- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper proposes a method to bridge integrated gradients (IG based on Aumann-Shapley) and baseline SHAP (based on Shapley-Shubik). Both methods explain the prediction of a black-box model at instance using a reference baseline point and decompose into attribution values per feature. The difference between both approaches is that IG uses the whole hypercube whereas baseline SHAP uses the discrete corners of the hypercube . The hypercube is spanned by the predictions of the model within the d-dimemnsional cube . In game-theoretic terms, IG works with divisible goods that can be continuous, whereas baseline SHAP relies on players that are either present or absent. Both agree, when the underlying model is multilinear.
The main contribution is called Aumann-SHAP, which additionally defines the attribution of feature to any interaction (Definition 3). The sum over all interactions gives an attribution value (Eq. 7), which satisfies efficiency and decomposes the difference between prediction and prediction of the baseline, similar to IG and baseline SHAP. The definition of the method relies on a discretization using resolutions which are defined per feature . The limiting cases recover the share of Möbius coefficients within baselineSHAP () and IG applied the continuous Möbius coefficient (see below), which build the foundation of the "bridge" of both methods (Theorem 1).
The method requires several steps:
- The method is constructed using the Möbius coefficients (Harsanyi dividends) of the baseline SHAP game, and introducing a continuous version of them (Def 1). On an intuitive level, this introduces the Möbius coefficients of the baseline game, where was moved around the hypercube defined via the resolutions .
- Instead of baseline SHAP going directly from to , Aumann-SHAP repeatedly goes along the resolutions and then defines a game at each step using "micro-contributions" of every feature (Def 2).
- These "micro-"Shapley values are then aggregated into Aumann-SHAP. The method for approximating Shapley values is permutation sampling.
- Since the resolutions increase the complexity of the computation, it is shown that there is redundancy in the computation, which for order- interactions becomes with uniform resolutions (Prop 3). The computational "trick" for approximation in practice is that each feature is revealed within the permutation and the resolution steps. So the resolution steps and number of permutations determines the complexity in practice, as well as the number of interactions used.
The paper then showcases AumannSHAP on a synthetic dataset and some real-world datasets, highlighting the application for counterfactuals.
Critical strengths
- Connection between path-methods and baseline SHAP: Exploring deeper connections between IG and baseline SHAP by looking at Möbius coefficients and and more fine-grained interaction contributions is an interesting research direction. The combination of discrete Shapley values and continuous path integrals, such as IG, have recently shown to be promising tools for interactions [2]. However, the interpretation of such measures is harder to justify.
Critical weaknesses
In my view, the current paper and the proposed method use a variety of constructions, which are not well motivated, which limits the interpretability of the approach. Additionally, while the method can be approximated in practice, it is unclear what exactly is being approximated and how well, which limits the practicability of the method. Lastly, the method relies on the Möbius coefficients (interaction pots), which has exponential many terms and the connection with IG and Baseline SHAP break down, once only a subset of pots are chosen.
-
Conceptual framework is not well constructed nor motivated: The paper makes several design choices which are not well motivated, nor justified. The final definition of Aumann-SHAP is not very intuitive: introducing the Möbius coefficients for the whole hypercube by effectively moving and the definition of micro-Shapley as perturbations along a small grid. Both construction appear quite arbitrary, and stacking them on top of each other obscures the interpretation. A clear formulation, such as the path integral in IG, which is then approximated using a discrete grid, is a much more desirable object to work with. Instead Aumann-SHAP's definition relies on the discretization and is only linked to existing approaches in the limit. I believe that there exists a more intuitive reformulation and precise definition, which would benefit the approach (see question below). Lastly, a comparison with other methods such as Integrated Hessians and approaches which also combine baseline SHAP and IG (see e.g. sum of power [1] or meta-attributions [2]) would be helpful to understand the practical differences. Overall, AumannSHAP defines an interaction attribution method and should be treated and evaluated as such.
-
Limited practicability: As discussed above, it is undesirable to use the resolution scales as the defining target. But even given the resolution scale , the method remains hard to compute, because it involves several Shapley values that have to be approximated. While the permutation sampling method aligns well with the definition of AumannSHAP, it is also know to be a relatively poor estimator of Shapley values, where more powerful algorithms exist. Moreover, and importantly, the connection to IG and baseline SHAP require all Möbius coefficients (interaction pots) which is impractical in most applications. Selecting the relevant ones is difficult itself, and restricting to some pots breaks the connection to IG and baseline SHAP.
Minor: The figures could be improved, and the captions seem to use a strange formatting.
[1] A unifying framework to the analysis of interaction methods using synergy functions. Lundstrom & Razaviyayn. ICML, 2023.
[2] The Metagame of Interpretability and Meta-Attributions. Baniecki et al., 2026.
Questions for authors
- The construction of AumannSHAP relies on the Möbius components (interaction surfaces) . These are functions that rely only on features in and AumannSHAP essentially computes where is an attribution that is defined as the "micro"-Shapley value. Due to linearity of both Möbius and the attributions (Shapley value), one could also write this as , i.e. first computing an attribution and then the corresponding Möbius interaction. Using this notion, can your method be related to the Möbius coefficients of the metagame [2], i.e. the Möbius coefficients of a cooperative game that is defined based on first-order attribution ? Making such connections would help to better understand what is actually being computed. Moreover, the introduced first-order attribution (micro-Shapley value) might be an interesting attribution itself worth of study.
- Could AumannSHAP be reformulated without the resolution scale and treat each resolution scale only as an approximation of the actual property that is underlying it (i.e. the integral), where gives the exact value (the intergal) and the "discrete derivative" approximation? I think this construction would mitigate the interpretation problem of choosing .
- Since AumannSHAP "does not require differentiability", how does AumannSHAP look like on well understood mode classes, such tree ensembles and TreeSHAP? Making these connections would help understanding the impact of the design choices.
Third review, from area chair
Dear authors,
As we could not secure a third review from the regular pool, I contribute one, written independently last week, before viewing the two regular reviews.
Paper summary
The Shapley value is a standard attribution technique. It compares a model's performance when a feature x is observed to when that feature is instead held to a generic baseline value x′ (e.g. a mean or conditional mean). This approach is therefore based on a discrete calculation: the specific feature is known or it is not.
Another approach, Sundararajan et al.'s integrated gradients, "are defined as the path integral of the gradients along the straightline path from the baseline x′ to the input x". This approach is therefore continuous.
Yet a third approach recognises that features interact, so that assigning values to individual features may mislead. This assigns values to sets of features.
This manuscript blends these approaches, considering interaction measures (3rd approach) along paths (2nd approach). Specifically, a subset of interacting features forms a hyperplane in feature space: consider e.g. the plane formed by features x1,x2 in the unit cube, when holding x3=1. The Aumann-SHAP presented here imposes an m-step grid on these surfaces, and determines an interaction value at each grid point.
Proposition 1 evaluates the interaction surface at its corners: at 0 it is equal to zero; at 1, it is equal to the Harsanyi dividend.
The paper then builds on this grid, referring to micro-players, micro-coalitions, micro-games, micro-moves... An induced cooperative game is defined, with a player set that is the product of the original player set with the grid points.
Proposition 2 presents the induced game for a set of features as a unanimity game, representing its value function in terms of the Harsanyi dividends of that feature set.
Theorem 1 presents a limit result on the 'Aumann-SHAP' share, a sum of Shapley values along the grid, as grid size becomes infinitessimal.
Proposition 3 takes advantages of symmetries to simplify the calculation of the Aumann-SHAP share.
Experiments then show that the Aumann-SHAP share reproduces known special cases (6.1), can flip signs relative to equal (rather than path-dependent) splits (6.2), can counterfactually move instances in credit scoring problems towards acceptance (6.3), and may identify relevant editable pixels in a transition from an MNIST '1' to a '7' more efficiently than an equal split approach (6.4).
Presentation: 2
Informativeness: 1
Soundness: 3
Critical Strengths
Understanding complex, non-linear, high-dimensional machine learning models is important! Thus, this manuscript is well positioned.
Further, from what I can tell, its results are correct.
Critical Weaknesses
At one end, the explainable AI literature seeks to help a user (whether a developer, a business owner, a customer,...) understand a problem better. At the other, it often has axiomatic foundations (e.g. Shapley value's four axioms, integrated gradients' two...) to formally define an 'explanation'.
Between those two ends, a large collection of papers adds mathematical complexity to an existing measure, without a strong basis in either user insight or axiomatised intuitions. This manuscript feels to me like it falls into this category.
For user insights, I do not see an argument explicitly made. Implicitly, the 6.4 example seems too specialist (a narrow task, and a specific transition); 6.3's example provides no reason to believe that the counterfactuals identified are better than those of other methods.
For axiomatised intuitions, there is no explicit foundation. Instead, it seems to mix two internally consistent approaches - the discrete Shapley and the continuous IG. Technically, this is possible to do, but I see no attempt to motivate it beyond that.
Expositionally, I found the paper difficult, primarily as it frequently uses terms that have not been properly introduced or defined (e.g. 'interaction pots', 'endpoint priorities'). This strengthened a feeling that a lot of the text had been drafted by an LLM. While trying to find a definition of a 'Harsanyi pot' online, I came across what may be an earlier version of the paper. That version introduced a running example on the first page, which - I suspect - would help anchor intuitions more firmly in readers' minds.
Finally, computationally, the Shapley value is already expensive. Even with Proposition 3's symmetries, augmenting the player set with (a possibly large number of) grid points seems to make Aumann-SHAP prohibitively expensive in practice.
Questions for Authors
Flag for Ethics Review: No
Rating: 1
Confidence: 3