: Festschrift for J. Neyman, pages 411–444. Wiley

    work page 1966

  20. [20]

    Yao, F., Coquery, J., and Lê Cao, K.-A. (2012). Independent principal component anal- ysis for biologically meaningful dimension reduction of large biological data sets.BMC Bioinformatics, 13:24. 27

    work page 2012