Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

The many Shapley values for model explanation

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read Shapley's uniqueness theorem does not select a single explanation method; Baseline Shapley gets its own uniqueness theorem.

desk verdict Sharp conceptual critique of Shapley-based attribution, but the advertised uniqueness theorem for BShap is misstated and needs an added axiom. read the letter →

arxiv 1908.08474 v2 pith:C7CCEK5F submitted 2019-08-22 cs.AI cs.LGecon.TH

classification cs.AIcs.LGecon.TH MSC 91A1291A80
keywords ShapleyvaluemodelexplanationaxiomaticattributionBaselineIntegratedGradientsconditionalexpectationscost-sharinguniqueness
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 argues that the textbook uniqueness of the Shapley value does not carry over to machine-learning explanation, because the value can be applied to the model in several nonequivalent ways. It shows that the conditional-expectation version (CES) can violate basic desiderata: it can give nonzero credit to features the model ignores and can break linearity, especially when continuous features rarely repeat in the training data. The paper then proposes Baseline Shapley (BShap), which models a feature's absence by substituting a baseline value, and proves that BShap is the unique attribution method satisfying a specific set of axioms. Integrated Gradients receives a parallel uniqueness theorem. The practical point is that a model-explanation tool should state which Shapley variant it implements and under which axioms it is justified.

What carries the argument

The load-bearing object is the set function that turns a model into a cooperative game. BShap defines $v(S) = f(x_S; x'_{N\setminus S})$, meaning features outside $S$ are pinned to baseline values, and then applies the usual Shapley average of marginal contributions. The argument's engine is Theorem 4.1, which shows that any attribution method obeying Linearity and Affine Scale Invariance reduces to a difference of two cost-sharing problems: the model is decomposed as $f_1 - f_2$ with both parts non-decreasing, using the infimum of its partial derivatives, provided that infimum exists (bounded derivatives). That reduction imports the uniqueness theorem for Shapley–Shubik sharing, giving Corollary 4.3, and the analogous theorem for Aumann–Shapley, giving Corollary 4.4. A second mechanism is Lemma 4.8, identifying CES over an independent distribution with Random Baseline Shapley, which localizes CES's failures to the dependence and sparsity of the chosen distribution.

What would settle it

The sharpest test is to search for an attribution rule that satisfies Linearity, Dummy, ASI, Demand Monotonicity, and Symmetry but differs from BShap on a simple problem like $\min(x_1, x_2)$; its existence would refute Corollary 4.3 as stated. A faster negative check is to compute the infimum of partial derivatives for a model like $f(x) = e^x$; because it is $-\infty$, the paper's decomposition into a difference of non-decreasing functions does not exist, so the theorem's domain is narrower than the general attribution problem.

Watch

Extended reading notes

Core claim

The paper's central claim is that Shapley's uniqueness theorem, often cited to justify attribution methods, does not select a single explanation method because the Shapley value can be applied to a model in several non-equivalent ways. The conditional-expectation operationalization (CES) is shown to be fragile: when continuous features have unique values in the empirical distribution, it hands out equal attributions to every feature, including ones the model does not use, and it can violate linearity even for simple two-variable functions. Against this, the paper proves that Baseline Shapley (BShap), defined by replacing absent features with baseline values, is the unique attribution method satisfying Linearity, Dummy, Affine Scale Invariance, Demand Monotonicity, and Symmetry, by reducing any attribution problem with bounded derivatives to a difference of two cost-sharing problems and importing the known uniqueness of Shapley–Shubik sharing. Integrated Gradients gets a parallel uniqueness theorem with Proportionality in place of Demand Monotonicity. The paper also shows that BShap is itself a limiting case of CES under a specially chosen distribution, so the practical difference between BShap and empirical CES is largely a difference in how the feature distribution is handled.

Load-bearing premise

The proof's reduction assumes bounded partial derivatives and silently uses an Efficiency condition that is not listed among the axioms; without both, the claimed uniqueness for BShap is not established.

Editorial extensions

If this is right

  • BShap assigns zero attribution to dummy features and respects linearity, so it avoids the counterintuitive CES behavior the paper documents on the diabetes case study.
  • Any use of CES must specify the feature distribution; with the empirical distribution on continuous data, it can collapse to equal attributions across all features.
  • IG and BShap are both axiomatically grounded but differ; IG's proportionality forces smooth interpolation along the path from baseline to explicand, while BShap mixes baseline and explicand feature values directly.
  • When feature distributions are independent, CES and RBShap coincide, so independence assumptions, not conditional expectation per se, drive axiom satisfaction.
  • The baseline in BShap is a modeling choice that can encode the explanation context, such as fixing immutable features to the explicand's values for loan decisions.

Reading between the lines

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

  • The reduction theorem suggests a general recipe: for any attribution method that is linear and scale-invariant, the search for a uniqueness theorem can be reframed as choosing a continuous cost-sharing rule, so the space of defensible explanation methods is larger than the two corollaries enumerate.
  • The bounded-derivative condition means the uniqueness result is formally silent for models such as deep networks with unbounded gradients; BShap is still defined for them, but its axiomatic backing is not proven.
  • Lemma 4.6 could be inverted as a diagnostic: by constructing the two-point distribution that makes CES approximate BShap, a practitioner can quantify how much of the disagreement between explanation methods is purely distributional choice.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper argues that the Shapley value's classical uniqueness theorem does not resolve the ambiguity of model explanation, because the value can be operationalized through different set functions (CES, BShap, RBShap, IG) that reference the model, the feature distribution, and the explanation context in different ways, leading to very different and sometimes counterintuitive attributions. It studies CES in detail, showing failures of Dummy and Linearity, and proposes BShap as a baseline-based method. The paper reduces model explanation to cost-sharing and claims uniqueness results for BShap and IG via known cost-sharing theorems, with an illustrative diabetes-prediction experiment.

Significance. The paper makes a valuable conceptual contribution by clarifying that existing uniqueness results for the Shapley value do not select a unique attribution method, because the choice of set function is left open. The concrete counterexamples to CES (e.g., dummy features receiving nonzero attributions in §3.3) are correct and practically important. The reduction from attribution to cost-sharing in §4.2 is an elegant and potentially useful bridge. However, the advertised uniqueness results are currently stated without an Efficiency axiom while their proofs rely on uniqueness theorems that include it, and the affine transformation in the reduction theorem has a sign error. Both issues are load-bearing for the paper's central claim but are fixable with reasonable changes.

major comments (2)
  1. [§4.2, Corollaries 4.3 and 4.4, and Appendix D.4] Both uniqueness results omit the Efficiency axiom from their stated axiom lists, yet their proofs invoke uniqueness theorems (Theorem 4.2 and Theorem 3 of [7]) whose hypotheses include Efficiency. The five listed axioms do not imply Efficiency: for any λ∈(0,1), the method M_λ = λ·BShap satisfies Linearity, Dummy, ASI, Demand Monotonicity, and Symmetry, but its attributions sum to λ(f(x)-f(x')), not f(x)-f(x'). Thus Corollary 4.3 is false as stated, and the same construction (λ·IG) refutes Corollary 4.4. The proof in Appendix D.4 silently uses Efficiency when it applies Theorem 4.2. The fix is to add Efficiency to the axiom list (matching Theorem 4.2) and to argue that the induced cost-sharing method inherits it; the present statement gives a uniqueness claim that is not true.
  2. [§4.2, Theorem 4.1 proof] The affine transformation step in the proof of Theorem 4.1 chooses c = 1 when the original explicand value x_i is non-negative and c = -1 otherwise, with d = -x'_i·c. The transformed explicand value is then c·x_i + d = c·(x_i - x'_i). This is non-negative if and only if c has the sign of x_i - x'_i, not the sign of x_i. For example, with x_i = 1 and x'_i = 2, the paper's rule gives c = 1 and transformed value -1 < 0, so the reduced problem is not a cost-sharing problem as claimed. The proof should set c = sign(x_i - x'_i) and d = -c·x'_i. As written, the reduction, and hence the transfer of the cost-sharing uniqueness result to all attribution problems, is not established.
minor comments (5)
  1. [§2.6, ASI definition] The equation defining Affine Scale Invariance is easy to misread: the relationship between f1, f2, c, d, and the explicand/baseline values should be stated in a way that makes clear which function is expressed in which coordinate system. Please rewrite it to avoid ambiguity.
  2. [§3.2, Remark 3.2] The sentence 'the first feature gets attribution f(x)-E[f] while all the other features get zero. Therefore all the variables get equal attributions' is confusing because equal attributions only arise after averaging over permutations; please clarify that the equality is the Shapley-value average.
  3. [§4.2, Theorem 4.1] The theorem relies on a 'minor technical condition that the derivatives are bounded.' This condition should be stated precisely (e.g., uniformly bounded partial derivatives on the relevant domain), and its restrictiveness should be discussed, since common models such as polynomials on unbounded domains do not satisfy it.
  4. [Appendix A] There is a typo: 'We introdoce' should be 'We introduce.'
  5. [§4.2, Corollary 4.3] For self-containedness, please state the technical conditions from [7] that are being suppressed, or at least indicate which conditions are needed for the transfer via Theorem 4.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: BShap uniqueness is imported from an external cost-sharing theorem after an independent reduction; self-citations are descriptive, not load-bearing.

full rationale

The paper's central derivation chain is not circular. The claimed uniqueness of BShap is obtained by first proving Theorem 4.1, which reduces any attribution method satisfying Linearity and ASI to a difference of two cost-sharing problems, and then invoking Friedman and Moulin's external uniqueness theorem for Shapley-Shubik. This is a genuine reduction, not an equation of the conclusion with an input: the proof transforms the baseline to zero, decomposes the function into monotone parts using the bounded-derivative infimum, and applies Linearity. The stated Corollary 4.3 omits Efficiency from its axiom list even though the proof uses Theorem 4.2, whose uniqueness hypothesis includes Efficiency; this is a correctness flaw in the theorem as stated, and scaled-BShap counterexamples show the stated claim is false. But this does not make the argument circular, since supplying the missing axiom restores the intended external result. The paper's self-citations to Integrated Gradients and prior baseline-attribution work are descriptive references to existing methods, not load-bearing justifications for the uniqueness claim. Lemma 4.6 and Lemma 4.8 are constructive equivalences, not tautologies. No fitted parameter is relabeled as a prediction, and no uniqueness theorem is imported from the authors' own prior work. The bounded-derivative caveat is an explicit technical condition, not a hidden assumption of the conclusion. Accordingly, the paper receives score 0 for circularity, with the noted correctness concern about Corollary 4.3 left to the correctness assessment.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper's central claims rest on external results from cooperative game theory and cost-sharing, plus a technical bounded-derivative assumption. No new physical entities are introduced. The axioms are assumed as desirable properties rather than derived from first principles.

free parameters (1)
  • smoothing parameter for CES = 0.1 and 0.2
    Used in the empirical case study (Figure 1) to define closeness in feature values; not part of the theoretical results, but influences the demonstrated sensitivity of CES.
assumptions (4)
  • standard math Shapley value definition and its uniqueness as the value satisfying Efficiency, Linearity, Dummy, and Symmetry for cooperative games (Shapley 1953).
    Background result used to define the methods and motivate the axioms.
  • domain assumption Friedman-Moulin uniqueness theorems for Shapley-Shubik and Aumann-Shapley in cost-sharing.
    External results (ref [7]) that underpin Corollaries 4.3 and 4.4; the paper assumes these apply to the attribution setting via reduction.
  • domain assumption Bounded partial derivatives of the model function (stated as 'minor technical condition' in Theorem 4.1).
    Ensures the infimum p exists and allows decomposition into non-decreasing functions; if violated, the reduction to cost-sharing is not justified.
  • domain assumption Attribution methods under study satisfy the desirability axioms (Dummy, Efficiency, Linearity, Symmetry, ASI, Demand Monotonicity, Proportionality) as formal properties.
    These axioms are chosen as the normative framework; the paper does not justify them beyond plausibility, and the choice drives which methods are 'unique'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The many Shapley values for model explanation." pith.science (2026). https://pith.science/paper/C7CCEK5F

@misc{pith2026190808474,
  author       = {Pith},
  title        = {Pith review of: The many Shapley values for model explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7CCEK5F}},
  note         = {Machine review of arXiv:1908.08474}
}
read the original abstract

The Shapley value has become a popular method to attribute the prediction of a machine-learning model on an input to its base features. The use of the Shapley value is justified by citing [16] showing that it is the \emph{unique} method that satisfies certain good properties (\emph{axioms}). There are, however, a multiplicity of ways in which the Shapley value is operationalized in the attribution problem. These differ in how they reference the model, the training data, and the explanation context. These give very different results, rendering the uniqueness result meaningless. Furthermore, we find that previously proposed approaches can produce counterintuitive attributions in theory and in practice---for instance, they can assign non-zero attributions to features that are not even referenced by the model. In this paper, we use the axiomatic approach to study the differences between some of the many operationalizations of the Shapley value for attribution, and propose a technique called Baseline Shapley (BShap) that is backed by a proper uniqueness result. We also contrast BShap with Integrated Gradients, another extension of Shapley value to the continuous setting.

Figures

Figures reproduced from arXiv: 1908.08474 by the authors.

Figure 1
Figure 1. Attribution distribution across 20 explicands for four methods, BShap, [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Overview and practical recommendations on using Shapley Values for identifying predictive biomarkers via CATE modeling

    stat.ME 2025-05 conditional novelty 5.0 of 10

    Surrogate SHAP fits an XGBoost model to estimated CATEs and uses TreeSHAP to rank predictive biomarkers, with simulations favoring S-learning in RCTs and R/DR-learning in observational settings.

Reference graph

Works this paper leans on

23 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [7]

    Three methods to share joint costs or surplus

    Friedman, E., and Moulin, H. Three methods to share joint costs or surplus. Journal of Economic Theory 87 , 2 (1999), 275 – 312

  2. [1]

    Explaining individual predictions when features are dependent: More accurate approximations to Shapley values

    Aas, K., Jullum, M., and Løland, A. Explaining individual predictions when features are dependent: More accurate approximations to Shapley values. arXiv e-prints (Mar 2019), arXiv:1903.10464

  3. [2]

    A new inter- action index inspired by the taylor series

    Agarwal, A., Dhamdhere, K., and Sundararajan, M. A new inter- action index inspired by the taylor series. CoRR abs/1902.05622 (2019)

  4. [3]

    Fast algorithms for mining association rules

    Agrawal, R., and Srikant, R. Fast algorithms for mining association rules. In Proc. of 20th Intl. Conf. on VLDB (1994), pp. 487–499

  5. [4]

    J., and Shapley, L

    Aumann, R. J., and Shapley, L. S. Values of Non-Atomic Games . Princeton University Press, Princeton, NJ, 1974

  6. [5]

    Algorithmic transparency via quanti- tative input influence: Theory and experiments with learning systems

    Datta, A., Sen, S., and Zick, Y. Algorithmic transparency via quanti- tative input influence: Theory and experiments with learning systems. In 2016 IEEE Symposium on Security and Privacy (SP) (Los Alamitos, CA, USA, may 2016), IEEE Computer Society, pp. 598–617

  7. [6]

    Least angle regression

    Efron, B., Hastie, T., Johnstone, I., and Tibshirani, R. Least angle regression. Annals of Statistics 32 (2004), 407–499

  8. [8]

    Estimators of relative importance in linear regression based on variance decomposition

    Gr¨omping, U. Estimators of relative importance in linear regression based on variance decomposition. The American Statistician 61, 2 (2007), 139–147

Show all 23 references
  1. [9]

    M., and Gold., R

    Lindeman, R.H., P. M., and Gold., R. Introduction to bivariate and multivariate analysis. Tech. rep., 1980

  2. [10]

    A unified approach to interpreting model predictions

    Lundberg, S., and Lee, S.-I. A unified approach to interpreting model predictions. In NIPS (2017)

  3. [11]

    M., Erion, G

    Lundberg, S. M., Erion, G. G., and Lee, S. Consistent individualized feature attribution for tree ensembles. CoRR abs/1802.03888 (2018)

  4. [12]

    Sobol’ indices and shapley value

    Owen, A. Sobol’ indices and shapley value. SIAM/ASA Journal on Uncertainty Quantification 2, 1 (2014), 245–251

  5. [13]

    On shapley value for measuring importance of dependent inputs

    Owen, A., and Prieur, C. On shapley value for measuring importance of dependent inputs. SIAM/ASA Journal on Uncertainty Quantification 5 , 1 (2017), 986–1002. 15

  6. [14]

    Scikit-learn: Machine learning in Python

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. Scikit-learn: Machine learning in Python. Journal of M...

  7. [15]

    T., Singh, S., and Guestrin, C

    Ribeiro, M. T., Singh, S., and Guestrin, C. ”why should I trust you?”: Explaining the predictions of any classifier. CoRR abs/1602.04938 (2016)

  8. [16]

    Shapley, L. S. A value of n-person games. Contributions to the Theory of Games (1953), 307–317

  9. [17]

    ˇStrumbelj, E., Kononenko, I., and ˇSikonja, M. R. Explaining instance classifications with interactions of subsets of feature values. Data & Knowledge Engineering 68 , 10 (2009), 886–904

  10. [18]

    Axiomatic attribution for multilinear functions

    Sun, Y., and Sundararajan, M. Axiomatic attribution for multilinear functions. CoRR abs/1102.0989 (2011)

  11. [19]

    Axiomatic attribution for deep networks

    Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017 (2017), D. Precup and Y. W. Teh, Eds., vol. 70 of Proceedings of Machin...

  12. [20]

    Regression shrinkage and selection via the lasso

    Tibshirani, R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological) 58 , 1 (1996), 267–288

  13. [21]

    Explaining prediction models and individual predictions with feature contributions

    ˇStrumbelj, E., and Kononenko, I. Explaining prediction models and individual predictions with feature contributions. Knowl. Inf. Syst. 41 , 3 (Dec. 2014), 647–665

  14. [22]

    Young, H. P. Monotonic solutions of cooperative games. International Journal of Game Theory 14 , 2 (Jun 1985), 65–72. 16 A Strong Monotonicity We introdoce an additional axiom called Strong Monotonicity that plays a role in the results in the Appendix. Strong Monotonicity if f...

  15. [23]

    It claims that Missingness, Local Accuracy and Consistency suffice

    claims to show that one of the standard axioms (Symmetry) used in the Shapley axiomatization is redundant within an earlier axiomatization of the Shapley value by [22]. It claims that Missingness, Local Accuracy and Consistency suffice. (The latter two axioms are called Efficiency...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.