acceptodds
Under review as a conference paper at ICLR 2027

A Paired Case Study of Conditional and Marginal Feature-Replacement Scores

Abstract

Feature-removal explanations measure how a predictor responds when an input feature is replaced. Because the predictor still requires a complete input, every such explanation depends on a replacement distribution. Conditional replacement is commonly motivated by its ability to preserve dependencies among observed features, but observational fidelity does not by itself determine how the predictor will respond to the resulting inputs. We study this gap with a paired evaluation that holds the predictor, evaluation examples, labels, and scoring rule fixed while changing the replacement distribution. The primary study compares conditional and marginal replacement on Dry Bean and evaluates coordinate prediction, covariance preservation, score magnitude, and feature ordering. Complementary studies examine stochastic draws, deterministic means, and label-informed replacements across several tabular domains. Conditional replacement more closely reproduces observed coordinate and covariance structure, yet it produces substantially different feature-removal scores and rankings from marginal replacement through the same predictors. The secondary comparisons further show that sensitivity to the replacement procedure depends on the dataset and predictor family. These findings establish the replacement distribution as part of the feature-removal estimand rather than a secondary implementation detail. A reproducible explanation should therefore report the replacement rule together with the predictor and scoring procedure.

Primary area:
interpretability and explainable AI
Keywords:
feature attribution, feature removal, conditional sampling, marginal sampling, explainable AI, model interpretation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_X1srOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how replacement distributions affect feature-removal explanations. The primary study compares ridge-based conditional Gaussian and moment-matched marginal Gaussian replacements on Dry Bean across three MLP runs on a common data split. Conditional replacement achieves lower coordinate-prediction negative log-likelihood and covariance discrepancy, yet produces substantially smaller loss-based importance scores and different feature rankings. Complementary experiments on CelebA, Adult, Spambase, and Bank Marketing examine accuracy sensitivity to stochastic draws, deterministic means, and label-informed replacements. The main contribution is a controlled empirical characterization of replacement sensitivity, showing that stronger observational fidelity can coexist with substantial disagreement between explanations.

Critical strengths

The observed differences affect substantive interpretation, beyond numerical score scales. On Dry Bean, conditional replacement achieves better coordinate-prediction and covariance diagnostics, yet conditional and marginal replacement select different top-ranked features in all three runs, with Kendall’s τ_b between 0.12 and 0.32. These results provide a concrete illustration that stronger performance on distributional diagnostics can coexist with substantial disagreement in feature prioritization. This is consequential for interpreting and reporting feature-removal explanations.

Critical weaknesses
  1. The novelty is limited The fact that conditional and marginal replacement define different estimands, and can therefore produce different importance scores for dependent features, is already recognized in the literature discussed by the paper. The paired study quantifies this distinction clearly, but provides limited additional analysis of the mechanisms or conditions governing the magnitude and ranking changes. Consequently, the manuscript’s advance beyond a controlled illustration of an established principle remains modest, which limits the strength of its contribution.

  2. The direct evidence for the central finding is narrow. The conditional–marginal comparison uses one Dry Bean split, three MLP runs, and two Gaussian replacement constructions. The secondary experiments use different replacement contrasts and an accuracy endpoint, so they do not directly establish the robustness of the primary score and ranking findings. Moreover, the ranking results are point estimates, and the reported bootstrap conditions on fitted models and realized draws. These limitations leave uncertainty about how strongly the reported disagreements depend on the particular dataset, replacement approximations, and experimental realization.

Questions for authors

How stable are the top-ranked feature identities and Kendall’s τ_b values when increasing the Monte Carlo draw count or resampling the evaluation examples? Since ranking disagreement is central to the paper’s contribution, quantifying its uncertainty would help establish the strength of this finding.

Reviewer_CRjROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Feature-removal-based explanations measure how a predictor responds when an input feature is replaced. The work contrasts two common replacement techniques: conditional and marginal replacement. The paper provides strong metrics with clear explanations and motivation for their use. The authors also experiment on multiple datasets to validate their findings. Clarification is needed on certain parameter choices and confounds prior to acceptance.

Critical strengths
  • Robust evaluation across multiple datasets. The methods section clearly explains the workings and motivation behind each metric used.

  • The paper explores an important but often overlooked aspect of feature removal-based explanations.

Critical weaknesses
  • Certain choices of hyperparameters were not well motivated and appear arbitary. (I.e. why only look at K ∈\in {1,3,6} and not K inin {1,2,3,4,5,6}? Why is the standardized clip interval [5,5]?)

  • The broader impact of the findings needs further discussion.

Questions for authors
  • What exactly is a 16-32-ReLU-7 MLP?

  • Marginal replacements yield a higher NLL than conditional replacements. Hence, the distribution of the replacement varies more in case of marginal replacement. Would this also not lead to higher magnitudes of scores? How are we isolating the impact of the shift from the impact of the scores (i.e., the two questions discussed in lines 428- 429)?

Reviewer_iTAXOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how the choice of feature-replacement distribution affects perturbation-based feature importance explanations. The authors compare conditional replacement, which preserves dependencies between features, with marginal replacement, which ignores these dependencies.

The main contribution is a paired experimental protocol that keeps the predictor, evaluation data, labels, and scoring procedure fixed while changing the replacement distribution. On the Dry Bean dataset, the authors show that conditional replacement better preserves the observed feature distribution, but produces substantially different importance scores and rankings from marginal replacement. Additional experiments investigate stochastic versus deterministic and label-informed replacement strategies on several other datasets.

The main conclusion is that feature importance depends on the replacement distribution, and that better preservation of the observed feature distribution does not necessarily imply agreement between explanations.

Critical strengths

Careful experimental isolation of the replacement mechanism.

The paired evaluation is well controlled: both approaches use the same predictors, examples, labels, and shared random variates. This reduces confounding and allows differences in importance scores to be attributed to the replacement procedure. The authors also examine multiple aspects of the results, including distributional fidelity, score magnitude, and ranking agreement, rather than focusing on a single metric.

This makes the empirical observation credible, although it does not by itself establish the novelty of the finding.

Critical weaknesses

1. Limited novelty and insufficient differentiation from closely related prior work.

The authors acknowledge that conditional and marginal feature replacement measure different quantities and position their contribution as a controlled paired evaluation of distributional fidelity, score differences, and ranking agreement. This is a reasonable experimental distinction.

However, the extent of the novelty remains unclear. Prior work has already investigated the effects of feature dependence and conditioning on feature-importance scores and rankings (e.g., Strobl et al., 2008; Debeer & Strobl, 2020). In particular, Debeer & Strobl (2020) is not discussed.

While combining distributional diagnostics and explanation agreement in a paired protocol may be useful, the paper does not sufficiently establish what new scientific insight this evaluation provides beyond previously known differences between conditional and marginal importance.

References

Strobl, C., et al. (2008). Conditional variable importance for random forests. BMC Bioinformatics, 9, 307. https://doi.org/10.1186/1471-2105-9-307

Debeer, D., & Strobl, C. (2020). Conditional permutation importance revisited. BMC Bioinformatics, 21, 307. https://doi.org/10.1186/s12859-020-03622-2

2. Limited experimental validation of the central claim.

The primary conditional-versus-marginal comparison is restricted to one dataset (Dry Bean), one predictor architecture (MLP), and one train-test split, with three random seeds.

Although the paper includes additional datasets, those experiments compare different replacement procedures (e.g., stochastic versus deterministic or label-informed replacements), use different evaluation metrics, and do not directly replicate the main conditional-versus-marginal comparison.

Consequently, the paper does not establish how broadly the reported disagreement generalizes across datasets, predictor architectures, or dependency structures. A more systematic evaluation, particularly using controlled synthetic data and multiple real datasets, would substantially strengthen the contribution.

3. Limited analysis of when and why the methods disagree.

The paper documents differences in feature-importance scores and rankings but provides limited insight into the mechanisms determining the magnitude of those differences.

For example, how does ranking disagreement depend on feature correlations, redundancy, predictor sensitivity, or the quality of the conditional approximation? The paper does not provide a theoretical characterization or a systematic experiment varying these factors.

Furthermore, the main ranking comparisons rely on point estimates, without uncertainty quantification for ranking stability. The reported absolute score separation is also an example-level finite-sample quantity, rather than a direct measure of differences between population mean importance scores.

As a result, the experiments establish that disagreement occurs in the studied setting, but provide limited guidance about when to expect it or how to interpret its significance.

Questions for authors

Please see weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.