- 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
-
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.
-
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
-
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.
-
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
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 and . We expanded the simulations to and , 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 at both dimensions. At , 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 and from 30 refits. Spearman and Top 10: ranking agreement between 30 and all refits. Recovered: true signals among the top 50.
XGBoost
| KS 30 refits | KS all | Coverage | Coverage | Spearman | Top 10 | Recovered | ||
|---|---|---|---|---|---|---|---|---|
| 100 | 600 | 0.080 | 0.036 | 0.938 | 0.877 | 0.985 | 9.2 | 45.3 |
| 100 | 1200 | 0.078 | 0.030 | 0.940 | 0.881 | 0.993 | 9.4 | 47.4 |
| 100 | 2400 | 0.076 | 0.024 | 0.942 | 0.884 | 0.996 | 9.6 | 49.2 |
| 1000 | 600 | 0.080 | 0.043 | 0.939 | 0.874 | 0.968 | 8.8 | 40.2 |
| 1000 | 1200 | 0.077 | 0.032 | 0.939 | 0.884 | 0.987 | 9.4 | 45.8 |
| 1000 | 2400 | 0.075 | 0.023 | 0.939 | 0.884 | 0.994 | 10.0 | 48.1 |
Logistic regression
| KS 30 refits | KS all | Coverage | Coverage | Spearman | Top 10 | Recovered | ||
|---|---|---|---|---|---|---|---|---|
| 100 | 600 | 0.075 | 0.024 | 0.941 | 0.882 | 0.980 | 9.2 | 48.0 |
| 100 | 1200 | 0.075 | 0.023 | 0.942 | 0.883 | 0.983 | 9.0 | 47.6 |
| 100 | 2400 | 0.076 | 0.021 | 0.939 | 0.882 | 0.986 | 9.0 | 46.3 |
| 1000 | 600 | 0.075 | 0.019 | 0.940 | 0.882 | 0.988 | 9.4 | 47.9 |
| 1000 | 1200 | 0.076 | 0.022 | 0.940 | 0.881 | 0.992 | 9.2 | 48.6 |
| 1000 | 2400 | 0.076 | 0.020 | 0.941 | 0.881 | 0.994 | 9.4 | 48.4 |
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.
| Data | Learner | KS 30 refits | KS all | Coverage | Coverage | Spearman | Top 10 |
|---|---|---|---|---|---|---|---|
| Bioresponse (, ) | XGBoost | 0.087 | 0.052 | 0.938 | 0.869 | 0.940 | 9.0 |
| LightGBM | 0.089 | 0.055 | 0.936 | 0.863 | 0.948 | 9.0 | |
| Logistic | 0.079 | 0.028 | 0.941 | 0.886 | 0.810 | 8.0 | |
| Christine (, ) | XGBoost | 0.089 | 0.052 | 0.938 | 0.867 | 0.929 | 9.0 |
| LightGBM | 0.087 | 0.056 | 0.938 | 0.866 | 0.934 | 9.0 | |
| Logistic | 0.081 | 0.028 | 0.940 | 0.884 | 0.815 | 8.0 |
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.
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∣, where the coefficient is fitted on the complement of O and cj 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), 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∣ with Z∼N(0,1), where δj measures the coefficient's separation from zero relative to its variability across refits; δj=0 gives the half-normal limit identified by the reviewer. The higher-order share is 1−η(δj), where η(δ)=Var(∣δ+Z∣){2Φ(δ)−1}2. It decreases from 1 at δj=0 to 0.02 at ∣δj∣=2 and 0.001 at ∣δj∣=3, characterizing when higher-order effects become negligible. The ratio μj/σj, already estimated by RoSHAP, also increases with ∣δj∣ (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=2400, d=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.
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.
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.