Pith. sign in

REVIEW 3 major objections 3 minor 44 references

Provable Recovery of Locally Important Signed Features and Interactions from Random Forest

T0 review · 3 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Under a spike-signed-interaction model, a random-forest interpretability method provably recovers the exact signed features and interactions that drive a single test prediction as sample size grows.

desk verdict A genuine local extension of LSSFind with coherent new consistency theorems, but the experiments do not verify the theorem's RF assumptions and the abstract overstates the practical scope. read the letter →

arxiv 2512.11081 v2 pith:W4MUSY2B submitted 2025-12-11 stat.ML cs.LGstat.ME

Kata Vuk , Nicolas Alexander Ihlo , Merle Behr This is my paper
classification stat.MLcs.LGstat.ME MSC 62G2062G0868T05
keywords randomforestslocalfeatureimportancesignedinteractionsBooleaninteractionmodelLSSconsistencyinterpretabilitytreeensembles
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes LocalLSSFind, a local model-specific importance method for random forests that reports which features, and which combinations of features, drive the prediction for an individual test point, together with the direction of each effect. Its central claim is a consistency theorem: if the data come from a Locally Spike Sparse (LSS) model—a regression function made of Boolean threshold terms—then, under speed constraints on random-forest tuning, the method's output exactly equals the true signed interactions active at the test point, with probability going to 1. This matters because most local explanation methods have no statistical guarantee about recovering the data-generating signal; here the target of explanation is defined by the model and the proof shows the method hits it. The method works by combining two prevalence statistics: how often a signed feature group appears on random tree paths globally, and how often it appears on the specific paths that the test point traverses. A simplified variant is shown to recover local signed feature importance alone.

What carries the argument

Depth-weighted prevalence (DWP) and test-point path prevalence (PP*). DWP is the probability, over randomly chosen depth-weighted paths in the forest and over tree randomness, that a signed feature set appears among splits with impurity decrease at least epsilon; PP* is the same probability but restricted to the unique path that the test point follows in each tree. The method thresholds both, feeding the global filter into a local filter, which is what converts a global interaction-recovery result into a local one.

What would settle it

Train a forest on LSS data with bootstrap enabled and mtry around the square root of p (standard defaults) rather than mtry p/2, run LocalLSSFind with thresholds as in Theorem 1, and measure recovery frequency. The theorem's conditions A3 and A4 are violated; if recovery still holds, the assumptions are stronger than needed, and if it fails, the no-bootstrap and mtry-order conditions are doing real work.

Watch

Extended reading notes

Core claim

LocalLSSFind defines a signed interaction as a set of feature–direction pairs, weights each tree path by depth, and counts co-occurrences of signed features along paths in the forest. It keeps candidate interactions only if their depth-weighted prevalence exceeds one threshold and their test-point-specific path prevalence exceeds another. Theorem 1 states that, for data from the LSS model with uniform features, non-overlapping interactions, bounded response, and sparse signal, if the forest is grown with balanced splits, no bootstrap, and mtry proportional to p, then for any fixed impurity threshold epsilon and thresholds eta_DWP and eta_PP chosen in the window between a small error term b(e

Load-bearing premise

The load-bearing premise is that the data really come from a Locally Spike Sparse model—independent uniform features, non-overlapping Boolean interactions, bounded response, sparse signal—and that the forest is grown without bootstrap, with balanced splits and mtry of order p; the paper itself flags the LSS assumption as a limitation of the theory in its discussion.

Editorial extensions

If this is right

  • The output of LocalLSSFind is an exact recovery statement, not just a ranking: with enough samples, the returned set equals the true signed interactions of the test point (Theorem 1).
  • The guarantee automatically yields consistent recovery of signed individual features, because any feature appearing in a recovered interaction is itself recoverable; a simplified variant, LocalFeatureLSSFind, does this directly (Theorem 2).
  • If an interaction is globally present in the LSS model but is not active at the test point, its path prevalence collapses to zero, so the local filter removes it while keeping the active ones.
  • The method's scores are model-specific and independent of marginal signal strength, so they can identify directional drivers such as 'young age combined with many prior offenses' rather than only additive contributions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: The proof relies on an idealized forest (no bootstrap, balanced volume-ratio splits, mtry of order p). If a practitioner trains a forest with standard bootstrap defaults, the theorem's conditions are not met; one testable extension would be to verify empirically whether bootstrap breaks the recovery or merely requires retuning.
  • Inference: The LSS model assumes independent uniform features and non-overlapping interactions; real data violate this. A natural stress test is to run LocalLSSFind on data with correlated features and overlapping interaction terms and compare recovery rates with the theorem's predicted threshold window.
  • Inference: The paper's own simulations show a trade-off: for long interactions (size 4), local path prevalence can hurt because few training points share the test point's full path. A hybrid that rescales or regularizes PP* based on path sample size might bridge this gap.
  • Inference: Because the target is an individual prediction, the approach points toward per-subject statements such as 'this defendant's high risk score is driven by the interaction of young age and moderate priors'—the kind of claim that personalized-medicine or credit-decision audits need.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper proposes LocalLSSFind, a local, RF-specific method for recovering signed feature and interaction importance for an individual test point. The method combines two prevalence measures: a global depth-weighted prevalence (DWP) over random decision paths and a local path prevalence (PP) over the specific path taken by the test point, both restricted to splits with impurity decrease above ε. Under a Locally Spike Sparse (LSS) model with constraints C1–C4 and random-forest assumptions A1–A4, the paper proves (Theorem 1) that the output of LocalLSSFind equals the set of basic signed interactions of the test point of size at most s_max, with probability converging to one as n→∞. Theorem 2 gives an analogous result for signed feature importance via a simplified variant. The proof strategy is to lower-bound the local prevalence of true test-point interactions (Proposition 1), show that global but not local interactions have vanishing local prevalence (Proposition 2), and combine these with the global consistency result from [8]. The paper also reports simulations and a COMPAS application, and compares the method with TreeSHAP.

Significance. If Theorem 1 is correct, this is the first consistency result for local, signed interaction recovery in random forests under a precise data-generating model. The LSS framework gives a defensible definition of 'true' local interactions, and the paper is transparent about the modeling assumptions. The proof strategy is coherent and builds on the established global result of [8] rather than introducing a circular argument. The paper also provides code for the proposed method, which supports reproducibility. The main value is the rigorous formalization of a local importance claim in a setting where most existing local interpretability tools lack statistical guarantees. The contribution is incremental relative to [8] but the local extension is nontrivial and potentially useful for personalized interpretations.

major comments (3)
  1. [§5.1, §5.3 and Assumptions A2/A4] Theorem 1 is conditional on A1–A4, including no bootstrap or subsampling (A4) and balanced splits with volume ratio bounded below by Cγ/(1−Cγ) (A2). The experiments are presented as confirming the theory, yet §5.1 only reports 'RF were trained with n=1000 or 10000 samples using mtry=p/2=10 and 500 trees', and §5.3 uses ranger with cross-validated min.node.size. Neither section states that bootstrap was disabled (bootstrap=False / replace=FALSE) or that splits were constrained to satisfy A2. Since scikit-learn and ranger default to bootstrap, Figures 1–4 and Table 1 may lie outside the theorem's scope. The authors should either enforce and report A2/A4 in the experiments, or explicitly reframe the empirical sections as heuristic illustrations rather than confirmation of Theorem 1.
  2. [§2, Eqs. (3)–(4); Algorithm 1; §5.1] DWP and PP* are defined as exact probabilities conditional on the data D, and Algorithm 1 uses these exact quantities. In practice, a finite forest with B trees only provides empirical frequencies. Theorem 1 contains no growth condition on B and no Monte Carlo concentration term. The convergence statement therefore applies to an oracle version of the algorithm, not to the finite-B procedure used in the simulations (500 trees). Please state that the theorem concerns the exact DWP/PP variant and add a convergence result or concentration bound for the empirical estimator as B→∞, or incorporate the Monte Carlo error into the threshold conditions.
  3. [Theorem 1, Eq. (8); §5.1] The threshold conditions in Theorem 1 involve unknown constants Cβ, Cγ, Cm, and s, and require b(ε) to lie in a specific interval that is not constructively specified. The simulations fix ε=η_DWP=η_PP=0.01 without checking whether the inequalities in (8) hold. Thus the theorem is conditional on an oracle choice of thresholds, and the practical plug-in version has no proven guarantee. The paper should state this limitation explicitly and, ideally, provide a data-dependent heuristic for choosing the thresholds or explain why the fixed choice is consistent with the theory.
minor comments (3)
  1. [Proposition 1 and Appendix E] The statement of Proposition 1 says PP*_ε(S*±) ≥ 1−b(ε)+r_n(D,ε), but the proof at the end of Appendix E concludes ≥ 1−b(ε)−r_n(D,ε). Since r_n → 0 in probability, the theorem is unaffected, but the statement and proof should be aligned.
  2. [Appendix B, Notation] The entry for S±_j writes b_k∈{0,1}; this should be b_k∈{−1,+1} to match the definition in the main text.
  3. [§5.1] The sentence 'In total, approximately (3p)^{L+1} = 60^{L+1} signed candidate interactions are possible' is unclear: for p=20, signed features give 2p=40 candidates, not 3p. Please explain the factor 3 or correct the formula.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the local recovery theorem is proven from explicit LSS-model assumptions together with the published global consistency result of [8]; no fitted parameter is renamed as a prediction and no definitional identity links the output to the target set.

full rationale

The central claim, Theorem 1, is that under LSS constraints C1-C4 and RF assumptions A1-A4, LocalLSSFind recovers exactly the basic signed interactions of xtest of size at most smax. The proof does not tune thresholds to the recovered set. The output SL is defined as SG ∩ V, where SG is the DWP-thresholded, minimal-interaction set inherited from LSSFind and V is the PP*-thresholded set. Recovery is established by three independent ingredients: Theorem 3 of [8] identifies SG with the global BSIs in the LSS model; Proposition 1 lower-bounds PP*(S*±) by 1 - b(epsilon) - r_n(D, epsilon) for BSIs that are local for xtest; Proposition 2 shows PP*(S±) -> 0 for BSIs that are not local for xtest. These propositions are proved in the appendix from CART impurity-decrease concentration, the LSS thresholds, and oracle definitions such as F(P*) and U(t); these oracle objects are not defined in terms of Algorithm 1's output. Although the paper relies heavily on the authors' own prior work [8] for the global consistency theorem and technical lemmas, that work is a published, parameter-free, externally checkable consistency result whose stated assumptions do not include the present local target, so under Rule 4 it counts as independent evidence rather than circularity. The non-constructive threshold condition and the lack of an explicit statement that bootstrap was disabled in the experiments are scope/correctness concerns, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a specific generative model (LSS), a set of strong distributional constraints (C1-C4), and an idealized RF configuration (A1-A4). The method's tuning parameters epsilon, eta_DWP, eta_PP, and smax are not fitted but are chosen by hand, with no data-driven guidance ensuring the theorem's inequalities. No new physical or computational entities are postulated.

free parameters (4)
  • epsilon (impurity decrease threshold) = 0.01 in simulations; no general selection rule
    Appears in (2), Algorithm 1, and threshold conditions in Theorem 1; the theorem requires 2^s b(epsilon) < eta thresholds, which involves unknown C_beta, C_gamma.
  • eta_DWP (global depth-weighted prevalence threshold) = 0.01 in simulations
    Algorithm 1 screening threshold for DWP; Theorem 1 requires it to lie between 2^s b(epsilon) and C_m^s/2.
  • eta_PP (local path prevalence threshold) = 0.01 in simulations
    Algorithm 1 local filtering threshold; Theorem 1 requires b(epsilon) < eta_PP < 1.
  • smax (maximum interaction size) = set by user; L+1 in simulations
    Restricts output size; the theorem only guarantees recovery of test-point BSIs of size at most smax.
assumptions (4)
  • domain assumption LSS model (Definition 1): E[Y|X] = beta_0 + sum_j beta_j prod_{k in S_j} 1(X_k <= gamma_k)
    The ground truth of 'true local interactions' is defined by this model; without it the notion of interaction is scale-dependent (Section 3, Discussion).
  • domain assumption C1-C4: X uniform on [0,1]^p, |Y|<1, non-overlapping interaction sets, sparsity s=O(1), log(p)/n -> 0
    Section 3; C3 (disjoint S_j) is needed for identifiability/minimality of the recovered set, C4 keeps the number of BSIs finite.
  • domain assumption A1-A4: full-depth trees, balanced splits (A2), mtry = C*p (A3), no bootstrap/subsampling (A4)
    Section 3; all proofs of local prevalence concentration (Lemmas 6-7, Props 1-2) require these. Standard RF implementations violate A4 by default.
  • standard math Theorem 3 of [8], Lemma S11/S13/S2 and Prop S6 from supplement of [8] are correct and applicable
    The proof of Theorem 1 imports the global LSSFind consistency result and key impurity-decrease concentration bounds from [8] without re-proving them (Section 4, Appendix E).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Provable Recovery of Locally Important Signed Features and Interactions from Random Forest." pith.science (2026). https://pith.science/paper/W4MUSY2B

@misc{pith2026251211081,
  author       = {Pith},
  title        = {Pith review of: Provable Recovery of Locally Important Signed Features and Interactions from Random Forest},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4MUSY2B}},
  note         = {Machine review of arXiv:2512.11081}
}
read the original abstract

Feature and Interaction Importance (FII) methods are essential in supervised learning for assessing the relevance of input variables and their interactions in complex prediction models. In many domains, such as personalized medicine, local interpretations for individual predictions are often required, rather than global scores summarizing overall feature importance. Random Forests (RFs) are widely used in these settings, and existing interpretability methods typically exploit tree structures and split statistics to provide model-specific insights. However, theoretical understanding of local FII methods for RF remains limited, making it unclear how to interpret high importance scores for individual predictions. We propose a novel, local, model-specific FII method that identifies frequent co-occurrences of features along decision paths, combining global patterns with those observed on paths specific to a given test point. We prove that our method consistently recovers the true local signal features and their interactions under a Locally Spike Sparse (LSS) model and also identifies whether large or small feature values drive a prediction. We illustrate the usefulness of our method and theoretical results through simulation studies and a real-world data example.

Figures

Figures reproduced from arXiv: 2512.11081 by the authors.

Figure 1
Figure 1. Histograms of PP for BSIs in the LSS model. BSIs for the test points are shown [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Evaluations of interaction rankings: y-axis shows rankings based on [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Ranks of local interaction importance scores of BSIs in the LSS model. BSIs for [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Relative frequency, with which BSIs for the test point are included in the top 10 [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: Global signed feature importance scores, [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Interaction map showing pairwise signed feature interaction scores. Each point [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: Relationship between local interaction importance ( [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 10 canonical work pages

  1. [8]

    Provable Boolean interaction recovery from tree ensemble obtained via random forests,

    M. Behr, Y. Wang, X. Li, and B. Yu, “Provable Boolean interaction recovery from tree ensemble obtained via random forests,”Proceedings of the National Academy of Sciences, vol. 119, no. 22, e2118636119, 2022.doi:10.1073/pnas.2118636119

    work page doi:10.1073/pnas.2118636119 2022

  2. [1]

    MegaSNPHunter: A learning approach to detect disease predisposition SNPs and high level interactions in 27 REFERENCES REFERENCES genome wide association study,

    X. Wan, C. Yang, Q. Yang, H. Xue, N. L. Tang, and W. Yu, “MegaSNPHunter: A learning approach to detect disease predisposition SNPs and high level interactions in 27 REFERENCES REFERENCES genome wide association study,”BMC Bioinformatics, vol. 10, no. 1, p. 13, 2009.doi: 10.1186/1471-2105-10-13

    work page doi:10.1186/1471-2105-10-13 2009

  3. [2]

    SNPInterForest: A new method for detecting epistatic interactions,

    M. Yoshida and A. Koike, “SNPInterForest: A new method for detecting epistatic interactions,”BMC Bioinformatics, vol. 12, no. 1, p. 469, 2011.doi: 10.1186/1471- 2105-12-469

    doi:10.1186/1471- 2011

  4. [3]

    Random forests,

    L. Breiman, “Random forests,”Machine Learning, vol. 45, pp. 5–32, 2001.doi: 10.1023/A:1010933404324

    doi:10.1023/a:1010933404324 2001

  5. [4]

    Iterative random forests to discover predictive and stable high-order interactions,

    S. Basu, K. Kumbier, J. B. Brown, and B. Yu, “Iterative random forests to discover predictive and stable high-order interactions,”Proceedings of the National Academy of Sciences, vol. 115, no. 8, pp. 1943–1948, 2018.doi:10.1073/pnas.1711236115

    work page doi:10.1073/pnas.1711236115 1943

  6. [5]

    Kumbier, S

    K. Kumbier, S. Basu, J. B. Brown, S. Celniker, and B. Yu,Refining interaction search through signed iterative Random Forests, 2018.doi:10.1101/467498

    work page doi:10.1101/467498 2018

  7. [6]

    Learning epistatic polygenic phenotypes with Boolean interactions,

    M. Behr et al., “Learning epistatic polygenic phenotypes with Boolean interactions,” PLOS ONE, vol. 19, no. 4, e0298906, 2024.doi:10.1371/journal.pone.0298906

    work page doi:10.1371/journal.pone.0298906 2024

  8. [7]

    Epistasis regulates genetic control of cardiac hypertrophy,

    Q. Wang et al., “Epistasis regulates genetic control of cardiac hypertrophy,”Nature Cardiovascular Research, vol. 4, no. 6, pp. 740–760, 2025.doi: 10.1038/s44161-025- 00656-8

    work page doi:10.1038/s44161-025- 2025

Show all 44 references
  1. [9]

    Representation Properties of Networks: Kolmogorov’s Theorem Is Irrelevant,

    F. Girosi and T. Poggio, “Representation Properties of Networks: Kolmogorov’s Theorem Is Irrelevant,”Neural Computation, vol. 1, no. 4, pp. 465–469, 1989.doi: 10.1162/neco.1989.1.4.465

    1989 doi

  2. [10]

    “Why Should I Trust You?

    M. T. Ribeiro, S. Singh, and C. Guestrin, ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Association for Computing Machinery, 2016, pp. 1135–1144.d...

    2016

  3. [11]

    Analysis of regression in game theory approach,

    S. Lipovetsky and M. Conklin, “Analysis of regression in game theory approach,” Applied Stochastic Models in Business and Industry, vol. 17, no. 4, pp. 319–330, 2001. doi:10.1002/asmb.446

    2001 doi

  4. [12]

    Explaining prediction models and individual predic- tions with feature contributions,

    E. ˇStrumbelj and I. Kononenko, “Explaining prediction models and individual predic- tions with feature contributions,”Knowledge and Information Systems, vol. 41, no. 3, pp. 647–665, 2014.doi:10.1007/s10115-013-0679-x

    2014 doi

  5. [13]

    A unified approach to interpreting model predictions,

    S. M. Lundberg and S. -I. Lee, “A unified approach to interpreting model predictions,” inAdvances in Neural Information Processing Systems, vol. 30, Curran Associates, Inc., 2017. [Online]. Available: https://proceedings.neurips.cc/paper_files/ paper/2017/hash/8a20a8621978632d...

    2017

  6. [14]

    An axiomatic approach to the concept of interaction among players in cooperative games,

    M. Grabisch and M. Roubens, “An axiomatic approach to the concept of interaction among players in cooperative games,”International Journal of Game Theory, vol. 28, no. 4, pp. 547–565, 1999.doi:10.1007/s001820050125

    1999 doi

  7. [15]

    The Shapley Taylor Interaction Index,

    M. Sundararajan, K. Dhamdhere, and A. Agarwal, “The Shapley Taylor Interaction Index,” inProceedings of the 37th International Conference on Machine Learning, PMLR, 2020, pp. 9259–9268. [Online]. Available: https://proceedings.mlr.press/ v119/sundararajan20a.html

    2020

  8. [16]

    From Shapley Values to Generalized Additive Models and back,

    S. Bordt and U. v. Luxburg, “From Shapley Values to Generalized Additive Models and back,” inProceedings of The 26th International Conference on Artificial Intelligence and Statistics, PMLR, 2023, pp. 709–745. [Online]. Available: https://proceedings. mlr.press/v206/bordt23a.html

    2023

  9. [17]

    Faith-Shap: The Faithful Shapley Interaction Index,

    C.-P. Tsai, C.-K. Yeh, and P. Ravikumar, “Faith-Shap: The Faithful Shapley Interaction Index,”Journal of Machine Learning Research, vol. 24, no. 94, pp. 1–42, 2023. [Online]. Available:http://jmlr.org/papers/v24/22-0202.html

    2023

  10. [18]

    Liang, Z

    Z. Liang, Z. T. Rewolinski, A. Agarwal, T. M. Tang, and B. Yu,Local MDI+: Local feature importances for tree-based models, 2025.doi:10.48550/arXiv.2506.08928

    2025 doi arXiv

  11. [19]

    From local explanations to global understanding with explain- able AI for trees,

    S. M. Lundberg et al., “From local explanations to global understanding with explain- able AI for trees,”Nature Machine Intelligence, vol. 2, no. 1, pp. 56–67, 2020.doi: 10.1038/s42256-019-0138-9 29 REFERENCES REFERENCES

    2020 doi

  12. [20]

    SHAFF: Fast and consistent SHApley eFfect estimates via random forests,

    C. B´ enard, G. Biau, S. Da Veiga, and E. Scornet, “SHAFF: Fast and consistent SHApley eFfect estimates via random forests,” inProceedings of The 25th International Conference on Artificial Intelligence and Statistics, PMLR, 2022, pp. 5563–5582. [Online]. Available:https://pro...

    2022

  13. [21]

    Beyond TreeSHAP: Effi- cient Computation of Any-Order Shapley Interactions for Tree Ensembles,

    M. Muschalik, F. Fumagalli, B. Hammer, and E. H¨ ullermeier, “Beyond TreeSHAP: Effi- cient Computation of Any-Order Shapley Interactions for Tree Ensembles,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 13, pp. 14 388–14 396, 2024.doi:10.1609/aaai...

    2024 doi

  14. [22]

    Definitions, methods, and applications in interpretable machine learning,

    W. J. Murdoch, C. Singh, K. Kumbier, R. Abbasi-Asl, and B. Yu, “Definitions, methods, and applications in interpretable machine learning,”Proceedings of the National Academy of Sciences, vol. 116, no. 44, pp. 22 071–22 080, 2019.doi: 10.1073/ pnas.1900654116

    2019

  15. [23]

    Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges,

    C. Rudin, C. Chen, Z. Chen, H. Huang, L. Semenova, and C. Zhong, “Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges,”Statistics Surveys, vol. 16, pp. 1–85, 2022.doi:10.1214/21-SS133

    2022 doi

  16. [24]

    Interpretable Machine Learning for Discovery: Sta- tistical Challenges and Opportunities,

    G. I. Allen, L. Gan, and L. Zheng, “Interpretable Machine Learning for Discovery: Sta- tistical Challenges and Opportunities,”Annual Review of Statistics and Its Application, vol. 11, no. 1, pp. 97–121, 2024.doi: 10.1146/annurev-statistics-040120-030919

    2024 doi

  17. [25]

    Explaining the explainer: A first theoretical analysis of LIME,

    D. Garreau and U. Luxburg, “Explaining the explainer: A first theoretical analysis of LIME,” inProceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR, 2020, pp. 1287–1296. [Online]. Available: https: //proceedings.mlr.press/v10...

    2020

  18. [26]

    Zhou and G

    Z. Zhou and G. Hooker,Unbiased measurement of feature importance in tree-based methods, 2020.doi:10.48550/arXiv.1903.05179

    2020 doi arXiv

  19. [27]

    Unbiased variable importance for random forests,

    M. Loecher, “Unbiased variable importance for random forests,”Communications in Statistics - Theory and Methods, vol. 51, no. 5, pp. 1413–1425, 2022.doi: 10.1080/ 03610926.2020.1764042 30 REFERENCES REFERENCES

    2022

  20. [28]

    Mean decrease accuracy for random forests: Inconsistency, and a practical solution via the Sobol-MDA,

    C. B´ enard, S. Da Veiga, and E. Scornet, “Mean decrease accuracy for random forests: Inconsistency, and a practical solution via the Sobol-MDA,”Biometrika, vol. 109, no. 4, pp. 881–900, 2022.doi:10.1093/biomet/asac017

    2022 doi

  21. [29]

    Statistical Inference for Variable Importance,

    M. J. Van Der Laan, “Statistical Inference for Variable Importance,”The International Journal of Biostatistics, vol. 2, no. 1, 2006.doi:10.2202/1557-4679.1008

    2006

  22. [30]

    A General Framework for Inference on Algorithm-Agnostic Variable Importance,

    B. D. Williamson, P. B. Gilbert, N. R. Simon, and M. Carone, “A General Framework for Inference on Algorithm-Agnostic Variable Importance,”Journal of the American Statistical Association, vol. 118, no. 543, pp. 1645–1658, 2023.doi: 10.1080/01621459. 2021.2003200

    2023

  23. [31]

    Testing conditional independence in supervised learning algorithms,

    D. S. Watson and M. N. Wright, “Testing conditional independence in supervised learning algorithms,”Machine Learning, vol. 110, no. 8, pp. 2107–2129, 2021.doi: 10.1007/s10994-021-06030-6

    2021 doi

  24. [32]

    L. Gan, L. Zheng, and G. I. Allen,Model-Agnostic Confidence Intervals for Feature Importance: A Fast and Powerful Approach Using Minipatch Ensembles, 2023.doi: 10.48550/arXiv.2206.02088

    2023 doi

  25. [33]

    SIRUS: Stable and Interpretable RUle Set for classification,

    C. B´ enard, G. Biau, S. Da Veiga, and E. Scornet, “SIRUS: Stable and Interpretable RUle Set for classification,”Electronic Journal of Statistics, vol. 15, no. 1, 2021.doi: 10.1214/20-EJS1792

    2021 doi

  26. [34]

    Breiman, J

    L. Breiman, J. H. Friedman, C. J. Stone, and R. A. Olshen,Classification and Regression Trees. New York: Chapman and Hall, 1984

    1984

  27. [35]

    Large Scale Prediction with Decision Trees,

    J. M. Klusowski and P. M. Tian, “Large Scale Prediction with Decision Trees,”Journal of the American Statistical Association, vol. 119, no. 545, pp. 525–537, 2024.doi: 10.1080/01621459.2022.2126782

    2024

  28. [36]

    Analysis of a random forests model,

    G. Biau, “Analysis of a random forests model,”Journal of Machine Learning Research, vol. 13, no. 38, pp. 1063–1095, 2012. [Online]. Available: http://jmlr.org/papers/ v13/biau12a.html

    2012

  29. [37]

    Estimation and Inference of Heterogeneous Treatment Effects using Random Forests,

    S. Wager and S. Athey, “Estimation and Inference of Heterogeneous Treatment Effects using Random Forests,”Journal of the American Statistical Association, vol. 113, no. 523, pp. 1228–1242, 2018.doi:10.1080/01621459.2017.1319839 31 B NOTATIONS

    2018

  30. [38]

    Scikit-learn: Machine Learning in Python,

    F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,”Journal of Machine Learning Research, vol. 12, no. 85, pp. 2825–2830, 2011. [Online]. Available: http: //jmlr.org/papers/v12/pedregosa11a.html

    2011

  31. [39]

    How we analyzed the COMPAS re- cidivism algorithm,

    J. Larson, S. Mattu, L. Kirchner, and J. Angwin, “How we analyzed the COMPAS re- cidivism algorithm,” ProPublica, 2016. [Online]. Available: https://www.propublica. org/article/how-we-analyzed-the-compas-recidivism-algorithm/

    2016

  32. [40]

    All Models are Wrong, but Many are Useful: Learning a Variable’s Importance by Studying an Entire Class of Prediction Models Simultaneously,

    A. Fisher, C. Rudin, and F. Dominici, “All Models are Wrong, but Many are Useful: Learning a Variable’s Importance by Studying an Entire Class of Prediction Models Simultaneously,”Journal of Machine Learning Research, vol. 20, no. 177, pp. 1–81,

  33. [41]

    Decomposing Global Feature Effects Based on Feature Interactions,

    J. Herbinger, M. N. Wright, T. Nagler, B. Bischl, and G. Casalicchio, “Decomposing Global Feature Effects Based on Feature Interactions,”Journal of Machine Learning Research, vol. 25, no. 381, pp. 1–65, 2024. [Online]. Available: http://jmlr.org/ papers/v25/23-0699.html

    2024

  34. [42]

    ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R,

    M. N. Wright and A. Ziegler, “ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R,”Journal of Statistical Software, vol. 77, pp. 1–17, 2017.doi:10.18637/jss.v077.i01

    2017 doi

  35. [43]

    Hyperparameters and tuning strategies for random forest,

    P. Probst, M. N. Wright, and A.-L. Boulesteix, “Hyperparameters and tuning strategies for random forest,”WIREs Data Mining and Knowledge Discovery, vol. 9, no. 3, e1301, 2019.doi:10.1002/widm.1301 A Code repository Implementations of LocalLSSFind inRand Python as well as scrip...

    2019 doi

  36. [2019]

    Available:http://jmlr.org/papers/v20/18-760.html

    [Online]. Available:http://jmlr.org/papers/v20/18-760.html

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.