acceptodds
Under review as a conference paper at ICLR 2027

RoSHAP: Efficient Uncertainty Quantification for Feature Attributions

Abstract

Feature attribution analysis is critical for interpreting machine learning models and supporting data-driven decisions. However, feature attribution values and rankings can vary substantially across train–test splits, random seeds, or model-fitting procedures. The nature of attribution uncertainty makes it challenging to assess the true importance of a feature. We introduce a robust distributional framework **RoSHAP** that treats feature attributions as random quantities. RoSHAP estimates the attribution distribution by refitting the predictive model across multiple data resamples. The resulting distribution captures both attribution magnitude and variability within a unified framework. We show that, under mild regularity conditions, the aggregated attribution score is asymptotically Gaussian, enabling a computationally efficient approach to achieve the desired estimation precision. Simulations and real-data experiments demonstrate that RoSHAP recovers the known feature ordering, quantifies uncertainty in feature importance, and achieves competitive signal identification and predictive performance relative to single-fit and refit-based approaches. By incorporating attribution stochasticity and a computationally efficient approach, RoSHAP provides a reliable and robust metric for assessing feature importance under uncertainty.

Primary area:
interpretability and explainable AI
Keywords:
Feature Attribution, XAI, Uncertainty Quantification, Resampling, SHAP, Asymptotic Normality

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ckoZOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This study addresses a key limitation of existing feature attribution methods: attribution scores are often treated as fixed despite their stochastic nature. The authors propose RoSHAP, a robust distributional framework for estimating attribution distributions over repeated resampling and refitting. By characterizing both the magnitude and variability of feature attributions, the framework reveals when strong attributions are accompanied by substantial uncertainty, while also identifying near-ties and interpretations sensitive to the particular data split. The Gaussian approximation and accompanying diagnostics make this distributional view computationally practical by identifying when the attribution distribution can be summarized parametrically, allowing reliable estimates with substantially fewer resampling rounds.

Critical strengths
  1. This paper is well written and clear which is useful for interpreting unstable feature rankings and allocating computational resources, and provides a coherent alternative to reporting a single fitted model’s attribution scores.

  2. The empirical evaluation is clear and make sense. Across simulations and three real datasets with multiple learners, the paper evaluates distributional approximation, agreement with full-run rankings, moment-estimation precision, and downstream predictive performance. The reduced-budget experiments provide useful evidence that relatively few refits can produce informative summaries in the examined settings which supports the framework’s practical utility.

Critical weaknesses
  1. The applicability of the theorem to the attribution statistics of interest remains insufficiently established. The proof appears sound conditional on A1–A4, but A2 assumes that higher-order split effects are negligible, while A4 assumes Gaussian learner noise with sufficiently stable variance across splits when that noise is non-negligible. These properties are not derived from the evaluated learners. In particular, absolute-value aggregation can invalidate first-order approximations near null features: even a one-feature null linear regression can produce a half-normal aggregate because all evaluation attributions share the same fitted coefficient. This does not contradict the conditional theorem, but it shows that its assumptions are substantive and may exclude practically important cases. The resulting gap limits the theorem’s value as a general justification for Gaussian attribution uncertainty.

  2. The experiment settings and datasets need to be more complex. The simulations fix the sample size at n = 600 and consider only three configurations: (d, s) = (100, 10), (100, 50), and (1000, 50). Consequently, they do not establish how approximation accuracy and uncertainty calibration evolve as n increases or as n and d grow together. The real benchmarks provide some sample-size variation, including approximately 20,000 training observations, but contain only 166–520 features and do not include a controlled scaling analysis or substantially higher-dimensional applications. Larger and more varied settings are needed to assess whether the Gaussian approximation, ranking stability, and low-refit-budget results persist beyond the current regimes. This is consequential because the theoretical justification is asymptotic in n, while the practical motivation includes high-dimensional interpretation and computational efficiency.

Questions for authors

1.Can you justify A2 from properties of at least one evaluated learner, or provide targeted evidence for its behavior on weak and null features? Absolute-value aggregation can produce non-Gaussian limits in these settings. Clarifying the signal regimes and learner properties under which the approximation is reliable would materially affect my assessment of the theoretical contribution.

2.Could you extend both the synthetic and real-data evaluations to larger sample sizes and higher feature dimensionalities?

3 comments
AuthorsOpenReview ↗

Response to Q1: Theoretical applicability and assumptions

We sincerely appreciate the reviewer's careful examination of our theoretical assumptions, which led to a sharper result.

We first identify when A2 theoretically for the linear learners, then discuss existing theory for first-order approximations in tree ensembles, and finally provide targeted diagnostics of A2 and A4 on simulated and real data, with attention to weak and null features.

Linear learners. For L2-penalized logistic and ridge regression, the aggregated linear SHAP attribution is Uj(O)=∣β^j(O)∣∑i∈O∣xij−cj∣,U_j(O)=|\hat\beta_j(O)|\sum_{i\in O}|x_{ij}-c_j|, where the coefficient is fitted on the complement of OO and cjc_j is the SHAP background mean. Under fixed-dimensional regularity conditions, when the limiting coefficient is nonzero, the absolute value is locally linear: the higher-order variance share in A2 vanishes at rate O(n−1)O(n^{-1}), A3 follows from the same moment conditions, and A4 holds trivially because these learners are deterministic.

For null and local-to-zero coefficients, the limit is folded normal, ∣δj+Z∣|\delta_j+Z| with Z∼N(0,1)Z\sim N(0,1), where δj\delta_j measures the coefficient's separation from zero relative to its variability across refits; δj=0\delta_j=0 gives the half-normal limit identified by the reviewer. The higher-order share is 1−η(δj)1-\eta(\delta_j), where η(δ)={2Φ(δ)−1}2Var⁡(∣δ+Z∣).\eta(\delta)=\frac{\{2\Phi(\delta)-1\}^2}{\operatorname{Var}(|\delta+Z|)}. It decreases from 1 at δj=0\delta_j=0 to 0.02 at ∣δj∣=2|\delta_j|=2 and 0.001 at ∣δj∣=3|\delta_j|=3, characterizing when higher-order effects become negligible. The ratio μj/σj\mu_j/\sigma_j, already estimated by RoSHAP, also increases with ∣δj∣|\delta_j| (1.32 at 0, 2.09 at 2), indicating the signal regime at no extra cost.

Tree learners. Tree-ensemble attributions have no closed form, but Mentch and Hooker (2016) and Wager and Athey (2018) show that higher-order effects become negligible for subsampled random forests through first-order Hájek projections, and Zhou and Hooker (2022) establish a Gaussian limit for a regularized form of stochastic gradient boosting. These results concern predictions, not SHAP attributions, so we test A2 directly.

Empirical assessment. We decompose each feature's attribution variance into first-order split effects, higher-order split effects (A2), and learner randomness, isolated by refits that change only the random seed with the split held fixed; the same refits test whether learner randomness is Gaussian (A4).

In simulation (Table 1), higher-order effects are small for leading features across all learners. For the bottom 50 features, all null, they dominate under logistic regression (0.97), as our theory predicts, but stay small for tree ensembles (0.02–0.23). Learner randomness is approximately Gaussian for leading features.

Table 1. Simulated data (n=2400n=2400, d=1000d=1000; 50 signal, 950 null features), one data set, 1,000 refits per learner. Shares: fraction of a feature's attribution variance across refits due to each source; medians over features, taken separately. A4: % of Shapiro–Wilk tests passing at the 5% level. Logistic regression is deterministic.

LearnerFeaturesFirst-orderHigher-order (A2)Learner randomnessGaussian (A4)
LogisticTop 100.820.180—
Bottom 500.030.970—
XGBoostTop 100.640.120.2992%
Bottom 500.020.090.860%
LightGBMTop 100.840.110.0690%
Bottom 500.030.210.7154%
CatBoostTop 100.850.070.0895%
Bottom 500.030.020.920%
Random forestTop 100.500.140.3692%
Bottom 500.050.230.7441%

On Musk (Table 2), higher-order effects for weak features are large under logistic regression (0.75 versus 0.11 for leading features) but small for tree ensembles (0.00–0.15); learner randomness is approximately Gaussian for leading features (82–98%).

Table 2. Musk (Version 2), 500 refits per learner; as in Table 1.

LearnerFeaturesFirst-orderHigher-order (A2)Learner randomnessGaussian (A4)
LogisticTop 100.890.110—
Bottom 500.250.750—
XGBoostTop 100.670.010.2692%
Bottom 500.240.150.5556%
LightGBMTop 100.630.000.3088%
Bottom 500.160.120.6550%
CatBoostTop 100.500.030.5782%
Bottom 500.040.000.9135%
Random forestTop 100.330.000.6898%
Bottom 500.270.000.7267%

Summary and implications. Our analysis establishes sufficient conditions for A2 in linear learners and identifies its limitations near null coefficients. For tree ensembles, existing theory and our experiments support A2 and approximately Gaussian learner noise for leading features. These findings clarify the learner properties and signal regimes under which RoSHAP's Gaussian approximation is reliable.

AuthorsOpenReview ↗

Response to Q2: Larger sample sizes and higher feature dimensionalities

We thank the reviewer for this suggestion. We extended both the synthetic and real-data evaluations to examine how Gaussian approximation accuracy, uncertainty estimation, and ranking stability change with increasing sample size and feature dimensionality. Detailed results are provided below.

Controlled scaling in nn and dd. We expanded the simulations to n∈{600,1200,2400}n\in\{600,1200,2400\} and d∈{100,1000}d\in\{100,1000\}, with 50 signals of decreasing strength, five independent datasets per setting, and 1,000 refits, using XGBoost and L2-penalized logistic regression (Table 3).

For XGBoost, Gaussian approximation accuracy and ranking stability improve with increasing nn at both dimensions. At d=1000d=1000, the KS distance decreases from 0.043 to 0.023, while signal recovery increases from 40.2 to 48.1 out of 50. Logistic regression maintains accurate Gaussian approximations (KS: 0.019–0.024) and stable rankings (Spearman: 0.980–0.994). Across both learners, 30 refits yield KS distances close to the Gaussian reference (0.075–0.080 versus 0.076), with approximately 94% coverage for the mean and 87–88% for the standard deviation.

Table 3. Scaling analysis on simulated data (mean over five datasets, 1,000 refits each). KS: Gaussian fit from 30 or all refits versus the full empirical distribution, evaluated on the top 10 signals (Gaussian reference: 0.076 and 0.019). Coverage: nominal 95% intervals for μj\mu_j and σj\sigma_j from 30 refits. Spearman and Top 10: ranking agreement between 30 and all refits. Recovered: true signals among the top 50.

XGBoost

ddnnKS 30 refitsKS allCoverage μj\mu_jCoverage σj\sigma_jSpearmanTop 10Recovered
1006000.0800.0360.9380.8770.9859.245.3
10012000.0780.0300.9400.8810.9939.447.4
10024000.0760.0240.9420.8840.9969.649.2
10006000.0800.0430.9390.8740.9688.840.2
100012000.0770.0320.9390.8840.9879.445.8
100024000.0750.0230.9390.8840.99410.048.1

Logistic regression

ddnnKS 30 refitsKS allCoverage μj\mu_jCoverage σj\sigma_jSpearmanTop 10Recovered
1006000.0750.0240.9410.8820.9809.248.0
10012000.0750.0230.9420.8830.9839.047.6
10024000.0760.0210.9390.8820.9869.046.3
10006000.0750.0190.9400.8820.9889.447.9
100012000.0760.0220.9400.8810.9929.248.6
100024000.0760.0200.9410.8810.9949.448.4

Higher-dimensional real-data evaluations.

We further evaluated Bioresponse and Christine, containing 1,636–1,776 features, using XGBoost, LightGBM, and L2-penalized logistic regression, with eight training splits and 500 refits per split (Table 4).

Across both datasets and all three learners, Gaussian fits from 30 refits achieve KS distances close to the Gaussian reference (0.079–0.089 versus 0.080), with approximately 94% coverage for the mean and 86–89% for the standard deviation. Ranking agreement remains high for XGBoost and LightGBM (Spearman: 0.929–0.948, with 9 of the top 10 features recovered), while logistic regression retains 8 of the top 10 features despite lower overall rank correlations (0.810–0.815).

Table 4. Higher-dimensional real-data evaluations (median over eight training splits, 500 refits each). Logistic: L2-penalized regression on standardized features. KS and coverage as in Table 2 (Gaussian reference: 0.080 and 0.027). Spearman and Top 10: ranking agreement between 30 and all refits.

DataLearnerKS 30 refitsKS allCoverage μj\mu_jCoverage σj\sigma_jSpearmanTop 10
Bioresponse (n=3751n=3751, d=1776d=1776)XGBoost0.0870.0520.9380.8690.9409.0
LightGBM0.0890.0550.9360.8630.9489.0
Logistic0.0790.0280.9410.8860.8108.0
Christine (n=5418n=5418, d=1636d=1636)XGBoost0.0890.0520.9380.8670.9299.0
LightGBM0.0870.0560.9380.8660.9349.0
Logistic0.0810.0280.9400.8840.8158.0

The additional experiments provide direct evidence of RoSHAP's scalability and reliability across larger sample sizes and feature dimensions. The Gaussian approximation improves with increasing sample size for XGBoost and remains accurate across the evaluated settings for logistic regression. Moreover, 30 refits provide useful uncertainty estimates and consistent feature rankings, supporting the practical efficiency of RoSHAP in higher-dimensional applications.

Thank you for your responses. I will raise my score.

Reviewer_EBiFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RoSHAP, a framework for quantifying feature attribution uncertainty through repeated data resampling and model refitting. It constructs distributions of aggregated absolute SHAP values and ranks features by their average rank across refits. Under certain regularity conditions, the authors establish a Gaussian approximation for feature-level attribution distributions and derive Monte Carlo precision formulas to guide the number of refits. Experiments on synthetic and real datasets demonstrate attribution variability, improved signal recovery over single-fit SHAP, and competitive downstream predictive performance. Reduced-budget experiments suggest that approximately 30 refits can provide useful distributional approximations and stable rankings in the evaluated settings.

Critical strengths
  1. The proposed RoSHAP addresses a practically important limitation of single-fit explanations by quantifying attribution variability across data splits and model refits.

  2. The theoretical analysis provides explicit conditions for Gaussian approximation and Monte Carlo precision formulas to guide computational budgets.

  3. Experiments across synthetic data, real datasets, and multiple learners demonstrate improved signal recovery over single-fit SHAP and competitive predictive performance.

Critical weaknesses
  1. Limited methodological novelty: The resample–refit procedure and rank aggregation largely follow established approaches, leaving the distributional analysis as the main potential contribution.

  2. Uncontrolled computational budgets: Baselines use default configurations without compute-matched comparisons, making performance differences difficult to separate from differences in computational investment.

  3. Disconnect between theory and downstream gains: Gaussian fitting does not enter the average-rank selection rule, so signal-recovery and predictive improvements do not demonstrate its utility.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Q1: Clarifications on novelty

We thank the reviewer for raising these important questions.

RoSHAP's methodological novelty lies in its distributional characterization of aggregated feature attributions. The method allows characterizing how feature importance varies across model refits and using this information to quantify the reliability of feature importance conclusions. Existing resampling-based methods use repeated attributions to select features, test significance, or verify rankings, or simply average them across resamples. In contrast, RoSHAP goes beyond these objectives by estimating not only how important each feature is on average, but also how much its importance varies and how likely one feature is to outrank another across refits.

Our contributions are threefold. First, we establish conditions under which aggregated feature attributions approximately follow a Gaussian distribution, allowing their variability to be characterized using a mean and standard deviation. Second, we quantify how accurately these quantities can be estimated from a limited number of refits and provide theoretical guidance on the number of refits needed. Third, we use this distributional information to obtain robust feature rankings through expected-rank aggregation and quantify ranking stability through pairwise ordering probabilities.

We validate these contributions through experiments across simulated and real-world datasets with different learners. Q–Q plots and Kolmogorov–Smirnov (KS) distances assess the accuracy of the Gaussian approximation, while confidence interval coverage evaluates the precision of parameter estimation. We further compare Spearman rank correlations and top-feature agreement between limited and full refit runs to assess ranking stability. Across the evaluated settings, as few as 30 refits can closely approximate attribution distributions and rankings obtained from 1,500–3,500 refits, demonstrating the practical utility and computational efficiency of RoSHAP.

AuthorsOpenReview ↗

Response to Q2: Computational budgets design

We thank the reviewer for this suggestion. In the original experiments, we intentionally used the default settings to ensure a fair and representative comparison under the methods' intended operating conditions. This design preserves the methods' established configurations rather than imposing an arbitrary common refitting budget that may not align with their individual procedures or recommended usage. Our objective was to compare the methods as they are typically applied in practice, rather than artificially standardizing a single computational parameter across fundamentally different approaches.

To directly address the reviewer's concern, we conducted additional experiments using a common budget of 30 model refits for RoSHAP, PowerSHAP, SHAPicant, and BorutaSHAP to evaluate their relative performance under the same refitting budget. The simulation results are shown in the table below.

Table. Number of true signals ranked in the top ss, mean and range (minimum–maximum) over thirty independent data sets per setting, n=600n = 600 rows each.

Method(d,s)=(100,10)(d,s)=(100,10)(d,s)=(100,50)(d,s)=(100,50)(d,s)=(1000,50)(d,s)=(1000,50)
Single-fit SHAP7.2 (5–9)40.4 (36–43)31.5 (28–37)
PowerSHAP (30)8.7 (8–9)45.1 (44–47)41.2 (39–44)
SHAPicant (30)8.7 (8–9)44.8 (43–47)38.1 (34–40)
BorutaSHAP (30)8.1 (7–9)43.0 (41–46)36.6 (34–41)
Goldwasser & Hooker8.9 (8–9)44.3 (42–46)38.7 (36–41)
RoSHAP8.8 (8–10)45.7 (44–48)41.4 (39–44)

Under the matched refit budget, RoSHAP remains competitive, achieving the highest average signal recovery in the dense and high-dimensional settings. For the real-data experiments, the selected top 10 features remained unchanged under the revised budgets, resulting in identical downstream prediction performance. These findings confirm that RoSHAP's competitive performance is maintained under equal refitting budgets and that our conclusions do not depend on differences in the original budget settings.

AuthorsOpenReview ↗

Response to Q3: Connection between theory and downstream gains

The Gaussian approximation and average-rank aggregation serve complementary purposes in RoSHAP: robust feature ranking and computationally efficient uncertainty assessment. Average-rank aggregation reduces sensitivity to individual model fits, while the Gaussian approximation enables the stability of feature orderings to be quantified with relatively few refits.

We illustrate this connection in Section 4.2 of the manuscript, where the Gaussian approximation enables estimation of pairwise ordering probabilities from the distributions of attribution differences. The estimated probabilities agree with empirical frequencies within 0.01 in our evaluations. For example, on UJIIndoorLoc, WAP087 outranks WAP051 with probability 0.97, whereas WAP155 outranks WAP156 with probability 0.61, distinguishing stable from uncertain orderings.

Average-rank aggregation further supports robust feature recovery. For example, on MADELON with LightGBM, V5 has a higher mean attribution than V319 (192 vs. 187), placing V5 seventh under mean-based ranking. However, V5 falls outside the top 50 in 22% of refits, compared with only 2% for V319. Average-rank aggregation captures this instability and places V5 14th, reducing the influence of occasionally high attributions and producing a more stable feature ranking.

These results demonstrate the value of RoSHAP's distributional framework: attribution variability informs robust feature rankings, while the Gaussian approximation efficiently quantifies the reliability of feature orderings derived from the same refitting process. This enables practitioners to identify important features while also assessing how consistently their relative importance is supported across plausible model refits.

Dear Reviewer,

As the discussion period draws to a close, we would like to kindly follow up on our rebuttal. We hope our responses have addressed your questions and clarified our contributions. We would appreciate it if you could revisit our responses and let us know whether any concerns remain.

Thank you again for your time and thoughtful feedback.

Best regards,
The Authors

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

This paper studies uncertainty in feature attribution arising from variation in data sampling and model fitting. Rather than summarizing repeated attributions only through statistics or rankings, the proposed framework treats the attributions obtained through repeated resampling and model refitting as distributions, capturing both attribution magnitude and variability. The paper further establishes that asymptotic the aggregated attribution score is asymptotically Gaussian under regularity conditions, and uses this result to quantify estimation precision under a finite computational budget. The framework is evaluated on synthetic and real-world datasets using multiple predictive models

Critical strengths
  1. The paper presents a principled distributional treatment of attribution uncertainty. It goes beyond a single fitted model or simple stability summaries by treating attributions obtained across resampling and model refitting as random quantities, providing a unified characterization of attribution magnitude, variability, and estimation precision.

  2. The paper provides an integration of theory and empirical evaluation. The theoretical analysis provides support for the Gaussian approximation and finite-replicate precision assessment, while the experiments comprehensively examine attribution variability, Gaussian diagnostics, computational efficiency, and downstream prediction/feature-selection performance across multiple learners and datasets.

Critical weaknesses
  1. The practical interpretation of the quantified uncertainty is somewhat limited. The proposed distributions characterize attribution variability induced by the particular resampling and model refitting procedure, rather than uncertainty in the attribution of a fixed deployed model or in a feature's underlying “true” importance. The practical value of this uncertainty characterization could therefore be explained more clearly, particularly for users interested in interpreting a specific fitted model.

  2. Although multiple datasets and learners are considered, the experimental evaluation is limited to relatively conventional tabular learning settings and uses models including LightGBM, XGBoost, CatBoost, random forests, and a penalized linear model. It therefore remains unclear how well the framework, its theoretical assumptions, and the claimed computational efficiency extend to more complex and stochastic models, where feature attribution may be particularly valuable and repeated model refitting considerably more expensive.

Questions for authors
  1. Could you clarify the practical interpretation and intended use of the attribution uncertainty quantified by the proposed resampling and model refitting procedure?

  2. Could you discuss how the proposed framework, its theoretical assumptions, and computational efficiency might extend to more complex and stochastic models?

  3. Could you clarify how the proposed distributional treatment differs from existing feature attribution uncertainty quantification approaches that also analyse distributions obtained through resampling?

4 comments
AuthorsOpenReview ↗

Response to Q1: Practical interpretation and intended use of attribution uncertainty

We appreciate the reviewer's thoughtful questions and constructive suggestions.

The uncertainty quantified by RoSHAP should be interpreted as the stability of feature importance conclusions across plausible model refits. As the reviewer notes, this uncertainty differs from uncertainty about a feature's "true" importance or about a fixed model's SHAP values. A fixed model's SHAP values vary only through approximation error in computing them.

The intended use of RoSHAP is to help practitioners assess the reliability of feature importance conclusions. By quantifying attribution variability and pairwise ranking probabilities, RoSHAP allows users to distinguish consistently important features from those whose importance or relative ranking is sensitive to model refitting. This supports more informed interpretation, feature selection, and reporting of feature importance rather than relying solely on a single fitted model's ranking.

For example, on UJIIndoorLoc, WAP155 has mean 8.66×1048.66\times10^4 and standard deviation 1.01×1041.01\times10^4 across refits, and WAP156 has mean 8.15×1048.15\times10^4 and standard deviation 1.00×1041.00\times10^4. When comparing them WAP155 outranks WAP156 in 61% of refits, indicating an uncertain ordering. In contrast, WAP087 (mean 7.62×1047.62\times10^4, standard deviation 0.90×1040.90\times10^4) outranks WAP051 (mean 5.26×1045.26\times10^4, standard deviation 0.89×1040.89\times10^4) in 97% of refits, suggesting a substantially more stable ordering. These probabilities can be reliably estimated with as few as 30 refits using the Gaussian approximation, making such uncertainty assessments computationally practical.

We will clarify this practical interpretation and the intended use of RoSHAP in Section 4.2 of the manuscript.

AuthorsOpenReview ↗

Response to Q2: Extension to complex and stochastic models

RoSHAP's resampling and refitting framework is model-agnostic and captures both training-data variability and internal learner randomness. The Gaussian approximation remains valid when A2–A3 characterize resampling variation and A4's conditions on learner randomness hold. Computational efficiency depends on training and attribution costs, but when Gaussianity holds, relatively few refits can provide reliable distributional estimates.

To evaluate its applicability on more complex models, we evaluated RoSHAP using an FT-Transformer on Musk (Version 2) with 500 refits and its approximation with 30 refits. Results are in the table below. For the FT-Transformer, 95% confidence intervals based on 30 refits achieve near-nominal coverage for the mean (0.932–0.940) and reasonably high coverage for the standard deviation (0.816–0.889). The Gaussian distributions fitted using 30 refits yield Kolmogorov–Smirnov (KS) distances of 0.086–0.101 from the empirical distributions of all 500 refits, close to the 0.080 benchmark for exactly Gaussian data. These results demonstrate that the Gaussian approximation remains effective for the FT-Transformer with a limited number of refits.

Table. FT-Transformer on Musk (Version 2), ten leading features by expected rank. Full run: mean μj\mu_j and standard deviation σj\sigma_j of the aggregated attribution over all 500 refits. Coverage: share of 20,000 samples of 30 refits whose nominal 95% interval for the mean, or for the standard deviation that contains the full-run value. KS: Kolmogorov–Smirnov distance between a Gaussian, with mean and standard deviation estimated from 30 refits (median over 200 draws, 5th–95th percentile in brackets) or from all 500, and the empirical distribution of all 500 refits.

FeatureMean μj\mu_jSD σj\sigma_jCoverage μj\mu_jCoverage σj\sigma_jKS (30 refits)KS (500 refits)
f961,3873340.9380.8480.093 [0.048, 0.176]0.069
f1241,0542760.9380.8650.096 [0.048, 0.186]0.057
f949643060.9360.8160.101 [0.052, 0.204]0.076
f699743130.9330.8420.087 [0.047, 0.161]0.061
f1259753240.9320.8320.099 [0.055, 0.194]0.071
f998482840.9400.8370.087 [0.041, 0.168]0.044
f888202580.9340.8640.087 [0.052, 0.184]0.058
f1477982320.9370.8890.096 [0.056, 0.182]0.066
f718162600.9360.8700.086 [0.041, 0.175]0.038
f927242040.9390.8580.091 [0.049, 0.174]0.060

To demonstrate applicability beyond tabular data and to more complex models, we fine-tuned a pretrained Vision Transformer (ViT-Base/16) on CIFAR-10 across 50 independent refits, achieving validation accuracies of 98.3–98.9%. For each refit, we computed SHAP attributions for all 576 patches of the same test images. Shapiro–Wilk tests did not reject normality for 70% of the leading patch attributions. The results and images are added in the appendix.

These experiments demonstrate RoSHAP's applicability to more complex and stochastic models. The Gaussian approximation enables accurate distributional estimation with much fewer refits when its assumptions hold. Even when Gaussianity fails, RoSHAP remains robust, reliably recovering attribution means, standard deviations, and feature rankings with relatively few refits.

AuthorsOpenReview ↗

Response to Q3: Differences from existing attribution uncertainty quantification approaches

Existing resampling-based feature attribution methods also examine variability in feature importance estimates, but primarily use this variability for feature selection, statistical significance testing, or ranking verification. Among our baselines, PowerSHAP, SHAPicant, and BorutaSHAP use repeated attributions to assess feature relevance against random, permuted-target, or shadow-feature references. Goldwasser and Hooker instead focus on the reliability of feature rankings under attribution estimation noise for a fixed model, rather than variability across model refits.

RoSHAP builds on the resampling and refitting framework but differs in its treatment and use of the resulting attribution distributions. RoSHAP treats each feature's attribution distribution across model refits as the object of inference. Specifically, our contributions are threefold:

First, we establish a Gaussian approximation under stated conditions, providing a theoretical foundation for characterizing attribution uncertainty. Second, we quantify the precision of Gaussian parameter estimates and provide guidance on the number of refits needed, demonstrating empirically that as few as 30 refits can closely approximate results from 1,500–3,500 refits. Third, RoSHAP combines robust feature rankings through expected-rank aggregation with pairwise ranking probabilities that quantify the stability of feature orderings across refits.

Thus, while existing methods primarily use resampling variability to determine feature relevance or verify rankings, RoSHAP provides a distributional characterization of feature importance across plausible model refits. This enables practitioners to assess not only which features are important, but also how reliable their relative importance conclusions are.

Dear Reviewer,

As the discussion period draws to a close, we would like to kindly follow up on our rebuttal. We hope our responses have addressed your questions and clarified our contributions. We would appreciate hearing whether any concerns remain.

Thank you again for your time and thoughtful feedback.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses to All Reviewers

We thank the reviewers and the AC for their time and constructive feedback. We would like to briefly summarize our main contributions, the strengths recognized by the reviewers, and the additional analyses and experiments conducted in response to their comments.

Main contributions. RoSHAP provides a computationally efficient distributional framework for quantifying feature attribution uncertainty across resampling and model refitting. Our contributions are threefold. First, we establish conditions under which aggregated feature attributions are approximately Gaussian. Second, we quantify the precision of estimating their means and standard deviations and provide guidance on the number of refits needed. Our experiments show that as few as 30 refits can closely approximate results from thousands of refits, making uncertainty quantification practical at a much lower computational cost. Third, we provide robust feature rankings, achieving the highest signal recovery among baselines, while enabling the estimation of pairwise ordering probabilities to assess ranking stability.

Strengths recognized by reviewers. Reviewer grRS highlighted the principled distributional treatment and the integration of theory and empirical evaluation. Reviewer EBiF recognized the explicit conditions for Gaussian approximation, Monte Carlo precision analysis, and improved signal recovery over single-fit SHAP. Reviewer ckoZ acknowledged that the theoretical proof is sound under A1–A4 and that the reduced-budget experiments support the framework's practical utility.

Additional analyses and experiments. In response to the reviewers' concerns, we conducted several new analyses and substantially expanded our experiments:

  • Theoretical justification and assumption diagnostics (ckoZ, Q1; Tables 1–2): We sharpened our theoretical results by deriving sufficient conditions for A2 in linear learners and characterizing the signal regimes where the Gaussian approximation is reliable. We further conducted variance decomposition experiments across five learners on simulated and real data, providing empirical support for the assumptions on leading features. For weak features, expected-rank aggregation remains robust even when the Gaussian approximation does not hold.
  • Larger-scale experiments (ckoZ, Q2; Tables 3–4): We expanded the simulations to larger sample sizes and feature dimensions and added two higher-dimensional real datasets. Using only 3–6% of the full refit budget, RoSHAP achieves accurate Gaussian approximations and recovers 80–100% of the top-ranked features identified by the full runs.
  • Baseline comparisons (EBiF, Q2): Our original comparisons used each baseline's default configuration by design, reflecting their intended use in practice. To address the reviewer's concern about computational budgets, we additionally evaluated all resampling-based methods under a common budget of refits. RoSHAP continues to achieve the highest average signal recovery on the simulation data. On real datasets, the selected top-10 features and downstream prediction performance remain strong.
  • Extension to complex models (grRS, Q2; appendix): We added experiments with an FT-Transformer on Musk and a fine-tuned Vision Transformer on CIFAR-10, demonstrating RoSHAP's applicability to more complex and stochastic learners. The results also support accurate distributional estimation with substantially fewer refits, reinforcing its computational efficiency.

Additional clarifications. We also clarified the practical interpretation of attribution uncertainty as the stability of feature importance conclusions across plausible model refits (grRS, Q1), how RoSHAP differs from existing resampling-based approaches (grRS, Q3; EBiF, Q1), and how the Gaussian approximation complements expected-rank aggregation by quantifying pairwise ranking probabilities (EBiF, Q3).

We appreciate the reviewers' suggestions, which have helped us strengthen both the theoretical analysis and empirical evaluation. We hope these additional results and clarifications address their main concerns and provide a clearer picture of RoSHAP's contributions and practical value.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.