Robust Feature Attribution for Redundant and Correlated Data

This direction is merged into another direction: feature independence assumption (feature correlation bias). Its last version stays readable and citable here.

Version 1 of 150 papers60 conceptsfrozen 2026-10-04

When we ask a machine learning model "why" it made a decision, we usually expect a ranking of features by their importance. However, most standard tools for this—like Permutation Importance or Mean Decrease Impurity—rely on a fragile assumption: that features are independent. In real-world datasets, from genomics to finance, features are often highly correlated. This redundancy creates a "masking" effect that can make even the most critical drivers of a model appear irrelevant.

The Failure of Marginal Importance

The intuition behind most feature importance metrics is simple: if you remove or shuffle a feature and the model's performance drops, that feature must be important. But as João P. B. Pereira and colleagues point out in Covered Information Disentanglement: Model Transparency via Unbiased Permutation Importance, this logic breaks down when features "cover" for one another.

Imagine two sensors measuring the same physical phenomenon. If you shuffle the data from Sensor A, the model can simply rely on the redundant information in Sensor B to maintain its accuracy. Consequently, both sensors might be assigned an importance score of near zero, even if the physical phenomenon they measure is the primary driver of the prediction. This "dilution" of signal is a recurring theme in the work of Giles Hooker and Lucas Mentch, who argue in Bridging Breiman's Brook that these heuristic measures produce biased results that hinder scientific inference.

From Marginal to Conditional Impact

To solve this, researchers have shifted from measuring marginal importance (how much a feature helps on its own) to conditional importance (how much unique information a feature provides that isn't already captured by others).

David Watson and Marvin N. Wright formalized this in Testing conditional independence in supervised learning algorithms. Their framework, cpi: Conditional Predictive Impact, uses "knockoff" variables—synthetic counterparts that mimic the correlation structure of the original data but are independent of the target. By comparing a feature to its knockoff, researchers can determine if a variable adds true predictive value.

However, even conditional methods have limits. Annie Liang and Richard Baumgartner demonstrated in Challenges in Variable Importance Ranking Under Correlation that when correlation exceeds 0.5, these methods can suffer from "importance deflation," where the similarity between the original and the knockoff makes it impossible to detect a signal.

Disentangling the Latent Space

A more recent and robust direction involves "untangling" the correlations before measuring importance. Instead of looking at the raw features, methods like Disentangled Feature Importance (DFI) map the data into an independent latent space.

In Disentangled Feature Importance, Jin-Hong Du and Larry A. Wasserman use optimal transport to create this mapping. By calculating importance in a space where variables are independent and then attributing that importance back to the original features, they avoid the "correlation distortion" that plagues standard methods. Similarly, Flow-Disentangled Feature Importance (FDFI) uses generative flow matching to achieve a similar disentanglement, providing statistically grounded p-values for feature relevance.

The Power of Grouping

When individual features are so tightly coupled that they cannot be meaningfully separated, the most honest approach may be to stop trying. Ahmad Chamma and Bertrand Thirion argue in Variable Importance in High-Dimensional Settings Requires Grouping that we should measure the importance of groups of variables.

Their BCPI (Block-Based Conditional Permutation Importance) framework identifies the collective signal of feature clusters. This is visualized effectively in the Triplot method by Przemyslaw Biecek, which combines hierarchical clustering with group-level importance. This prevents the "dilution" effect by showing that while individual features might look weak, the concept they represent is a primary driver of the model.

Adversarial Explanations and X-Hacking

Understanding these correlation failures isn't just a matter of mathematical curiosity—it is a security necessity. Eyke Hüllermeier and Rahul Sharma have identified a threat they call X-hacking (explanation hacking).

In X Hacking: The Threat of Misguided AutoML, they show that because many different models can achieve the same accuracy (the Rashomon Effect), an analyst can deliberately search for a model that hides a sensitive attribute (like race) by shifting its importance onto a correlated proxy (like zip code). Without robust, disentangled importance measures, these "fair" explanations can be easily manipulated to hide systemic biases.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_feature_attribution_for_redundant_dc1c4017,
  title = {Robust Feature Attribution for Redundant and Correlated Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-feature-attribution-for-redundant-and-correlated-data/rdr_dc1c401747084be89028a3775662c5a5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_feature_attribution_for_redundant_dc1c4017_v1,
  title = {Robust Feature Attribution for Redundant and Correlated Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-feature-attribution-for-redundant-and-correlated-data/rdr_dc1c401747084be89028a3775662c5a5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}