{"id":"c89418d7-158d-4d41-a429-51712dd5d664","arxiv_id":"2502.07153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Feature importance estimates from LIME, SHAP variants, and Tree Interpreter for tree models vary with data noise, correlation, and imbalance, and no method matches a hand-defined ground truth across settings.","lead":"This paper benchmarks six local feature-importance explainers for decision-tree models on controlled synthetic and real tabular datasets, and finds that their outputs disagree and shift with data noise, correlation, and imbalance. It offers practical guidance on which explainer to trust in which data regime, though the quantitative rankings rest on a hand-defined ground truth that needs correction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative evaluation compares explainers to a hand-defined ground truth based on the data-generating f*, not the fitted model f, so the rankings and guidance in Section 5.2.1 lack a valid reference.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: the hand-defined ground truth conflates predictive relevance in the data-generating process with the actual feature usage of the learned model f. My reading confirms this is the single most consequential threat to the paper's central claim. If the ground truth is wrong, the quantitative rankings of explainers in Section 5 lose their basis, and the paper's stated contribution—guidance for selecting explainers based on data characteristics—is not supported by the reported evidence. The qualitative claim that explainers disagree and are sensitive to noise, correlation, and imbalance is plausible and visible in the figures, so a full rejection is not warranted. The appropriate remedy is to recompute the comparison against a model-based ground truth; until then, the conditional verdict is correct. I therefore recommend no change to the reader's verdict, while emphasizing that the concrete test above is necessary to settle whether the specific rankings survive.","tokens_in":11129,"tokens_out":4449,"duration_ms":44266,"concrete_test":"Re-run the synthetic experiments with a reference ground truth defined on the fitted model f: for the single decision trees used in Section 5.2.1, compute exact TreeSHAP values of f (or exact Gini-decrease importances) as ϕ*; do not multiply by ε before normalizing; for NOT set ϕ*_X2 = 0 for all ρ. Then recompute mean consistency, feature agreement, and compactness in Figures 4–6 and Table 4. If Tshap's consistency improves to near-perfect and LSurro/LIME's 'overestimation' of X2 changes magnitude or sign, the paper's rankings and recommendations are artifacts of the flawed ground truth.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims rest on the ground truth ϕ* defined in Section 4. For NOT, ϕ*_X1 = 1, ϕ*_X2 = ρ even though the true model and the learned decision tree never split on X2; Table 3 shows DT feature importance for NOT has ϕX2 = 0 for all ρ when ε = 0. A faithful local explanation of f must assign X2 zero importance, so any method doing so is penalized against a target that no model-based explainer should match. This systematically biases the consistency and feature-agreement results in Figures 4–6 and Table 4: they measure agreement with an arbitrary data-generating target, not with the model being explained. The ε scaling is also internally inconsistent: explainer outputs are normalized to sum to one, but ϕ* for XOR becomes .5ε + .5ε = ε, which is not normalized; if ϕ* is renormalized, multiplying by ε cancels and noise cannot lower relative importance as claimed. Consequently, the conclusions that 'LSurro and LIME overestimate irrelevant variables', 'TI is most affected by noise', and 'SHAP explainers always favor one feature' are not established by this evaluation. The qualitative observation that explainers disagree and respond to data properties may still survive, but the quantitative rankings and practitioner recommendations depend on a reference that is not the learned model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an empirical study of six local feature-importance explainers (local surrogates, LIME, Kernel/Sampling/Tree SHAP, and Tree Interpreter) applied to decision trees and random forests. Using synthetic data generated from XOR and NOT functions with controlled feature correlation, label noise, and class imbalance, the authors compare each explainer's output to a hand-defined ground-truth feature importance and also report stability, compactness, consistency, and feature/rank agreement metrics. They supplement the synthetic study with experiments on four UCI datasets. The main claims are that explainers disagree substantially, that LSurro and LIME overestimate the importance of irrelevant variables, that Tree Interpreter is the most noise-sensitive, that SHAP explainers consistently favor one feature, and that practitioners should select explainers according to data properties.","tokens_in":11390,"tokens_out":8794,"duration_ms":73680,"significance":"The paper tackles a practically important problem: choosing among local explanation methods for tree-based models under different data conditions. Its strengths include the systematic control of data-generating mechanisms, the comparison of six methods on a common set of metrics, the use of both synthetic and real data, and explicit practitioner recommendations. If the quantitative results were reliable, the paper would be a useful reference for model interpretability. However, the validity of the quantitative rankings depends on the ground-truth definition in Section 4, which has internal inconsistencies and normalization problems. The qualitative observation that explainers disagree and respond to noise and correlation is plausible and partially supported by the stability and agreement results, but the specific rankings and recommendations require substantial rework.","major_comments":[{"comment":"For the NOT datasets, the true model f* is described as a depth-1 decision tree that splits only on X1, yet the ground truth assigns phi*_X2 = rho. Table 3 confirms that the learned decision tree gives phi_X2 = 0 for all rho when epsilon = 0. Because the paper's stated objective is to explain predictions of the learned model f (Abstract, Section 5), a faithful local explanation of f must assign zero importance to X2; penalizing methods that do so (Figures 4-6, Table 4) is a mismatch between the reference and the object of explanation. The recommendations in Section 6 to avoid LSurro and LIME on datasets with irrelevant variables are derived from this comparison and are therefore not established. Please redefine the ground truth as the feature importance of the fitted model (e.g., exact Shapley values of the learned tree), or explicitly change the evaluation target to 'recovery of data-generating importance' and justify why post-hoc local explainers should be judged against that target.","section":"Section 4, 'Ground truth feature importance'"},{"comment":"For XOR with noise epsilon, the ground truth is defined as phi*_X1 = phi*_X2 = 0.5*epsilon, so phi*_X1 + phi*_X2 = epsilon, while Section 5.2.1 states that all explainer outputs are normalized so their contributions sum to one. Comparing normalized explainer outputs to an unnormalized ground truth is not meaningful; if the ground truth is also renormalized, the factor epsilon cancels and noise cannot lower the relative importance of X1 and X2. Therefore the claims in Section 5.2.1 that 'noise lowers the importance of X1 and X2 for all the explainers' and that 'Tree Interpreter is the most affected by noise' are not supported by the current analysis. Please specify the exact normalization used for both estimates and ground truth, and recompute the noise comparisons accordingly.","section":"Section 4, epsilon scaling"},{"comment":"The claim that 'LSurro and LIME overestimate the importance of irrelevant variables' is ambiguous. Under the paper's own ground truth, X2 in the NOT datasets receives positive importance rho whenever rho > 0, so it is not 'irrelevant' by that definition; at rho = 0 the claim reduces to a special case. To support the practical recommendation in Section 6, report the rho = 0 results separately and define relevance explicitly relative to the fitted model f (e.g., features that never appear in any split), or use a model-based ground truth throughout.","section":"Section 5.2.1, overestimation claim"}],"minor_comments":[{"comment":"The reference [17] for Tree Interpreter is actually 'A Debiased MDI Feature Importance Measure for Random Forests' (Li et al.), which is not the Tree Interpreter; please cite the correct source for Tree Interpreter (e.g., Saabas's implementation) or adjust the table.","section":"Table 1"},{"comment":"The caption says 'for the XOR datasets' but the table includes NOT rows; update the caption to 'XOR and NOT datasets'.","section":"Table 4 caption"},{"comment":"Figures 2 and 3 are referenced in Section 5.1 but never described in the text; please add a sentence explaining the layout (e.g., rows/columns correspond to values of rho and epsilon).","section":"Figures 2 and 3"},{"comment":"The column headers '# Features for 90% Accuracy' and 'Accuracy with 5 feature(%)' are unclear; specify whether 90% refers to model accuracy or to the fraction of the explainer's approximation, and define '5 feature' (e.g., five features or 5% of features).","section":"Tables 5-8"},{"comment":"The sentence 'we compute the feature importance scores of the learned model f on datasets with 1 000 instances' is ambiguous about whether 1,000 is the training or total sample size; please clarify and use the conventional '1,000' separator.","section":"Section 5.1"},{"comment":"Section 2 introduces several metrics (stability, compactness, consistency, feature agreement) but does not give exact formulas or definitions of the implementation; add precise definitions (e.g., how the L2 distance and top-k feature agreement are computed) to make the experiments reproducible.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains an extensive empirical study but the central quantitative evaluation is compromised by the ground-truth definition and normalization inconsistencies. The qualitative finding that local explainers for trees disagree and respond to data properties may still be salvageable with a reworked or model-based ground truth. I also noticed a citation error for Tree Interpreter; this, together with the incomplete metric definitions, suggests the manuscript needs careful polishing. I recommend major revision rather than rejection because the problem is in principle fixable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a reasonable empirical comparison of six local explainers on tree ensembles, but the central quantitative claims rest on a hand-defined ground truth that does not correspond to the model being explained. The qualitative phenomenon—explainers disagree and their outputs shift with noise, correlation, and imbalance—is credible and visible in the raw figures. What is actually new is a controlled two-feature synthetic setup (XOR and NOT) with varying noise and correlation, plus consistency/stability/compactness comparisons on four UCI datasets. That is a legitimate extension of existing benchmarking work like Flora et al., not a revolution.\n\nThe paper does well in scope and spirit. It compares a sensible set of explainers, including tree-specific ones, on a common tabular binary classification task. The synthetic design isolates properties one at a time, and the real-data section gives a practical feel for the problem. The message that no method is universally reliable is worth repeating.\n\nThe soft spots are serious and concentrate in Section 4. The ground truth for NOT gives X2 importance ρ even though the true decision tree never splits on X2 and Table 3 shows the fitted tree's own importance for X2 is 0 at ε=0. That is not a minor detail: the consistency and feature-agreement metrics in Figures 4–6 and Table 4 compare explainers against a target that mixes data-generating relevance with model usage. Any method that correctly ignores X2 is penalized. The ε scaling is also internally inconsistent: explainer outputs are normalized to sum to one, but the ground truth is scaled by ε, so at ε=0.5 the XOR ground truth sums to 0.5. Unless the ground truth is renormalized, the 'noise lowers importance' finding is partly an artifact. The covariance matrix for ρ=1 is singular. There are no error bars, no repeated seeds, and no code or data release, so the numbers in Tables 3–8 are hard to verify. The abstract's claim that some hyperparameters don't matter is unsupported—there is no hyperparameter experiment. These flaws do not kill the qualitative takeaway, but they do undermine the specific rankings and the practitioner guidance (like 'avoid LSurro for irrelevant variables' and 'avoid TI for noisy data').\n\nOverall, I would call this a conditional pass, not a reject. The authors need to fix the ground truth definition, add uncertainty quantification, release artifacts, and either support or remove the hyperparameter claim. I would send it to peer review because the question is relevant and the flaws are fixable; a good referee could push them to a much stronger paper. I would not cite it in my own work until the evaluation is redone.","headline":"A well-intentioned but flawed benchmark: the ground truth for feature importance is defined against the data-generating process rather than the model being explained, so the quantitative rankings of explainers do not hold.","tokens_in":11911,"tokens_out":3181,"would_cite":false,"duration_ms":28836,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Feature-importance estimates from local explainers depend on data properties such as correlation, noise, and imbalance, so no single explanation method is reliable across all datasets.","keywords":["Explainability","Decision Tree Models","Feature Importance","Synthetic Data Generation","LIME","SHAP","Tree Interpreter","Local explanations"],"falsifier":"Recompute the benchmark with ground truth defined by the learned model's actual split usage (such as Gini decrease) rather than the generative function's hand-assigned importance; if LSurro and LIME no longer overestimate irrelevant features, or the SHAP preference for one feature disappears, then the paper's method-level conclusions depend on that ground-truth choice.","tokens_in":10844,"feed_emoji":"📊","tokens_out":6530,"duration_ms":54186,"temperature":0.7,"pith_summary":"The paper sets out to show that local feature-importance explanations for decision-tree models are not universally trustworthy: which method is accurate depends on properties of the data being explained. To make this point, the authors generate synthetic binary-classification datasets from two ground-truth functions (XOR and NOT) while varying feature correlation, label noise, and class imbalance, so that the correct feature importance is known by construction. Across these datasets and four public tabular datasets, the six explainers examined—local surrogates, LIME, three SHAP variants, and Tree Interpreter—disagree in magnitude and even sign. The paper concludes with practical guidance: avoid LSurro and LIME when irrelevant variables are present, avoid Tree Interpreter under heavy label noise, and treat the SHAP variants as largely interchangeable. The value of the claim, if correct, is that practitioners can choose an explanation method from measurable data properties rather than assuming one method is always right.","feed_headline":"Pick your explainer by data: no method wins everywhere","feed_subtitle":"LIME overrates irrelevant features, SHAP favors one feature, and tree interpreter degrades with noise.","key_machinery":"The machinery is a controlled synthetic-data generation framework built from the factorization $P(X_1, X_2, Y) = P(Y \\mid X_1, X_2)P(X_2 \\mid X_1)P(X_1)$, which lets the authors isolate three data properties: the marginal distribution of $X_1$, the conditional dependence of $X_2$ on $X_1$ (correlation $\\rho$), and the label noise $\\epsilon$ introduced by flipping a fraction of labels. Two logical target functions, XOR and NOT, fix the ground-truth feature importance $\\phi^*_X$: for XOR both features matter equally ($0.5$ each), while for NOT only $X_1$ matters directly and $X_2$ receives importance equal to its correlation with $X_1$. Against that ground truth, the paper compares six local explainers using consistency, stability, compactness, and feature and rank agreement metrics. This setup is what lets the paper attribute differences in explanations to data properties rather than to dataset-specific accident.","core_discovery":"The central claim is that feature importance estimates from local explanation methods inherit the properties of the data-generating distribution $P(X_1, X_2, Y)$, and that no single explainer is reliable across all data conditions. The paper's experiments show that LSurro and LIME systematically overestimate the importance of irrelevant variables, Tree Interpreter is the most affected by label noise, the SHAP family yields very similar explanations to each other but consistently favors one feature over the other, and feature correlation inflates the importance and instability of $X_2$ in XOR datasets for SHAP explainers. From these patterns the authors infer concrete selection rules: for data with irrelevant variables avoid LSurro and LIME; for highly noisy data avoid Tree Interpreter; and for speed or adaptability prefer Sampling SHAP or Tree SHAP over Kernel SHAP. The claim is accompanied by the caveat that the study is restricted to binary classification on tabular data with two synthetic features, so the qualitative conclusions are established for that setting.","pith_inferences":["The two-feature design isolates mechanisms cleanly, but in higher-dimensional tabular data the surrogate-based methods may diverge even more, because local neighborhoods become harder to populate and spurious correlations multiply; this is a testable extension rather than a paper claim.","Because the failure modes are tied to measurable data properties, a natural follow-up is a meta-predictor that inspects correlation, noise, and imbalance and recommends an explainer; the reported patterns provide the training signal for such a rule.","The SHAP family's consistent favoring of one feature in these experiments suggests that Shapley attributions on correlated inputs can be misleading for causal readings of feature importance, a distinction the paper does not develop.","The same synthetic protocol could be reused for regression and multi-class targets, though the ground-truth importance definitions would need to be re-derived, so the ranking of methods in those settings should not be assumed to carry over."],"forward_implications":["For datasets with irrelevant variables, practitioners should avoid LSurro and LIME, since both overestimate the importance of uninformative features.","For datasets with high label noise, Tree Interpreter should be avoided because it is the most unstable explainer in that setting.","Kernel SHAP, Sampling SHAP, and Tree SHAP produce near-identical explanations, so Sampling SHAP or Tree SHAP can be used interchangeably with faster computation on tree ensembles.","Feature correlation inflates the importance and instability of the correlated feature under SHAP explainers in XOR-type data, while label noise lowers feature importance across all methods.","Model hyperparameters such as tree depth appear to influence feature importance less than data properties, so data diagnostics matter more than hyperparameter tuning in explainer choice."],"supporting_citations":[{"why":"Supplies the LIME method whose behavior is evaluated across the synthetic and real datasets.","marker":"[7]"},{"why":"Supplies the unified SHAP framework underlying the Kernel SHAP and Sampling SHAP explainers.","marker":"[8]"},{"why":"Defines the local surrogate (LSurro) approach that the paper benchmarks and finds to overestimate irrelevant variables.","marker":"[15]"},{"why":"Supplies the Tree SHAP method specialized to tree ensembles, one of the six explainers compared.","marker":"[16]"},{"why":"The reference the paper attaches to the Tree Interpreter method, which it finds most affected by noise.","marker":"[17]"},{"why":"Provides the stability metric used to compare how much each explainer's output changes in a local neighborhood.","marker":"[18]"},{"why":"Source of the four public tabular datasets used for the real-world demonstration.","marker":"[37]"}],"fun_headline_variants":["No universal explainer: data traits decide which feature importance method works","Data properties, not model choice, determine which explainer to trust","No best explainer exists: data characteristics decide your feature importance tool","Feature importance truth varies with data: choose explainer accordingly","When to trust LIME vs SHAP? It depends on your data, not your model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison stands on the hand-assigned ground truth—equal importance to both features in XOR, and $X_2$'s importance equal to its correlation $\\rho$ with $X_1$ in NOT—being the right target for an explainer rather than the features the learned tree actually splits on.","fun_headline_variants_meta":{"raw":{"variants":["No universal explainer: data traits decide which feature importance method works","Data properties, not model choice, determine which explainer to trust","No best explainer exists: data characteristics decide your feature importance tool","Feature importance truth varies with data: choose explainer accordingly","When to trust LIME vs SHAP? It depends on your data, not your model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001635,"raw_usage":{"total_tokens":6497,"prompt_tokens":936,"completion_tokens":5561,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":5467}},"tokens_in":552,"tokens_out":5561,"duration_ms":33430,"temperature":1.0,"reasoning_tokens":5467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:38:24.128413+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the benchmark with ground truth defined by the learned model's actual split usage (such as Gini decrease) rather than the generative function's hand-assigned importance; if LSurro and LIME no longer overestimate irrelevant features, or the SHAP preference for one feature disappears, then the paper's method-level conclusions depend on that ground-truth choice.","supporting_citations":[{"cited_title":"In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining","cited_arxiv_id":null,"evidence_quote":"Supplies the LIME method whose behavior is evaluated across the synthetic and real datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the local surrogate (LSurro) approach that the paper benchmarks and finds to overestimate irrelevant variables."},{"cited_title":"arXiv (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the Tree SHAP method specialized to tree ensembles, one of the six explainers compared."},{"cited_title":"A Debiased MDI Feature Importance Measure for Random Forests","cited_arxiv_id":"1906.10845","evidence_quote":"The reference the paper attaches to the Tree Interpreter method, which it finds most affected by noise."},{"cited_title":"arXiv (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the stability metric used to compare how much each explainer's output changes in a local neighborhood."},{"cited_title":"http://archive.ics","cited_arxiv_id":null,"evidence_quote":"Source of the four public tabular datasets used for the real-world demonstration."}],"review_version":1}