pith. machine review for the scientific record. sign in

arxiv: 2604.10569 · v1 · submitted 2026-04-12 · 💻 cs.LG

Recognition: unknown

WOODELF-HD: Efficient Background SHAP for High-Depth Decision Trees

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:28 UTC · model grok-4.3

classification 💻 cs.LG
keywords Background SHAPDecision treesTree ensemblesExplainable AIMatrix multiplicationEfficient algorithmsInterpretabilityHigh-depth trees
0
0 comments X

The pith

WoodelfHD reduces the preprocessing cost of exact Background SHAP from 3^D to 2^D for decision trees.

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

Decision tree ensembles power many predictive models, and Background SHAP gives accurate explanations by treating missing features with data from a background set. Earlier efficient methods still required preprocessing that scaled as 3 to the power of tree depth D, which quickly exhausted memory for anything beyond shallow trees. WoodelfHD applies a Strassen-like multiplication scheme tailored to the structure of Woodelf matrices and merges path nodes that share the same feature. These steps lower the dominant cost to 2^D while keeping results exact. The change makes full Background SHAP practical on ordinary hardware for trees up to depth 21 and yields large speed gains on ensembles of depth 12 and 15.

Core claim

The paper shows that a Strassen-like multiplication scheme exploiting Woodelf matrix structure reduces matrix-vector multiplication from quadratic to O(k log k) in a fully vectorized non-recursive form, while merging nodes with identical features shrinks cache and memory requirements. Together these changes replace the 3^D preprocessing bottleneck with 2^D, enabling exact Background SHAP on trees up to depth 21 and delivering speedups of 33x and 162x for depth-12 and depth-15 ensembles over prior state-of-the-art methods.

What carries the argument

The Strassen-like multiplication scheme for Woodelf matrices that reduces matrix-vector multiplication to O(k log k) via vectorized non-recursive implementation, together with merging of path nodes that share identical features.

If this is right

  • Exact Background SHAP becomes feasible for individual decision trees of depth up to 21 on standard computing hardware.
  • Tree ensembles at depths 12 and 15 see speedups of 33 times and 162 times respectively over the previous best exact method.
  • Memory footprint drops enough to avoid the excessive usage that stopped earlier methods at shallower depths.

Where Pith is reading between the lines

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

  • The same matrix-structure optimizations might apply to other tree-explanation algorithms that face similar exponential preprocessing costs.
  • Practitioners could now choose deeper trees for accuracy without losing the option of precise, exact SHAP explanations.
  • Further reductions in the exponential base could become possible if the merging idea is combined with additional structural properties of decision paths.

Load-bearing premise

The Strassen-like multiplication scheme preserves exactness for the specific Woodelf matrices and that merging path nodes with identical features does not alter the computed SHAP values.

What would settle it

A side-by-side comparison of SHAP values produced by WoodelfHD against a brute-force exact reference on any tree of depth 10 or less where both methods fit in memory, or an attempt to run the algorithm on a depth-22 tree to check whether memory usage stays within standard limits.

Figures

Figures reproduced from arXiv: 2604.10569 by Alexander Nadel, Ron Wettenstein, Udi Boker.

Figure 1
Figure 1. Figure 1: Notation. complexities have a (3DD) component, where D is the tree depth. In practice, this makes the algorithms diffi￾cult to use once the maximum depth exceeds about 12, as runtimes increase sharply and memory requirements typ￾ically become prohibitive. Although limiting tree depth is a common way to con￾trol overfitting [5], deep trees are still frequently used in practice. For example, LIGHTGBM [19] em… view at source ↗
Figure 2
Figure 2. Figure 2: A consumer row, baseline row and root-to-leaf [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The M matrices corresponding to a root-to-leaf path with 6 unique features. The matrix M[xa] matches feature a from the first split, M[xb] matches feature b, and so on. Each cell is colored according to its value: 1 appears as dark red, −1 as dark blue, 0 as white, and intermediate values are shown with opacity proportional to their magnitude. appendix can help generalize the approach discussed in this sec… view at source ↗
Figure 4
Figure 4. Figure 4: Using the matrix properties to perform matrix–vector multiplication with only 2 inner multiplications instead [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Python implementation of StrassenLikeMult. Here, Mdiag is a NumPy array of length 2 D containing the secondary diagonal of M, f is a NumPy array of length 2 D, and D is the tree depth. The function computes M ·f, see Appendix B for intuition and further details. directly into the formal definition of decision patterns. Using this formulation, we derive an algorithm for computing decision patterns that sign… view at source ↗
Figure 6
Figure 6. Figure 6: Mathematical justification for merging all same-feature splits into a single constraint. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example of the Mcubes matrix for x1 and x2. add x1 add x2 add x2 add x3 add x3 add x3 add x3 add ¬x1 add ¬x2 add ¬x3 ∅ ∅ ∅ ∅ ∅ ∅ ∅ (¬x1 ∧ ¬x2 ∧ ¬x3) ∅ ∅ ∅ ∅ ∅ ∅ (¬x1 ∧ ¬x2 ∧ x3) (¬x1 ∧ ¬x2) add ¬x3 ∅ ∅ ∅ ∅ ∅ (¬x1 ∧ x2 ∧ ¬x3) ∅ (¬x1 ∧ ¬x3) ∅ ∅ ∅ ∅ (¬x1 ∧ x2 ∧ x3) (¬x1 ∧ x2) (¬x1 ∧ x3) (¬x1) add ¬x2 add ¬x3 ∅ ∅ ∅ (x1 ∧ ¬x2 ∧ ¬x3) ∅ ∅ ∅ (¬x2 ∧ ¬x3) ∅ ∅ (x1 ∧ ¬x2 ∧ x3) (x1 ∧ ¬x2) ∅ ∅ (¬x2 ∧ x3) (¬x2) add ¬x3 ∅… view at source ↗
Figure 8
Figure 8. Figure 8: Example of the Mcubes matrix for x1, x2, x3. Algorithm 4 Recursive M·v Strassen-Like Multiplication 1: function MVRECURSIVE(M, v) 2: if M and v are scalars then 3: return M · v ▷ Split v to halves and M to quarters. 4: v1, v2 = v 5: M1, M2, M3, M4 = M ▷ Assume M1 = 0 and M2 + M3 = M4. ▷ Downward Pass: 6: r1 = MVRECURSIVE(M2, v2) 7: r2 = MVRECURSIVE(M3, v1 + v2) ▷ Upward Pass: 8: r = np.concat([r1, r1 + r2]… view at source ↗
Figure 9
Figure 9. Figure 9: Recursive and vectorized multiplication, as formalized by Alg. 4 and Fig. 5, applied to a simple example. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
read the original abstract

Decision-tree ensembles are a cornerstone of predictive modeling, and SHAP is a standard framework for interpreting their predictions. Among its variants, Background SHAP offers high accuracy by modeling missing features using a background dataset. Historically, this approach did not scale well, as the time complexity for explaining n instances using m background samples included an O(mn) component. Recent methods such as Woodelf and PLTreeSHAP reduce this to O(m+n), but introduce a preprocessing bottleneck that grows as 3^D with tree depth D, making them impractical for deep trees. We address this limitation with WoodelfHD, a Woodelf extension that reduces the 3^D factor to 2^D. The key idea is a Strassen-like multiplication scheme that exploits the structure of Woodelf matrices, reducing matrix-vector multiplication from O(k^2) to O(k*log(k)) via a fully vectorized, non-recursive implementation. In addition, we merge path nodes with identical features, reducing cache size and memory usage. When running on standard environments, WoodelfHD enables exact Background SHAP computation for trees with depths up to 21, where previous methods fail due to excessive memory usage. For ensembles of depths 12 and 15, it achieves speedups of 33x and 162x, respectively, over the state-of-the-art.

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

3 major / 2 minor

Summary. The manuscript proposes WoodelfHD, an extension of Woodelf for exact Background SHAP on high-depth decision trees. It reduces the preprocessing complexity from 3^D to 2^D via a Strassen-like matrix multiplication scheme that exploits Woodelf matrix structure to achieve O(k log k) matrix-vector multiplication in a fully vectorized non-recursive form, together with merging of path nodes sharing identical features to lower memory. The work claims this enables exact SHAP for depths up to 21 (where prior methods exhaust memory) and delivers speedups of 33x (depth 12) and 162x (depth 15) over the state of the art.

Significance. If the transformations preserve exact SHAP values, the result would meaningfully extend the practical reach of accurate background SHAP to deeper trees that dominate many deployed ensembles, removing a long-standing memory barrier in tree-based interpretability.

major comments (3)
  1. [Abstract] Abstract: the central claim that the Strassen-like scheme and identical-feature node merging both return identical SHAP values to the original 3^D preprocessing is load-bearing for every reported speedup and depth-21 feasibility statement, yet the manuscript provides neither a formal invariance proof nor an exhaustive numerical comparison of output vectors for any depth where both methods can run.
  2. [§3] §3 (algorithm description): the O(k log k) routine is presented as derived from Woodelf matrix structure, but no explicit verification is given that the reduced multiplication produces the same contribution vectors as the naive O(k^2) form for the specific matrices arising in Background SHAP; a small-D example matrix and its transformed product would directly test this.
  3. [§4] §4 (experiments): the 33x and 162x speedups are reported without accompanying error metrics; for depths 12 and 15 a table of maximum absolute deviation between WoodelfHD and Woodelf SHAP scores (or at least for a subset of instances) is required to substantiate the exactness assertion.
minor comments (2)
  1. [Abstract] The abstract refers to 'standard environments' and 'k' without defining the latter or listing hardware/software details needed for reproducibility of the reported timings.
  2. Notation for the Woodelf matrices and the precise definition of the merge operation on path nodes would benefit from an explicit small example (e.g., D=3) to clarify the claimed memory reduction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive feedback. We address the concerns regarding verification of exact SHAP values point by point below, and will incorporate revisions to strengthen the evidence.

read point-by-point responses
  1. Referee: [Abstract] Abstract: the central claim that the Strassen-like scheme and identical-feature node merging both return identical SHAP values to the original 3^D preprocessing is load-bearing for every reported speedup and depth-21 feasibility statement, yet the manuscript provides neither a formal invariance proof nor an exhaustive numerical comparison of output vectors for any depth where both methods can run.

    Authors: The transformations are constructed to preserve exact equivalence by leveraging the recursive block structure of the Woodelf matrices, similar to how Strassen's algorithm preserves matrix multiplication results. To substantiate this, we will add a formal invariance proof in the appendix and include exhaustive numerical comparisons of SHAP output vectors for all depths up to 8, where both methods can execute. revision: yes

  2. Referee: [§3] §3 (algorithm description): the O(k log k) routine is presented as derived from Woodelf matrix structure, but no explicit verification is given that the reduced multiplication produces the same contribution vectors as the naive O(k^2) form for the specific matrices arising in Background SHAP; a small-D example matrix and its transformed product would directly test this.

    Authors: We will include a small-D example (for depth 2 or 3) in the revised §3, explicitly showing the Woodelf matrix, the naive product, the reduced O(k log k) computation, and verifying that the contribution vectors are identical. revision: yes

  3. Referee: [§4] §4 (experiments): the 33x and 162x speedups are reported without accompanying error metrics; for depths 12 and 15 a table of maximum absolute deviation between WoodelfHD and Woodelf SHAP scores (or at least for a subset of instances) is required to substantiate the exactness assertion.

    Authors: Direct comparison for depths 12 and 15 is not possible as Woodelf exhausts memory, which is the motivation for our work. However, we will add a table in §4 showing maximum absolute deviations for depths 5-10 (where both run), confirming zero or negligible error (within floating point precision). This, combined with the proof, supports exactness for higher depths. revision: partial

Circularity Check

0 steps flagged

No circularity: algorithmic derivation self-contained via matrix structure

full rationale

The paper derives a complexity reduction from 3^D to 2^D by presenting an explicit Strassen-like matrix-vector multiplication scheme that exploits the structure of Woodelf matrices, together with a node-merging step for identical features. These steps are described as direct consequences of the matrix properties and path structure rather than any fitted parameter, self-definition, or load-bearing self-citation. The central claims rest on the stated exactness of the transformations and the O(k log k) routine, which are presented as first-principles algorithmic constructions without reducing to prior results by construction or renaming. No load-bearing step equates a derived quantity to its own input.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The method rests on standard assumptions about decision tree structure and matrix algebra; no free parameters, ad-hoc axioms, or new invented entities are introduced in the abstract.

pith-pipeline@v0.9.0 · 5546 in / 1069 out tokens · 41502 ms · 2026-05-10T16:28:35.744717+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

39 extracted references · 16 canonical work pages · 1 internal anchor

  1. [1]

    Marcelo Arenas, Pablo Barcel ´o, Leopoldo Bertossi, and Mika ¨el Monet. The tractability of shap-score- based explanations for classification over determin- istic and decomposable boolean circuits.Proceed- ings of the AAAI Conference on Artificial Intelli- gence, 35(8):6670–6678, May 2021. doi: 10.1609/ aaai.v35i8.16825

  2. [2]

    A machine learning approach for credit card fraud detection in massive datasets using smote and ran- dom sampling

    Sreenivas Prasad B, Akash Babu N, Harthikeswar Reddy, Rimjhim Padam Singh, and Sneha Kanchan. A machine learning approach for credit card fraud detection in massive datasets using smote and ran- dom sampling. In2024 IEEE Recent Advances in Intelligent Computational Systems (RAICS), pages 1–8, 2024. doi: 10.1109/RAICS61201.2024. 10690025

  3. [3]

    John F. Banzhaf. Weighted voting doesn’t work: A mathematical analysis.Rutgers Law Review, 19: 317–343, 1965

  4. [4]

    Jour- nal of Automated Reasoning26(2), 107–137 (2001) https://doi.org/10.1023/A: 1026518331905

    Leo Breiman. Random forests.Machine Learn- ing, 45(1):5–32, Oct 2001. doi: 10.1023/A: 1010933404324

  5. [5]

    Wadsworth, 1984

    Leo Breiman, Jerome Friedman, Richard Olshen, and Charles Stone.Classification and Regression Trees. Wadsworth, 1984

  6. [6]

    De Carli, A

    L. De Carli, A. Echezabal, and I. Morell. On the sierpi ´nski triangle and its generalizations. Journal of Mathematics and the Arts, 19(3-4): 79–92, 2025. doi: 10.1080/17513472.2025. 2576836. URLhttps://doi.org/10.1080/ 17513472.2025.2576836

  7. [7]

    A stacknet based model for fraud detec- tion

    Liiie Chen, Qihan Guan, Ning Chen, and Zhou Yi- Hang. A stacknet based model for fraud detec- tion. In2021 2nd International Conference on Ed- ucation, Knowledge and Information Management (ICEKIM), pages 328–331, 2021. doi: 10.1109/ ICEKIM52309.2021.00079

  8. [8]

    Why Should I Trust You?

    Tianqi Chen and Carlos Guestrin. Xgboost: A scal- able tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 785–794, New York, NY , USA, 2016. Asso- ciation for Computing Machinery. doi: 10.1145/ 2939672.2939785

  9. [9]

    A data mining based system for trans- action fraud detection

    Wenkai Deng, Ziming Huang, Jiachen Zhang, and Junyan Xu. A data mining based system for trans- action fraud detection. In2021 IEEE International Conference on Consumer Electronics and Computer Engineering (ICCECE), pages 542–545, 2021. doi: 10.1109/ICCECE51280.2021.9342376

  10. [10]

    Robert L. Devaney. Measure, topology, and fractal geometry.SIAM review, 33(4):668–669, 1991. ISSN 0036-1445

  11. [11]

    Nep 38 – simd opti- mizations.https://numpy.org/neps/nep- 0038-SIMD-optimizations.html, 2019

    NumPy Developers. Nep 38 – simd opti- mizations.https://numpy.org/neps/nep- 0038-SIMD-optimizations.html, 2019

  12. [12]

    Simd optimizations documen- tation.https://numpy.org/doc/stable/ reference/simd/index.html, 2024

    NumPy Developers. Simd optimizations documen- tation.https://numpy.org/doc/stable/ reference/simd/index.html, 2024

  13. [13]

    CatBoost: unbiased boosting with categorical features

    Anna Veronika Dorogush, Andrey Gulin, Gleb Gusev, Nikita Kazeev, Liudmila Ostroumova Prokhorenkova, and Aleksandr V orobev. Fighting biases with dynamic boosting.CoRR, 2017. URL http://arxiv.org/abs/1706.09516

  14. [14]

    An axiomatic approach to the concept of interaction among play- ers in cooperative games.International Journal of Game Theory, 28(4):547–565, Nov 1999

    Michel Grabisch and Marc Roubens. An axiomatic approach to the concept of interaction among play- ers in cooperative games.International Journal of Game Theory, 28(4):547–565, Nov 1999. doi: 10.1007/s001820050125

  15. [15]

    Harris, K

    Charles R. Harris, K. Jarrod Millman, St ´efan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebas- tian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fern´andez del R´ıo, Mark Wiebe, Pearu Peterson, Pierre G ´erard- Marchant, Ke...

  16. [16]

    doi: 10.1038/s41586-020-2649-2

  17. [17]

    Up- dates on the complexity of shap scores

    Xuanxiang Huang and Joao Marques-Silva. Up- dates on the complexity of shap scores. InProceed- ings of the Thirty-Third International Joint Confer- ence on Artificial Intelligence, IJCAI ’24, 2024. doi: 10.24963/ijcai.2024/45

  18. [18]

    Single-instruction multiple- data execution.Synthesis Lectures on Computer Architecture, 10:1–121, 05 2015

    Christopher Hughes. Single-instruction multiple- data execution.Synthesis Lectures on Computer Architecture, 10:1–121, 05 2015. doi: 10.2200/ S00647ED1V01Y201505CAC032

  19. [19]

    Credit card fraud detection based on unsupervised attentional anomaly detection net- work.Systems, 11:305, 06 2023

    Shanshan Jiang, Ruiting Dong, Jie Wang, and Min Xia. Credit card fraud detection based on unsupervised attentional anomaly detection net- work.Systems, 11:305, 06 2023. doi: 10.3390/ systems11060305. 10

  20. [20]

    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. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fer- gus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Sys- tems, volume 30. Curran Associates, Inc., ...

  21. [21]

    TreeGrad-Ranker: Feature Ranking via $O(L)$-Time Gradients for Decision Trees

    Weida Li, Yaoliang Yu, and Bryan Kian Hsiang Low. Treegrad-ranker: Feature ranking viao(l)-time gradients for decision trees, 2026. URLhttps: //arxiv.org/abs/2602.11623

  22. [22]

    The shapley value of tuples in query answering.Logical Methods in Computer Sci- ence, V olume 17, Issue 3, September 2021

    Ester Livshits, Leopoldo Bertossi, Benny Kimelfeld, and Moshe Sebag. The shapley value of tuples in query answering.Logical Methods in Computer Sci- ence, V olume 17, Issue 3, September 2021. doi: 10.46298/lmcs-17(3:22)2021

  23. [23]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, ed- itors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

  24. [24]

    Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M

    Scott M. Lundberg, Gabriel Erion, Hugh Chen, Alex 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 Intel- ligence, 2:56–67, 2020. doi: 10.1038/s42256-019- 0138-9

  25. [25]

    Np- completeness for calculating power indices of weighted majority games.Theoretical Computer Science, 263(1):305–310, 2001

    Yasuko Matsui and Tomomi Matsui. Np- completeness for calculating power indices of weighted majority games.Theoretical Computer Science, 263(1):305–310, 2001. URLhttps: //www.sciencedirect.com/science/ article/pii/S0304397500002516. Com- binatorics and Computer Science

  26. [26]

    Gputreeshap: massively parallel exact calculation of shap scores for tree ensembles.PeerJ

    Rory Mitchell, Eibe Frank, and Geoffrey Holmes. Gputreeshap: massively parallel exact calculation of shap scores for tree ensembles.PeerJ. Computer science, 8:e880, 2022. ISSN 2167-9843. doi: 10. 7717/peerj-cs.880. URLhttps://europepmc. org/articles/PMC9044362

  27. [27]

    From de- cision trees to boolean logic: A fast and unified shap algorithm.Proceedings of the AAAI Con- ference on Artificial Intelligence, 40(29):24476– 24485, Mar

    Alexander Nadel and Ron Wettenstein. From de- cision trees to boolean logic: A fast and unified shap algorithm.Proceedings of the AAAI Con- ference on Artificial Intelligence, 40(29):24476– 24485, Mar. 2026. doi: 10.1609/aaai.v40i29.39630. URLhttps://ojs.aaai.org/index.php/ AAAI/article/view/39630

  28. [28]

    Tunability: Importance of hyperparameters of machine learning algorithms.Journal of Machine Learning Research, 20(53):1–32, 2019

    Philipp Probst, Anne-Laure Boulesteix, and Bernd Bischl. Tunability: Importance of hyperparameters of machine learning algorithms.Journal of Machine Learning Research, 20(53):1–32, 2019

  29. [29]

    L. S. Shapley. A value of n-person games.Contribu- tions to the Theory of Games, page 307–317, 1953

  30. [30]

    Sur une courbe dont tout point est un point de ramification.Comptes Rendus de l’Acad´emie des Sciences de Paris, 160:302–305, 1915

    Wacław Sierpi´nski. Sur une courbe dont tout point est un point de ramification.Comptes Rendus de l’Acad´emie des Sciences de Paris, 160:302–305, 1915

  31. [31]

    Gaussian elimination is not op- timal.Numerische Mathematik, 13(4):354–356, Aug 1969

    V olker Strassen. Gaussian elimination is not op- timal.Numerische Mathematik, 13(4):354–356, Aug 1969. ISSN 0945-3245. doi: 10.1007/ BF02165411. URLhttps://doi.org/10. 1007/BF02165411

  32. [32]

    On the tractability of shap explanations.J

    Guy Van den Broeck, Anton Lykov, Maximilian Schleich, and Dan Suciu. On the tractability of shap explanations.J. Artif. Int. Res., 74, September 2022. ISSN 1076-9757. doi: 10.1613/jair.1.13283

  33. [33]

    Ieee-cis fraud detection based on xgb

    Zhijia Xiao. Ieee-cis fraud detection based on xgb. In Xiaolong Li, Chunhui Yuan, and John Kent, ed- itors,Proceedings of the 7th International Confer- ence on Economic Management and Green Develop- ment, pages 1785–1796, Singapore, 2024. Springer Nature Singapore

  34. [34]

    Fast treeshap: Accelerating shap value computation for trees.arXiv preprint arXiv:2109.09847, 2021

    Jilei Yang. Fast treeshap: Accelerating shap value computation for trees, 2022. URLhttps:// arxiv.org/abs/2109.09847

  35. [35]

    Linear tree shap

    peng yu, Albert Bifet, Jesse Read, and Chao Xu. Linear tree shap. In S. Koyejo, S. Mo- hamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Informa- tion Processing Systems, volume 35, pages 25818–25828. Curran Associates, Inc., 2022. URLhttps://proceedings.neurips. cc/paper_files/paper/2022/file/ a5a3b1ef79520b7cd122d888673a3ebc-...

  36. [36]

    Interventional shap values and interaction val- ues for piecewise linear regression trees.Proceed- ings of the AAAI Conference on Artificial Intelli- gence, 37(9):11164–11173, Jun

    Artjom Zern, Klaus Broelemann, and Gjergji Kas- neci. Interventional shap values and interaction val- ues for piecewise linear regression trees.Proceed- ings of the AAAI Conference on Artificial Intelli- gence, 37(9):11164–11173, Jun. 2023. doi: 10. 1609/aaai.v37i9.26322. 11 A A Boolean Logic View of theM Matrices In this section, we define the WOODELFMma...

  37. [37]

    If thei most significant bit is 0 add the literalx i to the cube

    The boolean representation of the row index is treated as a list of positive literals indicators. If thei most significant bit is 0 add the literalx i to the cube

  38. [38]

    If the imost significant bit is 0 add the literal¬x i to the cube

    The boolean representation of the column index is treated as a list of negated literals indicators. If the imost significant bit is 0 add the literal¬x i to the cube

  39. [39]

    Remove cubes that have both the positive and the negated literal of the same variable as they are un- satisfiable. Def. 6 formalize these rule: Definition 6.Letx 1, . . . , xD be Boolean variables. For anya, b∈ {0,1} D, the cubes matrixM cubes is defined as follows (Notation:a∨bis a bitwise OR anda i is thei most significant bit ofa): Mcubes[a][b] = ( ^ i...