Abstract:Addressing the limitations of individual attribution scores via the Shapley value (SV), the field of explainable AI (XAI) has recently explored intricate interactions of features or data points. In particular, extensions of the SV, such as the Shapley Interaction Index (SII), have been proposed as a measure to still benefit from the axiomatic basis of the SV. However, similar to the SV, their exact computation remains computationally prohibitive. Hence, we propose with SVARM-IQ a sampling-based approach to efficiently approximate Shapley-based interaction indices of any order. SVARM-IQ can be applied to a broad class of interaction indices, including the SII, by leveraging a novel stratified representation. We provide non-asymptotic theoretical guarantees on its approximation quality and empirically demonstrate that SVARM-IQ achieves state-of-the-art estimation results in practical XAI scenarios on different model classes and application domains.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: In the field of Explainable Artificial Intelligence (XAI), the exact calculation of Shapley values (SV) and Shapley - based interaction indices (such as Shapley Interaction Index SII) becomes infeasible in practical applications due to their exponentially growing computational complexity. Therefore, the author proposes a new method named SVARM - IQ for efficiently approximating Shapley interaction indices of any order. ### Specific problem description: 1. **Limitations of Shapley values and their extensions**: - Shapley value (SV) is a concept introduced from cooperative game theory and is used for fair distribution of group benefits. It is widely used for feature attribution in XAI. - However, in the presence of strong feature correlations or high - order interactions, using only Shapley values is not sufficient to capture the reasoning process of a trained machine - learning model (Wright et al., 2016; Slack et al., 2020; Sundararajan and Najmi, 2020; Kumar et al., 2020, 2021). 2. **Computational complexity problem**: - The exact calculation of Shapley values and their extensions (such as Shapley Interaction Index SII) quickly becomes infeasible without further assumptions on the ML model because their complexity grows exponentially (Deng and Papadimitriou, 1994). 3. **Deficiencies of existing methods**: - Existing approximation methods (such as permutation - sampling - based methods, kernel - based methods, etc.) can solve the problem to a certain extent, but there is still room for improvement in terms of efficiency and accuracy. ### Proposed solution: To address the above problems, the author proposes SVARM - IQ, a new method based on stratified sampling that can efficiently approximate Shapley interaction indices of any order. Specific contributions include: 1. **Model - independent approximation algorithm**: - SVARM - IQ estimates Shapley - based interaction scores of any order through a stratified representation method and is applicable to a wide range of interaction index categories. 2. **Theoretical analysis**: - It is proved that under mild assumptions, SVARM - IQ is unbiased and provides bounds on the approximation error. 3. **Empirical evaluation**: - An open - source implementation is provided, and experiments prove the superior performance of SVARM - IQ in different model classes and application scenarios. ### Core ideas of the method: - **Stratified representation method**: - Decompose the interaction index \( I_K \) into multiple stratified terms \( I_W^{K,\ell} \) and efficiently estimate these terms through stratified sampling. - **Update mechanism**: - Sample a coalition \( A\subseteq N \) each time and update the stratified estimate \( \hat{I}_W^{K,\ell} \) of each subset \( K \) according to \( A \). - **Budget optimization**: - By pre - calculating the values of small - scale and large - scale coalitions, reduce repeated calculations, thereby using limited computational resources more efficiently. ### Conclusion: This paper solves the computational complexity problem of Shapley values and their extensions in practical applications by proposing SVARM - IQ, providing an efficient and accurate approximation method and significantly enhancing the interpretability in the XAI field.

SVARM-IQ: Efficient Approximation of Any-order Shapley Interactions through Stratification

SHAP-IQ: Unified Approximation of any-order Shapley Interactions

KernelSHAP-IQ: Weighted Least-Square Optimization for Shapley Interactions

Approximating the Shapley Value without Marginal Contributions

shapiq: Shapley Interactions for Machine Learning

Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles

Faith-Shap: The Faithful Shapley Interaction Index

Evaluate the Contribution of Multiple Participants in Federated Learning

Efficient Sampling Approaches to Shapley Value Approximation

Improving the Sampling Strategy in KernelSHAP

Explainable AI: Efficiency Sequential Shapley Updating Approach

A $k$-additive Choquet integral-based approach to approximate the SHAP values for local interpretability in machine learning

A Comprehensive Study of Shapley Value in Data Analytics

Calculation of exact Shapley values for explaining support vector machine models using the radial basis function kernel

Fast Approximation of the Shapley Values Based on Order-of-Addition Experimental Designs

A Quantum Algorithm for Shapley Value Estimation

Computing the Shapley Value in Allocation Problems: Approximations and Bounds, with an Application to the Italian VQR Research Assessment Program

Rational Shapley Values

Approximating the Shapley Value Using Stratified Empirical Bernstein Sampling

Quantifying Feature Importance of Games and Strategies via Shapley Values

From SHAP Scores to Feature Importance Scores