Shapley Interactions: Theory & Applications

• Shapley Interactions is a framework that generalizes Shapley values to measure higher-order feature interactions and decompose model outputs into synergistic, redundant, and independent components.
• It builds on game-theoretic foundations using discrete derivatives to uniquely attribute joint contributions, making it applicable to models such as deep nets, tree ensembles, and graph neural networks.
• Efficient computational methods like Monte Carlo sampling, kernel regression, and structure-aware algorithms enable practical estimation and visualization of interactions in real-world systems.

Shapley Interactions (SI) quantify how groups of features—rather than individual features alone—jointly contribute to the outcome of a machine learning model, in an axiomatic, game-theoretic fashion. Formally, SI generalizes the classical Shapley value to higher-order interactions (pairs, triplets, etc.), providing rigorous and unique decompositions of model outputs into synergistic, redundant, and independent components associated with arbitrary feature subsets. Modern SI theory and computational tools enable practical estimation and interpretation for deep neural networks, tree ensembles, survival models, graph neural networks, hyperparameter optimization, and multimodal systems.

1. Game-Theoretic Foundations and Formal Definitions

The SI framework arises from cooperative game theory, where the value function $v : 2^N \to \mathbb{R}$ assigns a “worth” to all possible coalitions (subsets) of a set $N$ of players (features) (Muschalik et al., 2024). The classical Shapley value for feature $i$ is: $$\phi_v(i \mid N) = \sum_{S \subseteq N \setminus \{i\}} \frac{(n-|S|-1)!~|S|!}{n!} [v(S \cup \{i\}) - v(S)]$$

Shapley Interactions extend this to groups $I \subseteq N$, using the discrete derivative: $$\Delta_S^I v = \sum_{U \subseteq I} (-1)^{|I| - |U|} v(S \cup U),$$ yielding the general Shapley interaction index for coalition $I$ as: $$\phi_I(v) = \sum_{S \subseteq N \setminus I} \frac{|S|!\,(n-|S|-|I|)!}{(n-|I|+1)!} \Delta_S^I v$$ (Muschalik et al., 2024, Fumagalli et al., 2024, Dhamdhere et al., 2019). For $|I|=1$, this recovers the standard Shapley value; for $|I|>1$, it quantifies the unique k-way contribution of $I$ beyond all proper subsets.

Among interaction indices, well-known formalizations include the Shapley interaction index (SII) [Grabisch & Roubens], the Shapley–Taylor index (Dhamdhere et al., 2019), and the Faithful Shapley Interaction index (Tsai et al., 2022, Kim et al., 5 Jan 2026). Each satisfies extensions of the classic Shapley axioms—efficiency, symmetry, dummy, linearity—and, depending on the formulation, additional uniqueness or faithfulness constraints.

2. Algorithms and Efficient Approximation Schemes

Exact SI computation is combinatorially expensive—requiring enumeration over $2^{n-|I|}$ coalitions per interaction $I$ (Muschalik et al., 2024, Fumagalli et al., 2023). Major advances in scalable SI algorithms include:

• Monte Carlo and Permutation Sampling: Classic SI estimation via random permutations or random coalitional subsets, with statistical concentration guarantees (Kolpaczki et al., 2024, Muschalik et al., 2024). Sampling variance can be substantial, especially for high-order interactions.
• KernelSHAP-IQ: SI is reframed as a weighted least-squares (WLS) regression estimation (generalizing KernelSHAP for feature attributions), exploiting the closed-form correspondence between SII and WLS optima (Fumagalli et al., 2024). For pairwise interactions, this reduces variance and can yield state-of-the-art accuracy.
• Stratified Sampling (SVARM-IQ): Samples are efficiently shared across all possible interacting feature sets, stratified by coalition size and overlap, leading to orders-of-magnitude improvements in mean squared error (MSE) for multiway SI estimation (Kolpaczki et al., 2024).
• SHAP-IQ Unified Representation: Any cardinal interaction index (CIIs including SII, STI, FSI) can be estimated from randomized shell sampling and a unified weighting scheme, maintaining unbiasedness and efficiency (Fumagalli et al., 2023).
• TreeSHAP-IQ and GraphSHAP-IQ: For tree-based and message-passing models, structure can be exploited to compute all SIs up to order k with polynomial, often linear, complexity in the number of features or nodes, independent of the graph or tree size (Muschalik et al., 2024, Muschalik et al., 28 Jan 2025).
• Tensor-Train Methods for FSI (TT-FSI): Faithful Shapley indices are computed via a matrix product operator (MPO)/tensor-train framework, reducing memory from $O(4^d)$ to $O(\ell d^2)$ and time to $O(\ell^2 d^3 2^d)$, scaling to $d=20$ with up to 280× speedup over prior algorithms (Kim et al., 5 Jan 2026).

These methods enable practical SI estimation for real-world applications, with strong guarantees of unbiasedness, efficiency preservation, and variance reduction.

3. Model-Specific Adaptations and Applications

SI methods have been tailored for a range of model classes:

• Deep Neural Networks (DNNs): Multivariate SI analysis reveals prototype feature coalitions learned by DNNs. The block-interaction metric:

$$B([A]) = \phi([A]\mid N_{[A]}) - \sum_{i\in A} \phi(i\mid N_i)$$

together with the “significance” score $T([A]) = B_{\max}([A]) - B_{\min}([A])$, summarizes arbitrary higher-order effects (Zhang et al., 2020). A three-level Monte Carlo algorithm discovers non-adjacent coalitions (e.g., linguistic constituents in BERT).

• Multimodal Models: The MultiSHAP framework attributes prediction to patch-token (image-text) pairs using SI, producing high-resolution, instance- and dataset-level heatmaps for synergy and suppression (Wang et al., 1 Aug 2025).
• Tree Ensembles: TreeSHAP-IQ uses recursive polynomial arithmetic to compute all SIs up to any order in a single tree traversal, facilitating exact, efficient explanation of gradient-boosted and random forests (Muschalik et al., 2024).
• Graph Neural Networks: By leveraging local message-passing and pooling structure, GraphSHAP-IQ restricts SI computation to receptive fields, reducing exponential to near-linear scaling, with empirical validation on chemical and infrastructure graphs (Muschalik et al., 28 Jan 2025).
• Survival and Time-to-Event Models: In survival analysis, standard additive SI fails due to the non-additivity of hazard or survival scales. SurvSHAP-IQ extends SI to time-indexed functional decompositions, quantifying genuinely time-dependent and time-independent effects (Langbein et al., 18 Feb 2026).
• Hyperparameter Optimization (HPO): HyperSHAP attributes predictive performance (locally or globally) to main and pairwise/interacting hyperparameters, empirically demonstrating that most performance is explained by lower-order SIs (Wever et al., 3 Feb 2025).

4. Axiomatic Properties, Variants, and Theoretical Insights

Shapley Interactions admit several mathematically rigorous variants, each characterized by axiomatic choices:

• SII (Grabisch-Roubens): The classical SI, satisfying efficiency, symmetry, linearity, and (recursive) dummy axioms (Muschalik et al., 2024). Directly yields Möbius decomposition, and underlies most practical SI computation.
• Shapley–Taylor Index (STI): Adds an “interaction distribution” axiom, connecting SI to the truncated Taylor expansion of the set function’s multilinear extension, distributing pure k-way interactions equally across all size-k coalitions (Dhamdhere et al., 2019).
• Faithful Shapley Index (FSI, Faith-Shap): The unique solution to a constrained weighted least-squares fit of a polynomial of order $\ell$ to the value function, ensuring that for $\ell$-additive games, the SI surrogate matches the original game exactly (Tsai et al., 2022, Kim et al., 5 Jan 2026).
• n-Shapley Values: A parametric family ($1 \leq n \leq d$) interpolating between first-order SHAP and the full Möbius expansion, with each order n exactly recovering the corresponding GAM decomposition (Bordt et al., 2022).

Distinct indices agree on unique recovery of order-$n$ generalized additive models (GAMs). For higher-order or non-additive functions, index selection determines how interactions are distributed among features and groups (Bordt et al., 2022).

5. Empirical Capabilities and Visualization

The practical utility of SI is now demonstrated in multiple domains:

• Interpretability: SIs expose hidden model logic (e.g., BERT’s grouping of syntactic constituents (Zhang et al., 2020); synergy/suppression patterns in VQA (Wang et al., 1 Aug 2025); structured pairing in speech (Singhvi et al., 2024)).
• Uncertainty and Error Modes: Negative SIs (suppression) pinpoint critical mislocalization or failed compositional reasoning in vision, language, and clinical diagnosis (Wang et al., 1 Aug 2025).
• Data Valuation and Feature Selection: SIs enable fair distribution of value to groups of data points, features, or ensemble members, informing pruning and robust selection (Muschalik et al., 2024).
• Hypergraph and Network Visualization: SIs up to order 3 are often visualized as networks or hypergraphs, with edges or hyperedges corresponding to the magnitude and sign of the SI (Muschalik et al., 2024, Muschalik et al., 28 Jan 2025).
• Survival Models: Time-resolved SIs reveal which variables or covariate groups have genuinely dynamic effects versus static or additive influence (Langbein et al., 18 Feb 2026).

Empirically, across domains and model classes, most explanatory power is often captured by SIs up to order 2 or 3 (as measured by explained variance $R^2$), with higher orders rarely contributing substantially (Wever et al., 3 Feb 2025, Muschalik et al., 2024).

6. Limitations, Future Directions, and Open Problems

While Shapley Interaction methodology has achieved broad impact, several key limitations are recognized:

• Computational Complexity: Exact SI calculation scales as $O(2^n)$, which is infeasible for $n \gtrsim 20$. Efficient approximations via stratification, kernel regression, or structure exploitation are still sensitive to sample budget, especially for large $k$ or $n$ (Kolpaczki et al., 2024, Fumagalli et al., 2023, Kim et al., 5 Jan 2026).
• Extending to Arbitrary Structures: Most current methods assume flat or sequential inputs. Extending SI computation to arbitrary graph/node structures, high-dimensional time series, or non-Euclidean manifolds remains an open frontier (Zhang et al., 2020, Muschalik et al., 28 Jan 2025).
• Index Selection and Interpretation: For higher-order interactions ($k \geq 4$), both the computational burden and the cognitive load for users increase. Selecting or visualizing meaningful higher-order SIs requires further methodological innovation (Muschalik et al., 2024).
• Model Faithfulness vs. Data Faithfulness: Different SI variants optimize for distinct properties (exact recovery, efficiency, distribution, faithfulness). The choice of index can materially affect explanations, especially under feature dependence or in observational estimation (Tsai et al., 2022, Bordt et al., 2022).
• Human Interpretation Barriers: As the number and order of displayed interactions increase, information overload and interpretability become practical concerns (Muschalik et al., 2024).
• Application to Non-Standard Prediction Tasks: Survival models, structured prediction, and multi-modal fusion present challenges for SI estimation and aggregation across processes, targets, or modalities (Langbein et al., 18 Feb 2026, Wang et al., 1 Aug 2025).

Active research directions include improved sampling schemes (variance reduction, importance sampling), integration of structure-aware imputation, index-guided interpretability visualization, extension to graph-structured data, and human-centered studies of SI presentation.

---

References:

• Interpreting Multivariate Shapley Interactions in DNNs (Zhang et al., 2020)
• MultiSHAP: A Shapley-Based Framework for Explaining Cross-Modal Interactions in Multimodal AI Models (Wang et al., 1 Aug 2025)
• shapiq: Shapley Interactions for Machine Learning (Muschalik et al., 2024)
• SVARM-IQ: Efficient Approximation of Any-order Shapley Interactions through Stratification (Kolpaczki et al., 2024)
• SHAP-IQ: Unified Approximation of any-order Shapley Interactions (Fumagalli et al., 2023)
• The Shapley Taylor Interaction Index (Dhamdhere et al., 2019)
• TT-FSI: Scalable Faithful Shapley Interactions via Tensor-Train (Kim et al., 5 Jan 2026)
• Faith-Shap: The Faithful Shapley Interaction Index (Tsai et al., 2022)
• KernelSHAP-IQ: Weighted Least-Square Optimization for Shapley Interactions (Fumagalli et al., 2024)
• Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles (Muschalik et al., 2024)
• Exact Computation of Any-Order Shapley Interactions for Graph Neural Networks (Muschalik et al., 28 Jan 2025)
• HyperSHAP: Shapley Values and Interactions for Hyperparameter Importance (Wever et al., 3 Feb 2025)
• Functional Decomposition and Shapley Interactions for Interpreting Survival Models (Langbein et al., 18 Feb 2026)
• From Shapley Values to Generalized Additive Models and back (Bordt et al., 2022)
• Using Shapley interactions to understand how models use structure (Singhvi et al., 2024)