pith. sign in

arxiv: 2605.22738 · v1 · pith:YDSIFA3Mnew · submitted 2026-05-21 · 💻 cs.LG · cs.AI· stat.ML

Proxy-Based Approximation of Shapley and Banzhaf Interactions

Santo M. A. R. Thies , Hubert Baniecki , R. Teal Witter , Eyke H\"ullermeier , Maximilian Muschalik , Fabian Fumagalli This is my paper

Pith reviewed 2026-05-22 07:12 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.ML
keywords Shapley interactionsBanzhaf interactionsproxy modelsTreeSHAPfeature interactionsmodel interpretabilityapproximation algorithmsresidual correction
0
0 comments X

The pith

ProxySHAP approximates Shapley and Banzhaf interactions more accurately by using tree-based proxies plus residual correction.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that combining sample-efficient tree proxy models with a residual adjustment step via Maximum Sample Reuse yields consistent estimators for higher-order feature interactions. A sympathetic reader would care because current methods force a sharp trade-off between computational cost and accuracy when models have thousands of features. If the central claim holds, practitioners can compute reliable interaction values in both low- and high-budget settings without the variance exploding as interaction order grows. The work also supplies an exact polynomial-time algorithm for interaction indices on tree ensembles, removing an exponential dependence on tree depth that limited earlier exact methods.

Core claim

ProxySHAP reconciles the high sample efficiency of tree-based proxy models with a principled path to consistency via residual correction. On a theoretical level, it derives a polynomial-time generalization of interventional TreeSHAP to compute exact interaction indices for tree ensembles, successfully bypassing exponential tree-depth dependencies in prior methods. The residual adjustment strategy is shown to correct proxy bias under conditions where Maximum Sample Reuse keeps variance from scaling exponentially with interaction size.

What carries the argument

Residual adjustment via Maximum Sample Reuse applied to tree-based proxy models, which corrects bias while controlling variance growth and enables a polynomial-time exact TreeSHAP generalization for interactions.

If this is right

  • ProxySHAP records the lowest approximation error among tested estimators in both small- and large-budget regimes.
  • The method scales to applications with thousands of features while still outperforming ProxySPEX and KernelSHAP-IQ.
  • Downstream explainability tasks such as interaction-based feature selection improve when using the new estimates.
  • Exact interaction indices for tree ensembles become computable in polynomial time rather than exponential time in tree depth.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same proxy-plus-residual pattern could be tested on non-tree models such as neural networks by substituting appropriate fast proxies.
  • If the variance-control property generalizes, interaction-based fairness audits become feasible for high-dimensional tabular data.
  • The polynomial-time TreeSHAP generalization might be adapted to compute other cooperative-game values beyond Shapley and Banzhaf.

Load-bearing premise

The residual adjustment strategy corrects proxy bias without its variance scaling exponentially with interaction size under the specific conditions that hold for the models and datasets evaluated in the paper.

What would settle it

An experiment on one of the paper's large-feature datasets in which the empirical variance of the residual-corrected estimator grows exponentially with interaction order instead of remaining controlled.

Figures

Figures reproduced from arXiv: 2605.22738 by Eyke H\"ullermeier, Fabian Fumagalli, Hubert Baniecki, Maximilian Muschalik, R. Teal Witter, Santo M. A. R. Thies.

Figure 1
Figure 1. Figure 1: Left: A ProxySHAP explanation of the SigLIP-2 model using only 2048 model calls. Right: In Phase 1, we fit a regression proxy model using sampled binary coalitions and game values. In Phase 2, we extract proxy interactions and, when appropriate, adjust them using residual estimates. sum of its marginal contributions across all possible subsets: ϕ p i (ν) := X T ⊆N\{i} pt(n)∆iν(T), (1) where ∆iν(T) := ν(T ∪… view at source ↗
Figure 2
Figure 2. Figure 2: Runtime improvement of extracting interactions using our Algo￾rithm 2 over Fourier-based extraction. Per-dataset speedups and the effect of tree depth on approximation quality are shown in Figures 13 and 14. Proposition 3.2 shows that the interactions of the tree proxy can be computed exactly by aggregating leaf-wise contributions. In particular, for a fixed interaction S, ex￾traction requires only a singl… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of ProxySHAP with and without MSR adjustment, measured by the MSE ratio. While MSR improves Shapley value approximation, it can degrade higher-order interaction estimates, as its variance scales as n k−1/|T | for interactions of order k (Theorem 3.3). This motivates the adjusted ProxySHAP estimator ϕˆProxySHAP S (ν; T ) = ϕ p S (ˆνT ) + ϕˆMSR S (ν − νˆT ; T ). While MSR often improves singleton … view at source ↗
Figure 4
Figure 4. Figure 4: Approximation quality (Relative MSE) for Shapley interactions of ProxySHAP across different configurations and state-of-the-art baselines. Additional results for Shapley and Banzhaf interactions on all 47 datasets can be found in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative MSE for pairwise Shapley interac￾tion approximation of ProxySHAP with HPO (top) and for large n (bottom). Further results in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Area between the insertion/deletion curves (AID) for explaining two CLIP ViT vari￾ants on the MS COCO dataset with ProxySHAP, ProxySPEX, and the FIxLIP baseline. Setup. Similarly to Baniecki et al. [2], we ana￾lyze the CLIP model in two vision transformer variants: ViT-32 and ViT-16. We explain a sample of 200 image–text pairs from the MS COCO dataset [33], which contain around 10– 30 text tokens per input… view at source ↗
Figure 7
Figure 7. Figure 7: Empirical variance scaling with sampling budget |T| for interaction orders |S| = k. For each order k, the plot shows the mean, minimum, and maximum empirical variance over all subsets S of size k. The black curve denotes the theoretical bound shape, namely proportional to ∥v∥ 2 ∞ log(n)/|T| for k = 1 and to ∥v∥ 2 ∞n k−1/|T| for k > 1. Since big-O bounds are defined only up to a multiplicative constant, the… view at source ↗
Figure 8
Figure 8. Figure 8: Faithfulness R2 for explaining CLIP (ViT-16) on the MS COCO dataset with ProxySHAP, ProxySPEX, and the FIxLIP baseline. D.2 XGBoost Default for Large Player Counts [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ablation on approximating the cross-modal FIxLIP estimator. Faithfulness R2 for explaining two CLIP variants on MS COCO with ProxySHAP and the FIxLIP baseline. Motivated by this observation, we evaluate ProxySHAP (XGBoost+HPO-Informed) as an alternative default proxy for large-scale games. The results show that this configuration improves approximation quality in low-budget regimes and for games with many … view at source ↗
Figure 10
Figure 10. Figure 10: Approximation quality of two different XGBoost defaults. We show that using 2000 trees with a maximum depth of 3 improves estimation quality in low- to medium-budget regimes. low-budget regimes and for games with many players, where the standard XGBoost default may be insufficient to capture the relevant interaction structure. D.3 Runtime We evaluate runtime by translating model evaluations into wall-cloc… view at source ↗
Figure 11
Figure 11. Figure 11: Approximation quality as a function of runtime for second- and third-order interaction estimation across different per-evaluation cost regimes. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ablations of sampling weights and residual approximators for ProxySHAP. We further investigate the effect of the resid￾ual approximator and sampling weights used in the adjustment step. Specifically, we compare SHAP-IQ [15] and KernelSHAP-IQ [16] as model-agnostic residual approxima￾tors. We also compare leverage weights, as used in LeverageSHAP [45], with KernelSHAP￾IQ weights [16]. As underlying games, … view at source ↗
Figure 13
Figure 13. Figure 13: Approximation quality (Relative MSE) of ProxySHAP and ProxySPEX using different maximum tree depth options across small, medium, and large player domains. Our method relies on the ability to efficiently extract exact cardinal-probabilistic interaction indices from the underlying tree-based model. We extend interventional TreeSHAP by Zern et al. [65] to extract the exact cardinal-probabilistic interaction … view at source ↗
Figure 14
Figure 14. Figure 14: Speedup of interventional extraction compared to Fourier extraction for extracting all interactions of order 1, 2, and 3 across different datasets. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Predicted versus ground-truth normalized interaction values for different approximation methods and sampling budgets. Each point represents one interaction value from one dataset and one benchmark run; points closer to the diagonal indicate better agreement with the exact interaction values. Columns compare ProxySHAP, ProxySPEX, SHAPIQ, PermutationSamplingSII, and KernelSHAPIQ, while rows correspond to in… view at source ↗
Figure 16
Figure 16. Figure 16: ProxySHAP with disjoint coalition sets for proxy fitting and residual adjustment. drawn as the multiplicative interval [¯r/s, r¯ · s], corresponding to one standard deviation in log-space. Hence, values below one indicate that adjustment improves approximation quality, whereas values above one indicate that it deteriorates. For interaction indices, however, the effect of adjustment is more nuanced. While … view at source ↗
Figure 17
Figure 17. Figure 17: Selection of representative approximation curves for SII and BII at second- and third-order interactions. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Winnermap comparing the best performing method for each dataset and budget for SII orders 2 and 3. Note that the HPO-Informed variants are considered only for datasets with more than 1000 features in this overview. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Winnermap comparing the best performing method for each dataset and budget for BII orders 2 and 3. Note that the HPO-Informed variants are considered only for datasets with more than 1000 features in this overview. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_19.png] view at source ↗
read the original abstract

Shapley and Banzhaf interactions capture the complex dynamics inherent in modern machine learning applications. However, current estimators for these higher-order interactions trade off between speed and accuracy. To overcome this limitation, we introduce ProxySHAP. ProxySHAP reconciles the high sample efficiency of tree-based proxy models with a principled path to consistency via residual correction. On a theoretical level, we derive a polynomial-time generalization of interventional TreeSHAP to compute exact interaction indices for tree ensembles, successfully bypassing exponential tree-depth dependencies in prior methods. Furthermore, we formally analyze the residual adjustment strategy, characterizing the specific conditions under which Maximum Sample Reuse (MSR) corrects proxy bias without its variance scaling exponentially with interaction size. Extensive benchmarking demonstrates that ProxySHAP sets a new state-of-the-art standard for approximation quality, including in large-scale applications with thousands of features. By achieving the lowest error in both small- and large-budget regimes, ProxySHAP significantly outperforms the prior best estimators ProxySPEX and KernelSHAP-IQ, while also delivering superior performance on downstream explainability tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper introduces ProxySHAP for approximating Shapley and Banzhaf interactions. It combines tree-based proxy models for sample efficiency with a residual correction strategy (Maximum Sample Reuse, MSR) claimed to ensure consistency. A key theoretical contribution is a polynomial-time generalization of interventional TreeSHAP that computes exact interaction indices for tree ensembles while avoiding exponential dependence on tree depth. The authors formally analyze the residual adjustment, characterizing conditions under which MSR corrects proxy bias without variance scaling exponentially in interaction order. Extensive benchmarks on small- and large-budget regimes, including datasets with thousands of features, show ProxySHAP achieving lower error than ProxySPEX and KernelSHAP-IQ and better performance on downstream explainability tasks.

Significance. If the formal characterization of MSR conditions holds and the empirical superiority is robust across diverse models and high-dimensional regimes, ProxySHAP would advance scalable higher-order interaction estimation, which is relevant for interpretability in modern ML. The polynomial-time TreeSHAP generalization for exact indices on ensembles is a clear technical strength that could be adopted independently.

major comments (1)
  1. [Theoretical analysis of residual adjustment / MSR] The formal analysis of the residual adjustment strategy (abstract and corresponding theoretical section) is load-bearing for the central SOTA claim in large-scale settings. The characterization of conditions under which MSR corrects proxy bias without variance scaling exponentially with interaction size must be stated explicitly, including any assumptions on proxy fidelity, feature correlations, or bounded higher-order effects. Without these details or a concrete verification that the conditions hold for models with thousands of features, the superiority over ProxySPEX and KernelSHAP-IQ in both budget regimes cannot be fully assessed.
minor comments (1)
  1. [Abstract and § on TreeSHAP generalization] The abstract claims 'polynomial-time' and 'exact' indices; ensure the complexity statement and any remaining exponential factors (e.g., in interaction order k) are clarified with precise big-O notation in the main text.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript. We address the major comment below and have revised the paper to strengthen the explicit presentation of our theoretical results on the residual adjustment strategy.

read point-by-point responses

  1. Referee: The formal analysis of the residual adjustment strategy (abstract and corresponding theoretical section) is load-bearing for the central SOTA claim in large-scale settings. The characterization of conditions under which MSR corrects proxy bias without variance scaling exponentially with interaction size must be stated explicitly, including any assumptions on proxy fidelity, feature correlations, or bounded higher-order effects. Without these details or a concrete verification that the conditions hold for models with thousands of features, the superiority over ProxySPEX and KernelSHAP-IQ in both budget regimes cannot be fully assessed.

    Authors: We thank the referee for highlighting the need for greater explicitness in our theoretical characterization, which is indeed central to supporting the large-scale claims. Section 4 of the manuscript already derives the conditions under which MSR achieves bias correction without exponential variance growth in interaction order, but we agree that a more enumerated presentation will improve clarity and allow better assessment of the SOTA results. In the revision, we will add a dedicated paragraph and a formal theorem statement that explicitly lists the assumptions: (1) proxy fidelity, requiring the tree proxy to approximate the target function with L2 error bounded by a small constant (empirically verified via validation MSE in our training procedure); (2) bounded higher-order effects, with the remainder term controlled by a factor independent of order k; and (3) feature correlations handled through the standard interventional distribution used in TreeSHAP computations, without additional restrictions. We will also include a brief verification discussion referencing our large-scale experiments (Section 5.3) on datasets with thousands of features, where observed error rates and lack of variance explosion align with the derived bounds, confirming the conditions hold for the evaluated models. These changes directly address the request and will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

Derivation chain is self-contained with independent theoretical contributions.

full rationale

The paper introduces ProxySHAP by deriving a polynomial-time generalization of interventional TreeSHAP for exact interaction indices on tree ensembles and by formally characterizing conditions under which MSR residual correction removes proxy bias without exponential variance scaling in interaction order. These steps are presented as new theoretical results. Benchmarking comparisons to ProxySPEX and KernelSHAP-IQ are empirical and separate from the derivations. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citation chains appear in the abstract or described claims.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review is based solely on the abstract; full details on parameters, assumptions, and any invented constructs are unavailable. The ledger therefore reflects only high-level elements inferable from the summary.

free parameters (1)
  • Proxy model hyperparameters
    Choice and fitting of the tree-based proxy model is central to the approximation but not specified in the abstract.
axioms (1)
  • domain assumption Tree-based models can serve as sufficiently accurate proxies for the target black-box model
    The entire proxy-plus-correction strategy rests on this modeling choice being effective in practice.

pith-pipeline@v0.9.0 · 5748 in / 1281 out tokens · 46859 ms · 2026-05-22T07:12:05.101573+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

76 extracted references · 76 canonical work pages · 4 internal anchors

  1. [1]

    Efficient and Accurate Explanation Estimation with Distribution Compression

    Hubert Baniecki, Giuseppe Casalicchio, Bernd Bischl, and Przemyslaw Biecek. Efficient and Accurate Explanation Estimation with Distribution Compression. InProceedings of the International Conference on Learning Representations (ICLR), 2025

    work page 2025

  2. [2]

    Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions

    Hubert Baniecki, Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer, Eyke Hüller- meier, and Przemyslaw Biecek. Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    work page 2025

  3. [3]

    Weighted voting doesn’t work: A mathematical analysis.Rutgers Law Review, 19:317, 1964

    John F Banzhaf III. Weighted voting doesn’t work: A mathematical analysis.Rutgers Law Review, 19:317, 1964

    work page 1964

  4. [4]

    Proxy-SPEX: Sample-efficient interpretability via sparse feature interactions in LLMs

    Landon Butler, Abhineet Agarwal, Justin Singh Kang, Yigit Efe Erginbas, Bin Yu, and Kannan Ramchandran. Proxy-SPEX: Sample-efficient interpretability via sparse feature interactions in LLMs. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    work page 2025

  5. [5]

    Yu-Xuan Cai, Hai-Yan Chen, Ya-Jing Qu, Wen-Hao Zhao, Mei-Ying Wang, Ying Chen, and Jin Ma. Improved vertical distribution prediction of soil vocs contamination in site-scale utilizing ensemble machine learning approach integrated with molecular descriptors.Journal of Hazardous Materials, page 139452, 2025

    work page 2025

  6. [6]

    Lozano, A

    Javier Castro, Daniel Gómez, and Juan Tejada. Polynomial calculation of the Shapley value based on sampling.Computers & Operations Research, 36(5):1726–1730, 2009. doi: 10.1016/j. cor.2008.04.004

    work page doi:10.1016/j 2009

  7. [7]

    Improving polynomial estima- tion of the Shapley value by stratified random sampling with optimum allocation.Computers & Operations Research, 82:180–188, 2017

    Javier Castro, Daniel Gómez, Elisenda Molina, and Juan Tejada. Improving polynomial estima- tion of the Shapley value by stratified random sampling with optimum allocation.Computers & Operations Research, 82:180–188, 2017. doi: 10.1016/j.cor.2017.01.019

    work page doi:10.1016/j.cor.2017.01.019 2017

  8. [8]

    Xgboost: A scalable tree boosting system,

    Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 785–794. ACM, 2016. doi: 10.1145/2939672.2939785

    work page doi:10.1145/2939672.2939785 2016

  9. [9]

    Improving KernelSHAP: Practical Shapley Value Estimation Using Linear Regression

    Ian Covert and Su-In Lee. Improving KernelSHAP: Practical Shapley Value Estimation Using Linear Regression. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 3457–3465, 2021

    work page 2021

  10. [10]

    Stochastic amortization: A unified approach to accelerate feature and data attribution

    Ian Connick Covert, Chanwoo Kim, Su-In Lee, James Zou, and Tatsunori Hashimoto. Stochastic amortization: A unified approach to accelerate feature and data attribution. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024

    work page 2024

  11. [11]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InProceedings of the International Conference on Learning Representation...

    work page 2021

  12. [12]

    InstaSHAP: Interpretable additive models explain shapley values instantly

    James Enouen and Yan Liu. InstaSHAP: Interpretable additive models explain shapley values instantly. InProceedings of the International Conference on Learning Representations (ICLR), 2025

    work page 2025

  13. [13]

    Tabarena: A living benchmark for machine learning on tabular data

    Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzmüller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. Tabarena: A living benchmark for machine learning on tabular data. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), 2026

    work page 2026

  14. [14]

    Axiomatic characterizations of probabilistic and cardinal-probabilistic interaction indices.Games and Economic Behavior, 55(1):72–99, 2006

    Katsushige Fujimoto, Ivan Kojadinovic, and Jean-Luc Marichal. Axiomatic characterizations of probabilistic and cardinal-probabilistic interaction indices.Games and Economic Behavior, 55(1):72–99, 2006. doi: 10.1016/j.geb.2005.03.002

    work page doi:10.1016/j.geb.2005.03.002 2006

  15. [15]

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

    Fabian Fumagalli, Maximilian Muschalik, Patrick Kolpaczki, Eyke Hüllermeier, and Barbara Hammer. SHAP-IQ: Unified Approximation of any-order Shapley Interactions. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 11515–11551, 2023

    work page 2023

  16. [16]

    KernelSHAP-IQ: Weighted Least Square Optimization for Shapley Interactions

    Fabian Fumagalli, Maximilian Muschalik, Patrick Kolpaczki, Eyke Hüllermeier, and Barbara Hammer. KernelSHAP-IQ: Weighted Least Square Optimization for Shapley Interactions. In Proceedings of the International Conference on Machine Learning (ICML), pages 14308–14342, 2024

    work page 2024

  17. [17]

    Unifying Feature-Based Explanations with Functional ANOV A and Cooperative Game Theory

    Fabian Fumagalli, Maximilian Muschalik, Eyke Hüllermeier, Barbara Hammer, and Julia Herbinger. Unifying Feature-Based Explanations with Functional ANOV A and Cooperative Game Theory. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 5140–5148, 2025

    work page 2025

  18. [18]

    SHAP values via sparse fourier repre- sentation

    Ali Gorji, Andisheh Amrollahi, and Andreas Krause. SHAP values via sparse fourier repre- sentation. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), 2025

    work page 2025

  19. [19]

    An axiomatic approach to the concept of interaction among players in cooperative games.International Journal of Game Theory, 28(4):547–565,

    Michel Grabisch and Marc Roubens. An axiomatic approach to the concept of interaction among players in cooperative games.International Journal of Game Theory, 28(4):547–565,

    work page

  20. [20]

    doi: 10.1007/s001820050125

    work page doi:10.1007/s001820050125

  21. [21]

    TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

    Léo Grinsztajn, Klemens Flöge, Oscar Key, Felix Birkel, Philipp Jund, Brendan Roof, Benjamin Jäger, Dominik Safaric, Simone Alessi, Adrian Hayler, Mihir Manium, Rosen Yu, Felix Jablon- ski, Shi Bin Hoo, Anurag Garg, Jake Robertson, Magnus Bühler, Vladyslav Moroshan, Lennart Purucker, Clara Cornu, Lilly Charlotte Wehrhahn, Alessandro Bonetto, Bernhard Schö...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2025

  22. [22]

    Naofumi Hama, Masayoshi Mase, and Art B. Owen. Deletion and insertion tests in regression models.Journal of Machine Learning Research, 24:290:1–290:38, 2023

    work page 2023

  23. [23]

    Probability inequalities for sums of bounded random variables.Journal of the American Statistical Association, 58:13–30, 1963

    Wassily Hoeffding. Probability inequalities for sums of bounded random variables.Journal of the American Statistical Association, 58:13–30, 1963

    work page 1963

  24. [24]

    Accurate predictions on small data with a tabular foundation model,

    Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tab- ular foundation model.Nature, 637(8045):319–326, 2025. doi: 10.1038/s41586-024-08328-6

    work page doi:10.1038/s41586-024-08328-6 2025

  25. [25]

    Fast- SHAP: Real-Time Shapley Value Estimation

    Neil Jethani, Mukund Sudarshan, Ian Connick Covert, Su-In Lee, and Rajesh Ranganath. Fast- SHAP: Real-Time Shapley Value Estimation. InProceedings of the International Conference on Learning Representations (ICLR), 2022

    work page 2022

  26. [26]

    Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J. Spanos. Towards efficient data valuation based on the shapley value. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 1167–1176, 2019

    work page 2019

  27. [27]

    Hierarchical Banzhaf interaction for general video-language representation learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):2125–2139, 2025

    Peng Jin, Hao Li, Li Yuan, Shuicheng Yan, and Jie Chen. Hierarchical Banzhaf interaction for general video-language representation learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):2125–2139, 2025. 11

    work page 2025

  28. [28]

    SPEX: Scaling feature interaction explanations for LLMs

    Justin Singh Kang, Landon Butler, Abhineet Agarwal, Yigit Efe Erginbas, Ramtin Pedarsani, Bin Yu, and Kannan Ramchandran. SPEX: Scaling feature interaction explanations for LLMs. InProceedings of the Conference on Machine Learning (ICML), pages 28878–28903, 2025

    work page 2025

  29. [29]

    Derivation and validation of toxicophores for mutagenicity prediction.Journal of Medicinal Chemistry, 48(1):312–320, 2005

    Jeroen Kazius, Ross McGuire, and Roberta Bursi. Derivation and validation of toxicophores for mutagenicity prediction.Journal of Medicinal Chemistry, 48(1):312–320, 2005. doi: 10.1021/jm040835a

    work page doi:10.1021/jm040835a 2005

  30. [30]

    LightGBM: A Highly Efficient Gradient Boosting Decision Tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 3146–3154, 2017

    work page 2017

  31. [31]

    Approximating the shapley value without marginal contributions

    Patrick Kolpaczki, Viktor Bengs, Maximilian Muschalik, and Eyke Hüllermeier. Approximating the shapley value without marginal contributions. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 13246–13255, 2024

    work page 2024

  32. [32]

    SV ARM-IQ: efficient approximation of any-order shapley interactions through stratification

    Patrick Kolpaczki, Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer, and Eyke Hüllermeier. SV ARM-IQ: efficient approximation of any-order shapley interactions through stratification. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 3520–3528, 2024

    work page 2024

  33. [33]

    Datasets: A community library for natural language processing

    Quentin Lhoest, Albert Villanova del Moral, Patrick von Platen, Thomas Wolf, Mario Šaško, Yacine Jernite, Abhishek Thakur, Lewis Tunstall, Suraj Patil, Mariama Drame, Julien Chaumond, Julien Plu, Joe Davison, Simon Brandeis, Victor Sanh, Teven Le Scao, Kevin Canwen Xu, Nicolas Patry, Steven Liu, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Nat...

    work page 2021

  34. [34]

    Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In European Conference on Computer Vision ECCV, volume 8693, pages 740–755, 2014

    work page 2014

  35. [35]

    Smac3: A versatile bayesian optimization package for hyperparameter optimization.Journal of Machine Learning Research, 23(54):1–9, 2022

    Marius Lindauer, Katharina Eggensperger, Matthias Feurer, André Biedenkapp, Difan Deng, Carolin Benjamins, Tim Ruhkopf, René Sass, and Frank Hutter. Smac3: A versatile bayesian optimization package for hyperparameter optimization.Journal of Machine Learning Research, 23(54):1–9, 2022. URLhttp://jmlr.org/papers/v23/21-0888.html

    work page 2022

  36. [36]

    Disentangling environmental effects on perovskite solar cell performance via interpretable machine learning.ACS Energy Letters, 11: 1609–1617, 2026

    Tianran Liu, Nicky Evans, Kangyu Ji, Ronaldo Lee, Aaron Zhu, Vinn Nguyen, James Serdy, Elizabeth M Wall, Yongli Lu, Florian A Formica, et al. Disentangling environmental effects on perovskite solar cell performance via interpretable machine learning.ACS Energy Letters, 11: 1609–1617, 2026

    work page 2026

  37. [37]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. A Unified Approach to Interpreting Model Predictions. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 4765– 4774, 2017

    work page 2017

  38. [38]

    Lundberg, Gabriel G

    Scott M. Lundberg, Gabriel G. Erion, Hugh Chen, Alex J. DeGrave, Jordan M. Prutkin, Bala Nair, Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to global understanding with explainable AI for trees.Nature Machine Intelligence, 2(1):56–67,

    work page

  39. [39]

    doi: 10.1038/s42256-019-0138-9

    work page doi:10.1038/s42256-019-0138-9

  40. [40]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y . Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the Association for Computational Linguistics: Human Language Technologies (HLT), pages 142–150, 2011

    work page 2011

  41. [41]

    Axiomatic characterizations of generalized values.Discrete Applied Mathematics, 155(1):26–43, 2007

    Jean-Luc Marichal, Ivan Kojadinovic, and Katsushige Fujimoto. Axiomatic characterizations of generalized values.Discrete Applied Mathematics, 155(1):26–43, 2007. doi: 10.1016/J.DAM. 2006.05.002. 12

    work page doi:10.1016/j.dam 2007

  42. [42]

    Amortized Linear-time Exact Shapley Value for Product-Kernel Methods

    Majid Mohammadi, Siu Lun Chau, and Krikamol Muandet. Computing exact Shapley values in polynomial time for product-kernel methods.arXiv preprint, arXiv:2505.16516, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    General pitfalls of model-agnostic interpretation methods for machine learning models

    Christoph Molnar, Gunnar König, Julia Herbinger, Timo Freiesleben, Susanne Dandl, Chris- tian A Scholbeck, Giuseppe Casalicchio, Moritz Grosse-Wentrup, and Bernd Bischl. General pitfalls of model-agnostic interpretation methods for machine learning models. InxxAI-Beyond Explainable AI: International Workshop, Held in Conjunction with ICML 2020, pages 39–68, 2022

    work page 2020

  44. [44]

    shapiq: Shapley Interactions for Machine Learning

    Maximilian Muschalik, Hubert Baniecki, Fabian Fumagalli, Patrick Kolpaczki, Barbara Ham- mer, and Eyke Hüllermeier. shapiq: Shapley Interactions for Machine Learning. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 130324–130357, 2024

    work page 2024

  45. [45]

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

    Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer, and Eyke Hüllermeier. Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 14388–14396,

    work page

  46. [46]

    doi: 10.1609/aaai.v38i13.29352

    work page doi:10.1609/aaai.v38i13.29352

  47. [47]

    Exact Computation of Any- Order Shapley Interactions for Graph Neural Networks

    Maximilian Muschalik, Fabian Fumagalli, Paolo Frazzetto, Janine Strotherm, Luca Hermes, Alessandro Sperduti, Eyke Hüllermeier, and Barbara Hammer. Exact Computation of Any- Order Shapley Interactions for Graph Neural Networks. InProceedings of the Conference on Learning Representations (ICLR), 2025

    work page 2025

  48. [48]

    Teal Witter

    Christopher Musco and R. Teal Witter. Provably Accurate Shapley Value Estimation via Leverage Score Sampling. InProceedings of the International Conference on Learning Repre- sentations (ICLR), 2025

    work page 2025

  49. [49]

    From decision trees to boolean logic: A fast and unified SHAP algorithm

    Alexander Nadel and Ron Wettenstein. From decision trees to boolean logic: A fast and unified SHAP algorithm. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 24476–24485, 2026. doi: 10.1609/AAAI.V40I29.39630

    work page doi:10.1609/aaai.v40i29.39630 2026

  50. [50]

    Lars H. B. Olsen, Ingrid K. Glad, Martin Jullum, and Kjersti Aas. Using Shapley values and variational autoencoders to explain predictive models with dependent mixed features.Journal of Machine Learning Research, 23(213):1–51, 2022

    work page 2022

  51. [51]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InProceed- ings of the International Conference on Machine Learning ICML, pages 8748–8763, 2021

    work page 2021

  52. [52]

    The shapley value in machine learning

    Benedek Rozemberczki, Lauren Watson, Péter Bayer, Hao-Tsung Yang, Oliver Kiss, Sebastian Nilsson, and Rik Sarkar. The shapley value in machine learning. InProceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 5572–5579, 2022

    work page 2022

  53. [53]

    Evaluating attribution for graph neural networks

    Benjamin Sanchez-Lengeling, Jennifer Wei, Brian Lee, Emily Reif, Peter Wang, Wesley Qian, Kevin McCloskey, Lucy Colwell, and Alexander Wiltschko. Evaluating attribution for graph neural networks. InThe Thirty-third Annual Conference on Neural Information Processing Systems, volume 33, pages 5898–5910, 2020

    work page 2020

  54. [54]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter.CoRR, abs/1910.01108, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  55. [55]

    Investigating the impact of conceptual metaphors on LLM-based NLI through shapley interactions

    Meghdut Sengupta, Maximilian Muschalik, Fabian Fumagalli, Barbara Hammer, Eyke Hüller- meier, Debanjan Ghosh, and Henning Wachsmuth. Investigating the impact of conceptual metaphors on LLM-based NLI through shapley interactions. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 17393–17403, 2025

    work page 2025

  56. [56]

    L. S. Shapley. A Value for n-Person Games. InContributions to the Theory of Games (AM-28), Volume II, pages 307–318. Princeton University Press, 1953

    work page 1953

  57. [57]

    Adaptive prompting: Ad-hoc prompt composition for social bias detection

    Maximilian Spliethöver, Tim Knebler, Fabian Fumagalli, Maximilian Muschalik, Barbara Hammer, Eyke Hüllermeier, and Henning Wachsmuth. Adaptive prompting: Ad-hoc prompt composition for social bias detection. InProceedings of the Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 13

    work page 2025

  58. [58]

    An Efficient Explanation of Individual Classifications using Game Theory.Journal of Machine Learning Research, 11:1–18, 2010

    Erik Strumbelj and Igor Kononenko. An Efficient Explanation of Individual Classifications using Game Theory.Journal of Machine Learning Research, 11:1–18, 2010

    work page 2010

  59. [59]

    Explaining prediction models and individual predictions with feature contributions.Knowledge and Information Systems, 41(3):647–665, 2014

    Erik Strumbelj and Igor Kononenko. Explaining prediction models and individual predictions with feature contributions.Knowledge and Information Systems, 41(3):647–665, 2014. doi: 10.1007/s10115-013-0679-x

    work page doi:10.1007/s10115-013-0679-x 2014

  60. [60]

    The Shapley Taylor Interaction Index

    Mukund Sundararajan, Kedar Dhamdhere, and Ashish Agarwal. The Shapley Taylor Interaction Index. InProceedings of the International Conference on Machine Learning (ICML), pages 9259–9268, 2020

    work page 2020

  61. [61]

    Faith-Shap: The Faithful Shapley Interaction Index.Journal of Machine Learning Research, 24(94):1–42, 2023

    Che-Ping Tsai, Chih-Kuan Yeh, and Pradeep Ravikumar. Faith-Shap: The Faithful Shapley Interaction Index.Journal of Machine Learning Research, 24(94):1–42, 2023

    work page 2023

  62. [62]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Al- abdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. SigLIP 2: Multilingual vision-language encoders with improved semantic understanding, local- ization, and dense features.arXiv preprint arXiv:2502.14786, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  63. [63]

    Wang and Ruoxi Jia

    Jiachen T. Wang and Ruoxi Jia. Data banzhaf: A robust data valuation framework for machine learning. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 6388–6421, 2023

    work page 2023

  64. [64]

    Wang, Prateek Mittal, and Ruoxi Jia

    Jiachen T. Wang, Prateek Mittal, and Ruoxi Jia. Efficient data Shapley for weighted nearest neighbor algorithms. InProceedings of the International Conference on Artificial Intelligence and Statistics (AISTATS), pages 2557–2565, 2024

    work page 2024

  65. [65]

    HyperSHAP: Shapley Values and Interactions for Hyperparameter Importance

    Marcel Wever, Maximilian Muschalik, Fabian Fumagalli, and Marius Lindauer. HyperSHAP: Shapley Values and Interactions for Hyperparameter Importance. InAAAI, 2026

    work page 2026

  66. [66]

    Teal Witter, Yurong Liu, and Christopher Musco

    R. Teal Witter, Yurong Liu, and Christopher Musco. Regression-adjusted monte carlo esti- mators for shapley values and probabilistic values. InProceedings of Advances in Neural Information Processing Systems (NeurIPS), 2025. URL https://openreview.net/forum? id=Qabko39AS5

    work page 2025

  67. [67]

    How powerful are graph neural networks? InProceedings of the International Conference on Learning Representations (ICLR),

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InProceedings of the International Conference on Learning Representations (ICLR),

    work page

  68. [68]

    URLhttps://openreview.net/forum?id=ryGs6iA5Km

    work page

  69. [69]

    Interventional SHAP values and interac- tion values for piecewise linear regression trees

    Artjom Zern, Klaus Broelemann, and Gjergji Kasneci. Interventional SHAP values and interac- tion values for piecewise linear regression trees. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 11164–11173, 2023

    work page 2023

  70. [70]

    Proxy-based Approximation of Shapley and Banzhaf Interactions

    Chenyang Zhao, Kun Wang, Janet H. Hsiao, and Antoni B. Chan. Grad-ECLIP: Gradient-based visual and textual explanations for CLIP. InProceedings of the International Conference on Machine Learning ICML, 2024. 14 Appendix for “Proxy-based Approximation of Shapley and Banzhaf Interactions” A Proofs 16 A.1 Proof of Proposition 3.2 . . . . . . . . . . . . . . ...

    work page 2024

  71. [71]

    We also compare leverage weights, as used in LeverageSHAP [45], with KernelSHAP- IQ weights [ 16]

    as model-agnostic residual approxima- tors. We also compare leverage weights, as used in LeverageSHAP [45], with KernelSHAP- IQ weights [ 16]. As underlying games, we use VIT4BY4PATCHES, BIKESHARINGLO- CALXAI, CALIFORNIAHOUSINGLOCALXAI, CORRGROUPS60LOCALXAI, and COMMUNI- TIESANDCRIMELOCALXAI; details on these datasets are provided in Section C.1. For each...

    work page

  72. [72]

    Sampling and evaluation.Coalitions T ⊆2 N are sampled and evaluated, yielding the dataset D={(T, ν(T))} T∈T

    work page

  73. [73]

    Proxy fitting.A gradient-boosted tree model, by default LightGBM, is fitted on D by minimizing the mean squared error

    work page

  74. [74]

    ProxySPEX then keeps a minimal subset C ⋆ ⊆ F of coefficients that explains at least95%of the total squared Fourier mass, C ⋆ = arg min C⊆F |C|s.t

    Fourier extraction and truncation.Fourier coefficients are extracted from the fitted tree proxy. ProxySPEX then keeps a minimal subset C ⋆ ⊆ F of coefficients that explains at least95%of the total squared Fourier mass, C ⋆ = arg min C⊆F |C|s.t. P F∈C F 2 P F∈F F 2 ≥0.95, whereFdenotes the set of Fourier coefficients extracted from the tree

    work page

  75. [75]

    Limitations

    Adjustment.Given the truncated coefficient set C ⋆, ProxySPEX applies a refinement step to improve the extracted Fourier coefficients. It constructs a design matrix X∈ {−1,+1} |T |×|C ⋆| with entries Xi,j = (−1)|Ti∩Cj |, and solves the regularized regression problem F ⋆ = arg min F∈R |C⋆ | ∥ν−XF∥ 2 2 +λ∥F∥ 2 2. The truncation step is essential for making ...

    work page

  76. [76]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

    work page