Group Shapley Effects Explained

• Group Shapley Effects are a generalization of the traditional Shapley value that fairly allocate contributions to pre-specified groups of variables, satisfying axioms like efficiency, symmetry, dummy, and additivity.
• They leverage computational methods such as Möbius inversion, Monte Carlo estimation, and spectral decomposition to efficiently manage the combinatorial complexity inherent in group-based analyses.
• Practical applications include sensitivity analysis, fairness attribution, and feature selection, where grouping leads to more interpretable, robust, and equitable model insights.

Group Shapley Effects generalize the classic Shapley value from single entities (“players,” features, or parameters) to pre-specified sets, assigning a unique, fair, and interpretable attribution of model output, variance, or payoff to groups of variables or parameters. This construct is fundamental for sensitivity analysis, explainable AI, fairness studies, simulation-based attribution, and cooperative game theory when the objects of interest are inherently grouped.

1. Definition and Axiomatic Foundation

Let $F = \{1,\ldots,p\}$ denote features, parameters, or players partitioned into disjoint groups $\mathcal{G} = \{G_1, \ldots, G_K\}$, and let $v: 2^F \to \mathbb{R}$ be a set function quantifying coalition “worth” (e.g., explained variance, expected prediction, model outcome difference). The Group Shapley value for group $G \subset F$ is

$$\phi_G(v) = \sum_{S \subseteq F \setminus G} \frac{|S|!\,(p - |S| - |G|)!}{(p - |G| + 1)!} \bigl[v(S \cup G) - v(S)\bigr]$$

This attribution uniquely satisfies (i) efficiency ($\sum_{j=1}^K\phi_{G_j}(v) = v(F) - v(\varnothing)$), (ii) symmetry, (iii) dummy group, and (iv) additivity properties, directly generalizing the Shapley axioms to collective entities. For $|G|=1$, this reduces to the original Shapley value (Wang et al., 6 Jan 2025, Kwon et al., 2024, Plischke et al., 2020). Specializations include the Shapley–Owen effect in variance-based sensitivity analysis (Plischke et al., 2020), and the “Union Shapley value” in cooperative games (Kępczyński et al., 27 May 2025).

2. Theoretical Variants and Relations

Several theoretical constructions extend or complement Group Shapley effects:

• Shapley–Owen Effects: For variance decompositions, the Shapley–Owen effect of a group $G$ is defined as

$$\Phi_G = \sum_{S \subseteq F \setminus G} \frac{|S|!\,(k - |S| - |G|)!}{k!|G|!}\,[v(S \cup G) - v(S)]$$

It uniquely decomposes total variance, captures both main effects and all joint interactions involving $G$, and is well-defined even under dependent inputs (Plischke et al., 2020, Ruess, 2024).

• Union and Synergy Values: The Union Shapley value $\mathcal{G} = \{G_1, \ldots, G_K\}$0 measures a group's overall impact as the effect of collectively removing $\mathcal{G} = \{G_1, \ldots, G_K\}$1 from the game, and satisfies a parallel set of efficiency, linearity, and symmetry axioms. Its dual, the Intersection Shapley ($\mathcal{G} = \{G_1, \ldots, G_K\}$2), complements this by focusing on group synergies (Kępczyński et al., 27 May 2025).
• Weighted Least Squares and Constrained Attributions: In counterfactual decomposition, the group Shapley value can be formulated as the solution to a weighted constrained least squares problem, guaranteeing uniqueness and efficiency even with incomplete coalition-worth information (Kwon et al., 2024).

3. Computational Algorithms

Efficient computation of Group Shapley effects is essential given their combinatorial complexity.

• Möbius Inversion and Dynamic Programming: For variance-based effects (Shapley–Owen), the core algorithm computes for each subset $\mathcal{G} = \{G_1, \ldots, G_K\}$3 the conditional variance $\mathcal{G} = \{G_1, \ldots, G_K\}$4 (or analogous value), then applies Möbius inversion:

$\mathcal{G} = \{G_1, \ldots, G_K\}$5

The group effect is then

$\mathcal{G} = \{G_1, \ldots, G_K\}$6

This achieves $\mathcal{G} = \{G_1, \ldots, G_K\}$7 model evaluations, rather than explicit summation over all permutations ($\mathcal{G} = \{G_1, \ldots, G_K\}$8) (Plischke et al., 2020).

• Monte Carlo and Quasi-Monte Carlo Estimation: Conditional variances or model outputs for each subset are approximated via “pick-n-freeze” methods or QMC designs, including dependency-respecting Rosenblatt transforms for non-independent inputs (Plischke et al., 2020, Iooss et al., 2017).
• Spectral Decomposition (PCE-based): For models that admit polynomial chaos expansions,

$\mathcal{G} = \{G_1, \ldots, G_K\}$9

group effects decompose as

$v: 2^F \to \mathbb{R}$0

separating model-independent weights (precomputed) and model-dependent terms (PCE coefficients), yielding scalable and error-controlled computation (Ruess, 2024).

• Sampling and Weighted Least Squares: For arbitrary set functions (e.g., counterfactual analysis in economics), coalition subsets are sampled according to kernel-weights, and group effects recovered by solving a constrained least squares system (Kwon et al., 2024).
• Shapley Aggregation via SHAP: Empirically, tree-based SHAP values can be aggregated across features to approximate group Shapley importance for scalable feature selection and group-wise penalization (Luo et al., 11 Feb 2026, Wang et al., 6 Jan 2025).

4. Statistical Properties and Robustness

Group Shapley attributions inherit the statistical properties of the underlying value function and aggregation scheme.

• Significance Testing: A robust procedure for testing the nullity of group effects employs a three-cumulant chi-square approximation, controlling empirical size and enhancing power under sparse or dense alternatives—even for $v: 2^F \to \mathbb{R}$1, skewed, or dependent data (Wang et al., 6 Jan 2025).
• Error Control and Convergence: For PCE-based approaches, rigorous error bounds are established: truncation error in variance propagates linearly to Shapley–Owen error, and both converge as more basis terms are included. Variance estimation error in Monte Carlo blocks also translates quantitatively into effect confidence intervals (Plischke et al., 2020, Ruess, 2024).
• Handling Missing Data: When some coalition-worths cannot be evaluated, linear programming yields upper/lower bounds for each group effect (SUB/SLB), and minimum-norm imputations restore efficiency and interpretability (Kwon et al., 2024).

5. Practical Applications and Empirical Results

Group Shapley effects have been applied to diverse domains:

• Sensitivity Analysis: For both models with independent and correlated inputs, group Shapley effects robustly attribute variance contributions even when classical Sobol’ indices fail, and clarify the share of uncertainty due to individual versus joint (correlated) feature groups (Iooss et al., 2017, Plischke et al., 2020).
• Fairness and Attributive Equity: In contexts demanding equitable attribution (e.g., algorithmic fairness, risk allocation), Shapley–Owen group effects are foundational. Empirical studies in financial risk prediction demonstrate that group aggregation leads to more equitable (lower Gini index), interpretable, and less collinear attributions compared to individual-level SHAP (Wang et al., 6 Jan 2025, Ruess, 2024).
• Feature Selection: Frameworks such as GRASP utilize group SHAP-attributed importance to drive structured sparsity, yielding compact, stable, and interpretable feature sets in clinical and tabular prediction tasks, outperforming classical LASSO and SHAP-thresholding in feature stability and redundancy (Luo et al., 11 Feb 2026).
• Counterfactual Structural Attribution: In economic simulation, group Shapley decomposition uniquely attributes the total effect of parameter interventions (e.g., changes in inequality) to meaningful blocks (e.g., returns, variances), with applications validated in Roy models (Kwon et al., 2024).

6. Connections, Limitations, and Extensions

• Relations to Other Group Values: The union, intersection, and merge Shapley values, as well as semivalue- and interaction-index-based generalizations, span a rich taxonomy. A key result is that the inclusion–exclusion principle links the union value to the intersection value, and that the sum of group effects recovers total value or variance (Kępczyński et al., 27 May 2025, Plischke et al., 2020).
• Computational Complexity: Despite algorithmic advances ($v: 2^F \to \mathbb{R}$2 regimes), high-dimensional settings (e.g., $v: 2^F \to \mathbb{R}$3) challenge both memory and sampling; sampling-based or hybrid approximations become mandatory. For arbitrary set functions, only partial coalition worths may be computationally feasible (Plischke et al., 2020, Kwon et al., 2024, Ruess, 2024).
• Synergy and Interaction Measurements: Group Shapley effects quantify both main and higher-order contributions, and the difference between the sum of individual and group effects (the “synergy” or “interaction index”) measures non-additive influences (Kępczyński et al., 27 May 2025, Plischke et al., 2020).
• Limitations: The effects presume a well-specified grouping, and non-uniqueness arises if groups are not disjoint. Under strong dependence or for ill-behaved models, variance estimation and block-wise sampling demand tailored approaches (e.g., Monte Carlo fallback) (Plischke et al., 2020).

7. Summary Table: Key Group Shapley Variants

Name	Formula	Notable Application
Shapley–Owen Effect	$v: 2^F \to \mathbb{R}$4	Variance-based SA (Plischke et al., 2020)
Union Shapley Value	$v: 2^F \to \mathbb{R}$5	Coalition impact (Kępczyński et al., 27 May 2025)
PCE Spectral Shapley–Owen	$v: 2^F \to \mathbb{R}$6	Fairness/SA (Ruess, 2024)
Group SHAP (empirical)	$v: 2^F \to \mathbb{R}$7	Feature selection (Luo et al., 11 Feb 2026)

All forms uniquely generalize Shapley attribution, providing an axiomatic, interpretable decomposition of group contributions in multivariate models and games, with rigorous computational and statistical formalism.