Unrestricted permutation forces extrapolation: variable importance requires at least one more model, or there is no free variable importance
Open accessThe DOI matches its Crossref record; 50 of the 52 references it cites were found in a registry or an index.
Abstract
This paper reviews and advocates against the use of permute-and-predict (PaP) methods for interpreting black box functions. Methods such as the variable importance measures proposed for random forests, partial dependence plots, and individual conditional expectation plots remain popular because they are both model-agnostic and depend only on the pre-trained model output, making them computationally efficient and widely available in software. However, numerous studies have found that these tools can produce diagnostics that are highly misleading, particularly when there is strong dependence among features. The purpose of our work here is to (i) review this growing body of literature, (ii) provide further demonstrations of these drawbacks along with a detailed explanation as to why they occur, and (iii) advocate for alternative measures that involve additional modeling. In particular, we describe how breaking dependencies between features in hold-out data places undue emphasis on sparse regions of the feature space by forcing the original model to extrapolate to regions where there is little to no data. We explore these effects across various model setups and find support for previous claims in the literature that PaP metrics can vastly over-emphasize correlated features in both variable importance measures and partial dependence plots. As an alternative, we discuss and recommend more direct approaches that involve measuring the change in model performance after muting the effects of the features under investigation.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“However, these methods can be misleading when correlations exist between input variables 107,109 .”
From Machine learning in concrete science: applications, challenges, and best practices (2022) · cites this paper as [107]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Breiman, L. (2001) · Machine Learning · doi:10.1023/a:1010933404324 · Cited by 133,538 · cited 3×
“Breiman (2001) designed the method specifically for use with random forests and considered out-of-bag (OOB) loss in which variable importance is averaged over trees, measured with that part of the data not used to construct the corresponding tree.”
MethodGreedy function approximation: A gradient boosting machine
Friedman, J. H. (2001) · The Annals of Statistics · doi:10.1214/aos/1013203451 · Cited by 30,731
A Unified Approach to Interpreting Model Predictions
Lundberg, S., & Lee, S.-I. (2017) · arXiv · doi:10.48550/arxiv.1705.07874 · Cited by 23,707 · cited 3×
“Saliency (Simonyan et al., 2013) obtains derivatives with respect to x , while LIME (Ribeiro et al., 2016) approximates this with a LASSO method localized to the point of interest by a kernel; Shapley values can also be used for individual predictions (Lundberg and Lee, 2017).”
BackgroundMethod"Why Should I Trust You?": Explaining the Predictions of Any Classifier
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016) · Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2939672.2939778 · Cited by 17,555
Generalized Additive Models: An Introduction with R
Wood, S. N. (2017) · doi:10.1201/9781315370279 · Cited by 13,852
Rudin, C. (2019) · Nature Machine Intelligence · doi:10.1038/s42256-019-0048-x · Cited by 10,561 · cited 1×
“As a consequence, many researchers have suggested methods to “X-ray the black box” in order to provide insight into which input features are important and how they effect predictions; a task so precarious that others have advocated against the practice altogether (Rudin, 2019).”
BackgroundNelsen, R. B. (1999) · Lecture Notes in Statistics · doi:10.1007/978-1-4757-3076-0 · Cited by 7,744
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
Simonyan, K., Vedaldi, A., & Zisserman, A. (2013) · arXiv · doi:10.48550/arxiv.1312.6034 · Cited by 5,370 · cited 1×
“Saliency (Simonyan et al., 2013) obtains derivatives with respect to x , while LIME (Ribeiro et al., 2016) approximates this with a LASSO method localized to the point of interest by a kernel; Shapley values can also be used for individual predictions (Lundberg and Lee, 2017).”
MethodTesting Statistical Hypotheses
(2021) · Wiley Series in Probability and Statistics · doi:10.1002/9781119243830.ch12 · Cited by 5,106
Wood, S. N. (2006) · doi:10.1201/9781420010404 · Cited by 4,953
Research works printed but not linked: 1
- Tan, S., Caruana, R., Hooker, G., Koch, P., Gordo, A.: Learning global additive explanations for neural nets using model distillation. arXiv preprint arXiv:1801.08640 (2018) · arXiv
Cited by
Interpretable Machine Learning – A Brief History, State-of-the-Art and Challenges
Molnar, C., Casalicchio, G., & Bischl, B. (2020) · Communications in Computer and Information Science · doi:10.1007/978-3-030-65965-3_28 · Cited by 677
Explainable machine learning in deployment
Bhatt, U., Xiang, A., Sharma, S., Weller, A., Taly, A., Jia, Y., Ghosh, J., Puri, R., Moura, J. M. F., & Eckersley, P. (2020) · Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency · doi:10.1145/3351095.3375624 · Cited by 633
Big-Data Science in Porous Materials: Materials Genomics and Machine Learning
Jablonka, K. M., Ongari, D., Moosavi, S. M., & Smit, B. (2020) · Chemical Reviews · doi:10.1021/acs.chemrev.0c00004 · Cited by 599
Explaining by Removing: A Unified Framework for Model Explanation
Covert, I., Lundberg, S., & Lee, S. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2011.14878 · Cited by 116
Machine learning and deep learning—A review for ecologists
Pichler, M., & Hartig, F. (2023) · Methods in Ecology and Evolution · doi:10.1111/2041-210x.14061 · Cited by 454
Metadata
- Authors
- Hooker, G., Mentch, L., & Zhou, S.
- Affiliations
- University of California, Berkeley (US)
- University of Pittsburgh (US)
- Corresponding author
- Hooker, Giles
- Venue
- Statistics and Computing
- Published
- 2021-10-29
- Volume
- 31
- Issue
- 6
- ISSN
- 0960-3174 (ISSN-L)
- 1573-1375
- OpenAlex
- W3209434414
- Cited by
- 264
- References
- 52
- Citations by year
- 2026: 73 · 2025: 51 · 2024: 49 · 2023: 61 · 2022: 28 · 2021: 2
- FWCI
- 12.32
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Statistics and Computing (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- Directorate for Mathematical and Physical Sciences (1712554, 1712041, 2015400)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Physical Sciences › Mathematics › Statistics and Probability › Statistical Methods and Inference
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning and Data Classification
- Keywords
- feature importance, partial dependence plots, individual conditional expectation, random forest, feature dependence, extrapolation, model interpretability, black-box models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Unrestricted permutation forces extrapolation: variable importance requires at least one more model, or there is no free variable importance. https://app.cochat.ai/papers/10.1007/s11222-021-10057-z
- The paper
- Hooker, G., Mentch, L., & Zhou, S. (2021). Unrestricted permutation forces extrapolation: variable importance requires at least one more model, or there is no free variable importance. Statistics and Computing. https://doi.org/10.1007/s11222-021-10057-z
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.