📊 statistics

Cellwise Outliers

This paper reviews recent advancements in detecting and handling cellwise outliers—individual erroneous values within data matrices that can contaminate entire cases—across various statistical methods including estimation, regression, and principal component analysis, highlighting their growing dominance over traditional casewise approaches in high-dimensional settings.

Original authors: Mia Hubert, Jakob Raymaekers, Peter J. Rousseeuw

Published 2026-04-17

📖 6 min read🧠 Deep dive

CC0 1.0

Original authors: Mia Hubert, Jakob Raymaekers, Peter J. Rousseeuw

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The "Bad Apple" vs. The "Bad Bite": A Guide to Cellwise Outliers

Imagine you are a chef preparing a massive banquet for 1,000 guests. You have a giant spreadsheet listing every ingredient for every dish: flour, sugar, eggs, spices, etc.

The Old Way: Throwing Away the Whole Dish (Casewise Outliers)

For decades, statisticians treated data like this: If one guest's dish had a mistake, they assumed the entire dish was ruined.

The Scenario: If the "Eggs" column for Guest #42 said "1,000 eggs" instead of "2," the old method would say, "Oh no, Guest #42 is a disaster! Let's throw away their entire order (all their ingredients) and ignore them completely."
The Problem: What if Guest #42 only messed up the eggs, but their flour, sugar, and spices were perfect? By throwing away the whole order, you lose a lot of good information. In high-dimensional data (where you have hundreds of variables), this is like throwing away a whole pizza just because one slice has a burnt crust.

The New Way: Fixing Just the Bad Bite (Cellwise Outliers)

This paper introduces a new way of thinking called Cellwise Outliers. Instead of looking at the whole row (the whole guest/dish), we look at individual cells (individual ingredients).

The Scenario: We see the "1,000 eggs" error. Instead of tossing the whole order, we just say, "Hey, the eggs are wrong. Let's fix that one cell or ignore just that one ingredient, but keep the flour and sugar."
The Challenge: This is much harder. If you have 100 variables, even a tiny percentage of bad ingredients can ruin most of your guests' orders. If 5% of your ingredients are wrong, and you have 20 ingredients per dish, statistically, almost every single guest will have at least one bad ingredient. The old "throw away the whole dish" method fails here because you'd end up throwing away almost everyone.

The Core Problem: Why Can't We Just Use the Old Tools?

The authors explain that the mathematical rules we used for the "Old Way" (Casewise) don't work for the "New Way" (Cellwise).

The Analogy of the Rigid Ruler:
Imagine you have a rigid ruler that must stay perfectly straight no matter how you rotate it (this is called equivariance in math).

The Old Rule: If you have a straight line of data, the ruler must stay on that line.
The Cellwise Trap: If you have a straight line of data, but someone sneaks in a few "bad bites" (outliers) in different spots, the rigid ruler snaps. To fix the bad bites without breaking the ruler, you have to let go of the rule that the ruler must stay perfectly straight in every situation. You have to be flexible.

The paper proves that to handle these "bad bites," we have to abandon some of our favorite, intuitive mathematical rules. It's painful, but necessary.

The Solutions: How Do We Fix It?

The paper reviews a decade of progress in building new tools that can handle these bad bites. Here are the main characters in this story:

1. The Detective (DDC - Detect Deviating Cells)

Before we can fix the data, we need to find the bad bites.

How it works: Imagine you are checking a car's specs. If a car is listed as having a tiny width but a massive top speed, that's suspicious. The DDC algorithm acts like a detective. It looks at how variables relate to each other. If "Width" usually goes with "Top Speed," and a car breaks that pattern, the detective flags that specific cell as suspicious, even if the width number looks normal on its own.
The Visual: They use a "Cellmap" (like a heat map). Blue cells are lower than expected, red cells are higher. It highlights the bad bites in a sea of good data.

2. The Surgeon (CellMCD & CellLTS)

Once we find the bad bites, we need to estimate the "true" average and spread of the data without letting the bad bites distort the results.

CellMCD: This method acts like a surgeon. It tries to find the largest group of "clean" cells that fit together perfectly. It ignores the bad bites and calculates the average based only on the healthy ones.
CellLTS (The Prediction Machine): This is special for making predictions. Imagine you are predicting a car's price based on its features. A new car comes in, but its "Acceleration" number is a typo (a bad bite).
- Old Way: You use the typo, and your prediction is wild.
- CellLTS: It says, "Wait, that acceleration number is weird. Let's guess what it should be based on the other features, fix it, and then make the prediction." It imputes (fills in) the bad data before calculating the result.

3. The High-Dimensional Hero (For Massive Data)

When you have thousands of variables (like DNA data or image pixels), the old methods get slow and confused.

The Solution: New methods like FastDDC and CellPCA use clever shortcuts. They don't look at every single relationship; they use "wrapping" techniques and fast screening to find the bad bites quickly, even in data with millions of dimensions.
Tensor Data (The 3D Block): Sometimes data isn't a flat sheet (2D) but a 3D block (like a video, which is frames x height x width). The paper discusses ROMPCA, which treats a video frame like a 3D block. If a few pixels in the video are glitching (bad cells), ROMPCA fixes just those pixels without throwing away the whole video frame.

Why Does This Matter?

In the past, if your data had a few messy numbers, you might have had to throw away half your dataset to get a clean result. That's wasteful.

This paper shows us that we can be smart and surgical. We can:

Identify exactly which numbers are wrong.
Fix or ignore only those specific numbers.
Keep all the good information from the rest of the row.

This allows us to analyze messy, real-world data (which is full of errors) without losing the valuable insights hidden inside. It's the difference between throwing away a whole apple because of one bruise, versus just cutting out the bruise and eating the rest.

1. Problem Definition and Context

The Shift from Casewise to Cellwise Contamination
Traditional robust statistics operates under the casewise contamination model (Tukey-Huber model), where an entire observation (row/case) is either clean or entirely corrupted. If a case is an outlier, all its variables are treated as suspect, leading to the deletion or downweighting of the entire case.

However, in modern high-dimensional datasets, this assumption is often unrealistic. The paper introduces the cellwise contamination model, where individual values (cells) within the data matrix can be corrupted independently.

The "Curse of Dimensionality" in Contamination: Even a small proportion of outlying cells ( $\epsilon_{cell}$ ) can contaminate the majority of cases as the number of variables ( $d$ ) increases. The probability that a case contains at least one outlying cell is $1 - (1 - \epsilon_{cell})^d$ . For example, with $\epsilon_{cell} = 0.05$ and $d=70$ , over 97% of cases will contain at least one outlier.
Limitations of Casewise Methods: Applying casewise robust methods to cellwise contaminated data leads to excessive data loss. A case with only one bad cell is discarded entirely, wasting the information contained in its other clean cells.
Mixed Contamination: Real-world data often involves a mix of casewise outliers, cellwise outliers, and missing values (NA).

2. Methodological Framework

The paper reviews a decade of progress in developing methods that handle cellwise outliers without discarding entire cases. A central theme is the trade-off between robustness and equivariance.

2.1 Theoretical Breakdown and Equivariance

The authors demonstrate that standard properties desirable in casewise robustness are incompatible with cellwise robustness:

Affine Equivariance: Standard estimators (like MCD) are affine equivariant. However, the paper proves that any location or covariance estimator satisfying affine equivariance (or even weaker properties like orthogonal equivariance) has a cellwise breakdown value bounded by approximately $1/d$ .
Implosion vs. Explosion: For covariance matrices, "implosion" (singularity) is a critical failure mode. The paper shows that affine equivariant estimators have a very low cellwise implosion breakdown value ( $\approx 1/d$ ), meaning they fail quickly when even a few cells are corrupted.
Paradigm Shift: To achieve high cellwise breakdown values, methods must abandon strict affine equivariance. Instead, they often rely on coordinate-wise robustness or specific weighting schemes that do not transform naturally under all linear transformations.

2.2 Detection of Cellwise Outliers

Marginal vs. Multivariate: Simple marginal filtering (e.g., robust z-scores) fails to detect outliers that are only anomalous in a multivariate context (e.g., a car with normal width but inconsistent with its speed/acceleration).
DDC (Detect Deviating Cells): A key algorithm introduced by Rousseeuw and Van den Bossche (2018). It fits robust simple linear regressions of each variable against all others. The residuals from these regressions identify cells that deviate from the multivariate structure.
FastDDC: A computationally efficient version of DDC for high-dimensional data using nearest-neighbor screening.

2.3 Estimation of Location and Covariance

Two-Step Approaches: Early methods (e.g., TSGS) filtered outliers first (setting them to missing) and then applied casewise robust estimators (like Generalized S-estimators) on the cleaned data.
CellMCD: A major contribution is the Cellwise Minimum Covariance Determinant (CellMCD). Unlike the standard MCD which selects a subset of cases, CellMCD selects a subset of cells. It minimizes an objective function combining the log-likelihood of clean cells and a penalty for the number of flagged cells. It iteratively updates the set of clean cells and the covariance estimate.
Weighted Approaches: Recent work moves beyond hard-thresholding (discarding cells) to soft-weighting, where cells receive weights in $[0, 1]$ based on their deviation.

2.4 Regression

Cellwise Robust Regression: The goal is to use clean cells of a regressor vector while downweighting contaminated ones.
Shooting S-Estimator: Uses coordinate descent, fitting one slope at a time using casewise robust S-estimators, while modifying the response variable to prevent outlier propagation.
CellLTS: A method that first imputes cellwise outliers in the predictor matrix $X$ (using CellMCD logic) and then applies Least Trimmed Squares (LTS) regression. Crucially, it handles out-of-sample prediction by imputing outliers in new data vectors before applying the model, preventing the "red arrow" effect where a single bad cell skews the prediction.

2.5 High-Dimensional Data and PCA

Cellwise Robust PCA (cellPCA): Standard PCA fails because projecting a case with an outlier propagates the error to all dimensions. CellPCA uses a single objective function combining two bounded loss functions: one for cellwise residuals and one for casewise total deviation. It provides imputed data and weights.
Regularization: For $d > n$ , methods like CellRCov decompose the covariance matrix into a low-rank component (estimated via cellPCA) and a diagonal/regularized residual component.
Precision Matrix: Robust precision matrices are estimated by plugging robust covariance estimates (based on pairwise correlations) into the Graphical Lasso objective.

2.6 Tensor Data

Tensor Decompositions: Extends robustness to 3D+ arrays (e.g., video frames).
ROMPCA: A method for Multilinear PCA that handles both casewise and cellwise outliers simultaneously using bounded $\rho$ functions at both the case and cell levels. It provides imputed tensors and weights.
Applications: Demonstrated on video surveillance data (Dog Walker dataset), successfully flagging the moving subject as the source of cellwise outliers while ignoring background noise.

3. Key Contributions

Theoretical Limits: Rigorous proof that affine equivariance is incompatible with high cellwise breakdown values, necessitating a shift in statistical paradigm.
Algorithmic Development: Introduction and refinement of CellMCD, CellLTS, cellPCA, and ROMPCA, which integrate outlier detection directly into the estimation process rather than treating it as a separate preprocessing step.
Handling Missing Data: These methods naturally handle missing values (NA) by treating them similarly to flagged outliers during the estimation process, a significant advantage in real-world data.
Out-of-Sample Prediction: Development of mechanisms (specifically in CellLTS and ROMPCA) to robustly predict on new data containing outliers, a feature often missing in traditional robust methods.
Unified Framework: The paper synthesizes methods for location, covariance, regression, PCA, tensors, and other settings (clustering, time series, functional data) under the cellwise paradigm.

4. Results and Performance

Breakdown Values: The proposed methods (like CellMCD and CellLTS) achieve cellwise breakdown values significantly higher than $1/d$ , often approaching the casewise breakdown limit depending on the tuning parameters.
Simulation and Real Data:
- In the TopGear dataset, DDC successfully identified a car with a "small width" that was not an outlier marginally but was inconsistent with its speed and acceleration.
- In the Octane dataset (high-dimensional NIR spectra), cellPCA produced residual maps that clearly highlighted outliers without distorting the principal subspace.
- In the Dog Walker video, ROMPCA isolated the moving person as the source of cellwise outliers in the tensor data, distinguishing them from wind-blown leaves.
Comparison: The methods generally outperform traditional casewise robust methods (which lose too much data) and simple marginal filtering (which misses multivariate outliers).

5. Significance

This paper marks a maturation of robust statistics for the era of big data.

Efficiency: By preserving clean cells within contaminated cases, these methods retain statistical power and efficiency that would otherwise be lost.
Applicability: The framework is essential for high-dimensional fields like genomics, chemometrics, and image analysis, where $d \gg n$ and data quality is often variable.
Software Ecosystem: The authors highlight the availability of these methods in the R package cellWise, facilitating adoption by the broader statistical and machine learning community.

In conclusion, the paper argues that while cellwise robustness requires sacrificing some classical mathematical properties (like affine equivariance), the resulting methods are necessary and superior for analyzing modern, complex, and imperfect datasets.