Pith. sign in

REVIEW 2 major objections 2 minor 33 references

SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths

T0 review · 2 major / 2 minor · reviewed 2026-06-27 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read SAILS fits GAM surrogates to local effects of black-box models to detect pairwise interactions and categorize their functional forms as linear, product-separable, or non-product-separable.

desk verdict SAILS adds a concrete way to categorize pairwise interaction shapes via GAM surrogates on local effects, but its detection and typing steps rest on an approximation whose reliability the authors themselves flag under correlation or higher-order effects. read the letter →

arxiv 2606.09404 v1 pith:2KAIZVUK submitted 2026-06-08 stat.ML cs.AIcs.LG

Timo Hei{ss} , Julia Herbinger , Bernd Bischl , Giuseppe Casalicchio This is my paper · ORCID
classification stat.MLcs.AIcs.LG
keywords featureinteractionsexplainableAIgeneralizedadditivemodelslocaleffectssurrogateinteractiondetectionXAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper presents SAILS as a way to move past simple detection of feature interactions in machine learning models toward understanding their specific shapes. It does this by training generalized additive model surrogates on the local effects computed from any black-box model, then using the smooth terms to isolate interaction signals at the derivative level. This setup supports a significance-test heuristic for detection, a three-way categorization of interaction types, and type-specific visualizations. A reader would care because knowing that two features interact is less useful than knowing whether the interaction is additive, multiplicative, or more complex. The authors test the approach on controlled simulations and one real-world dataset, noting limits when features are strongly correlated or higher-order effects dominate.

What carries the argument

GAM surrogates fitted to local effects, whose smooth terms isolate interaction components on the derivative level for each interval of a feature of interest.

What would settle it

A controlled simulation in which known pairwise interactions of each type are injected into a black-box model, yet the SAILS heuristic either misses them or assigns the wrong category after the GAM fit.

Watch

Extended reading notes

Core claim

SAILS is a model-agnostic framework that analyzes pairwise interactions through interpretable generalized additive model (GAM) surrogates fitted to the local effects of a black-box model. For each interval of a feature of interest, the surrogate smooth terms isolate the interaction components on derivative level, enabling (i) interaction detection through a heuristic derived from significance tests on smooth terms, (ii) interaction form categorization into linear, product-separable, and non-product-separable types, and (iii) tailored, interpretable visualizations for each interaction type.

Load-bearing premise

The local effects extracted from the black-box model can be captured accurately enough by the GAM surrogates that significance tests on the smooth terms correctly identify and classify genuine interactions.

Editorial extensions

If this is right

  • Pairwise interactions can be detected via a heuristic based on significance tests of the smooth terms.
  • Detected interactions can be sorted into linear, product-separable, or non-product-separable categories.
  • Each category receives its own tailored visualization derived from the surrogate smooths.
  • The framework works on any black-box model because it operates only on the computed local effects.
  • Performance degrades when strong feature correlations or higher-order interactions are present.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the categorization proves stable, practitioners could use the type labels to decide whether to prune certain interactions during model simplification.
  • The interval-wise fitting approach might be adapted to produce partial dependence plots that explicitly separate main effects from interactions.
  • Because the method is model-agnostic, it could be applied post-hoc to ensembles or neural networks without retraining.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces SAILS, a model-agnostic XAI framework that fits GAM surrogates to local effects extracted from a black-box model. For each interval of a feature of interest, smooth terms on the surrogate isolate interaction components at the derivative level. This enables (i) interaction detection via a heuristic based on p-values of the smooth terms, (ii) categorization of detected interactions into linear, product-separable, and non-product-separable forms, and (iii) type-specific visualizations. The approach is validated on controlled simulations and one real-world task; the authors explicitly note degraded performance under strong feature correlations and higher-order interactions.

Significance. If the central assumption holds—that local effects can be faithfully recovered by the per-interval GAM surrogates so that significance tests on the smooth terms correctly detect and type pairwise interactions—SAILS would address a genuine gap by moving beyond interaction detection or restricted visualization to functional-form characterization. The provision of both simulation controls and an explicit limitations statement is a strength; the framework is falsifiable via the reported degradation regimes.

major comments (2)
  1. [Method description of the heuristic and categorization scheme] The central claim that the significance-based heuristic on GAM smooth terms reliably isolates and categorizes pairwise interactions rests on the untested assumption that local-effect surfaces remain free of omitted-variable bias when feature correlations or higher-order terms are present. The manuscript flags degraded performance in these regimes but does not quantify how often the heuristic misclassifies interaction type or produces false positives under controlled correlation strengths (e.g., ρ > 0.6).
  2. [Empirical validation (simulations)] The empirical validation section reports effectiveness on controlled simulations, yet the simulation design is not shown to include the exact stress regimes (strong correlations, higher-order interactions) that the authors themselves identify as problematic. Without those results, the claim that the framework works for pairwise interactions cannot be evaluated for robustness.
minor comments (2)
  1. [Notation section] Notation for the local-effect extraction and the per-interval GAM fitting should be unified; currently the same symbol appears to be reused for the black-box partial dependence and the surrogate smooth without explicit disambiguation.
  2. [Figure captions] Figure captions for the tailored visualizations should state the exact p-value threshold and smoothing-parameter selection method used, so that readers can reproduce the categorization boundaries.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which identify opportunities to strengthen the robustness evaluation of SAILS. We address each major comment below and will incorporate the suggested extensions in the revised manuscript.

read point-by-point responses

  1. Referee: [Method description of the heuristic and categorization scheme] The central claim that the significance-based heuristic on GAM smooth terms reliably isolates and categorizes pairwise interactions rests on the untested assumption that local-effect surfaces remain free of omitted-variable bias when feature correlations or higher-order terms are present. The manuscript flags degraded performance in these regimes but does not quantify how often the heuristic misclassifies interaction type or produces false positives under controlled correlation strengths (e.g., ρ > 0.6).

    Authors: We agree that the manuscript would benefit from explicit quantification of misclassification and false-positive rates under controlled correlation strengths. While the current text notes degradation in these regimes, it does not report specific error frequencies. In revision we will add simulations that systematically vary feature correlation (ρ = 0 to 0.8) and higher-order interaction strength, reporting detection accuracy, categorization error rates, and false-positive frequencies for the significance heuristic. revision: yes

  2. Referee: [Empirical validation (simulations)] The empirical validation section reports effectiveness on controlled simulations, yet the simulation design is not shown to include the exact stress regimes (strong correlations, higher-order interactions) that the authors themselves identify as problematic. Without those results, the claim that the framework works for pairwise interactions cannot be evaluated for robustness.

    Authors: The existing simulations were constructed to isolate the core behavior of the per-interval GAM surrogates under pairwise interactions without confounding correlations or higher-order terms. We acknowledge, however, that the stress regimes flagged in the limitations statement should be explicitly tested to bound the method’s applicability. We will therefore expand the empirical section with additional simulation designs that incorporate strong correlations and higher-order interactions, presenting the resulting performance metrics alongside the original results. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SAILS derives interaction detection and categorization from independent GAM fits to local effects

full rationale

The paper's core chain computes local effects from a black-box model, fits GAM surrogates per interval, and applies significance tests on smooth terms to produce a heuristic for detection plus a categorization into linear/product-separable/non-product-separable forms. These steps introduce new components (the heuristic and type scheme) that do not reduce by construction to the input local effects or to any self-citation. No equations equate a claimed prediction back to a fitted parameter, no uniqueness theorem is imported from prior author work, and no ansatz is smuggled via citation. The framework is self-contained against external benchmarks (simulations and real data) with explicitly stated limitations under correlations or higher-order effects, satisfying the criteria for a non-circular derivation.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only view yields no explicit free parameters, axioms, or invented entities; the method implicitly relies on standard GAM fitting assumptions and local effect definitions from prior literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths." pith.science (2026). https://pith.science/paper/2KAIZVUK

@misc{pith2026260609404,
  author       = {Pith},
  title        = {Pith review of: SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KAIZVUK}},
  note         = {Machine review of arXiv:2606.09404}
}
read the original abstract

Feature interactions drive much of the predictive power of machine learning models, yet existing explanation methods only detect and quantify interactions without revealing their functional form, or visualize only restricted interaction types. We propose Surrogate-based Analysis of Interactions via Local effect Smooths (SAILS), a model-agnostic framework that analyzes pairwise interactions through interpretable generalized additive model (GAM) surrogates fitted to the local effects of a black-box model. For each interval of a feature of interest, the surrogate smooth terms isolate the interaction components on derivative level, enabling (i) interaction detection through a heuristic derived from significance tests on smooth terms, (ii) interaction form categorization into linear, product-separable, and non-product-separable types, and (iii) tailored, interpretable visualizations for each interaction type. We empirically validate the framework through controlled simulations and a real-world task, demonstrating its effectiveness for pairwise interactions, with limitations under strong feature correlations and higher-order interactions. SAILS fills a notable gap in the XAI toolbox, going beyond detection of interactions alone to characterizing their functional form.

Figures

Figures reproduced from arXiv: 2606.09404 by the authors.

Figure 1
Figure 1. Feature effect plots for the motivating example. Dashed lines are PD [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Sails framework. ALE Local Effects. ALE [2] is a feature effect method that respects feature correlations by integrating the conditional expectation of the local derivative along an FOI. With local effects h(xj , x−j ) := ∂j ˆf(xj , x−j ), the 1D-ALE is: ALEj (xj ) = Z xj xmin,j EX−j |Xj=z [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Smooth terms sˆ (k) l (1st row) & ratios r (k) l (2nd) for ˆf(x) = 3x1 + 2x1x 2 2 + x 2 1 exp(x3) + sin(πx1x4) (FOI: X1); features are indep. U(−1, 1). Linear interac￾tion (X2): all sˆ (k) 2 are identical and so are the ratios. Product-separable interac￾tion (X3): all sˆ (k) 3 have the same shape but scaled differently, ratios are identical. Non-product-separable interaction (X4): shapes of both sˆ (k) 4 and r (k) 4… view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Interaction detection for the oracle across 30 repetitions. Green boxes: [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Categorization measures for the oracle across 30 repetitions. Colors indi [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Linear and product-separable interaction visualizations for the oracle (last [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: General interaction visualization (integrated smooths) for the oracle (last [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Interaction visualization for general_diffuse_flows as FOI of XG￾Boost on the Tetouan city dataset. Each curve corresponds to one interval of general_diffuse_flows, colored from low (purple) to high (yellow) radiation. 6 Conclusion We presented Sails, a model-agnostic …
Figure 9
Figure 9. Figure 9: Visualization of the true underlying two-way interaction effects in our [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: Full simulation results for Sails interaction detection. Approximate p￾values across 30 repetitions are visualized as boxplots. Dashed line: threshold of α = 0.05 [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Full simulation results for Sails linear interaction categorization mea￾sure. R2 lin across 30 repetitions visualized as boxplots. x2 x3 x4 x5 x6 x7 x8 x9 0.0 0.2 0.4 0.6 0.8 1.0 R 2 pro d model = Oracle x2 x3 x4 x5 x6 x7 x8 x9 model = GAM x2 x3 x4 x5 x6 x7 x8 x9 Inte…
Figure 12
Figure 12. Figure 12: Full simulation results for Sails product-separable interaction catego￾rization measure. R2 prod across 30 repetitions visualized as boxplots. For cases where all surrogate smooth terms were filtered out (below variance threshold) or the reference point did not exist …
Figure 13
Figure 13. Figure 13: Linear interaction visualizations for the oracle (last repetition). True is [PITH_FULL_IMAGE:figures/full_fig_p031_13.png]
Figure 14
Figure 14. Figure 14: Product-separable interaction visualizations (ratio-based) for the oracle [PITH_FULL_IMAGE:figures/full_fig_p032_14.png]
Figure 15
Figure 15. Figure 15: General interaction visualization (integrated smooths) for setting [PITH_FULL_IMAGE:figures/full_fig_p034_15.png]
Figure 16
Figure 16. Figure 16: General interaction visualization (integrated smooths) for setting [PITH_FULL_IMAGE:figures/full_fig_p036_16.png]
Figure 17
Figure 17. Figure 17: General interaction visualization (integrated smooths) for setting [PITH_FULL_IMAGE:figures/full_fig_p038_17.png]
Figure 18
Figure 18. Figure 18: General interaction visualization (integrated smooths) for setting [PITH_FULL_IMAGE:figures/full_fig_p040_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 3 canonical work pages

  1. [1]

    IEEE Access6, 52138–52160 (2018)

    Adadi, A., Berrada, M.: Peeking inside the black-box: A survey on Explainable Artificial Intelligence (XAI). IEEE Access6, 52138–52160 (2018)

    2018

  2. [2]

    Journal of the Royal Statistical Society Series B: Sta- tistical Methodology82(4), 1059–1086 (2020)

    Apley, D.W., Zhu, J.: Visualizing the effects of predictor variables in black box supervised learning models. Journal of the Royal Statistical Society Series B: Sta- tistical Methodology82(4), 1059–1086 (2020)

    2020

  3. [3]

    Journal of the Royal Statistical Society

    Benjamini, Y., Hochberg, Y.: Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B (Methodological)57(1), 289–300 (1995)

    1995

  4. [4]

    In: Shawe-Taylor, J., et al

    Bergstra, J., Bardenet, R., Bengio, Y., Kégl, B.: Algorithms for hyper-parameter optimization. In: Shawe-Taylor, J., et al. (eds.) Advances in Neural Information Processing Systems. vol. 24. Curran Associates, Inc. (2011)

    2011

  5. [5]

    Statistical Science16(3), 199–231 (2001)

    Breiman,L.:Statisticalmodeling:Thetwocultures(withcommentsandarejoinder by the author). Statistical Science16(3), 199–231 (2001)

    2001

  6. [6]

    Springer, Berlin, Heidelberg (2013)

    Fahrmeir, L., Kneib, T., Lang, S., Marx, B.: Regression: Models, Methods and Applications. Springer, Berlin, Heidelberg (2013)

    2013

  7. [7]

    The Annals of Statistics29(5), 1189–1232 (2001)

    Friedman, J.H.: Greedy function approximation: A gradient boosting machine. The Annals of Statistics29(5), 1189–1232 (2001)

    2001

  8. [8]

    The Annals of Applied Statistics2(3) (2008)

    Friedman, J.H., Popescu, B.E.: Predictive learning via rule ensembles. The Annals of Applied Statistics2(3) (2008)

    2008

Show all 33 references
  1. [9]

    In: Oh, A., et al

    Fumagalli, F., Muschalik, M., Kolpaczki, P., Hüllermeier, E., Hammer, B.: SHAP- IQ: Unified approximation of any-order shapley interactions. In: Oh, A., et al. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 11515– 11551. Curran Associates, Inc. (2023)

    2023

  2. [10]

    In: ECAI 2023, pp

    Gkolemis, V., Dalamagas, T., Ntoutsi, E., Diou, C.: RHALE: Robust and heterogeneity-aware accumulated local effects. In: ECAI 2023, pp. 859–866. IOS Press (2023)

    2023

  3. [11]

    Journal of Computational and Graphical Statistics24(1), 44–65 (2015)

    Goldstein, A., Kapelner, A., Bleich, J., Pitkin, E.: Peeking inside the black box: Visualizing statistical learning with plots of individual conditional expectation. Journal of Computational and Graphical Statistics24(1), 44–65 (2015)

    2015

  4. [12]

    International Journal of Game Theory28(4), 547–565 (1999)

    Grabisch, M., Roubens, M.: An axiomatic approach to the concept of interaction among players in cooperative games. International Journal of Game Theory28(4), 547–565 (1999)

    1999

  5. [13]

    arXiv:1805.04755 [stat.ML] (2018)

    Greenwell, B.M., Boehmke, B.C., McCarthy, A.J.: A simple and effective model- based variable importance measure. arXiv:1805.04755 [stat.ML] (2018)

    2018 arXiv

  6. [14]

    arXiv:2603.15057 [stat.ML] (2026)

    Heiß, T., Bögel, C., Bischl, B., Casalicchio, G.: Analyzing error sources in global feature effect estimation. arXiv:2603.15057 [stat.ML] (2026)

    2026

  7. [15]

    In: Proceedings of The 25th International Conference on Artificial Intelligence and Statistics

    Herbinger, J., Bischl, B., Casalicchio, G.: REPID: Regional effect plots with im- plicit interaction detection. In: Proceedings of The 25th International Conference on Artificial Intelligence and Statistics. pp. 10209–10233. PMLR (2022) 16 T. Heiß et al

    2022

  8. [16]

    In: Nowaczyk, S., et al

    Herbinger, J., Dandl, S., Ewald, F.K., Loibl, S., Casalicchio, G.: Leveraging model- based trees as interpretable surrogate models for model distillation. In: Nowaczyk, S., et al. (eds.) Artificial Intelligence. ECAI 2023 International Workshops. pp. 232–249. Springer Nature S...

    2023

  9. [17]

    Journal of Machine Learning Research25(381), 1–65 (2024)

    Herbinger, J., Wright, M.N., Nagler, T., Bischl, B., Casalicchio, G.: Decomposing global feature effects based on feature interactions. Journal of Machine Learning Research25(381), 1–65 (2024)

    2024

  10. [18]

    In: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining

    Hooker, G.: Discovering additive structure in black box functions. In: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. pp. 575–580. KDD ’04, Association for Computing Machinery, New York, NY, USA (2004)

    2004

  11. [19]

    Journal of Computational and Graphical Statistics 16(3), 709–732 (2007)

    Hooker, G.: Generalized functional ANOVA diagnostics for high-dimensional func- tions of dependent variables. Journal of Computational and Graphical Statistics 16(3), 709–732 (2007)

    2007

  12. [20]

    arXiv:2503.16445 [cs.HC] (2025)

    Kleinau, A., Preim, B., Meuschke, M.: FINCH: Locally visualizing higher-order feature interactions in black box models. arXiv:2503.16445 [cs.HC] (2025)

    2025

  13. [21]

    Journal of Mathematical Chemistry50(1), 99–130 (2012)

    Li, G., Rabitz, H.: General formulation of HDMR component functions with in- dependent and correlated variables. Journal of Mathematical Chemistry50(1), 99–130 (2012)

    2012

  14. [22]

    Nat Mach Intell2(1), 56–67 (2020)

    Lundberg, S.M., Erion, G., Chen, H., DeGrave, A., Prutkin, J.M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., Lee, S.I.: From local explanations to global under- standing with explainable AI for trees. Nat Mach Intell2(1), 56–67 (2020)

    2020

  15. [23]

    In: Globerson, A., et al

    Muschalik, M., Baniecki, H., Fumagalli, F., Kolpaczki, P., Hammer, B., Hüller- meier, E.: shapiq: Shapley interactions for machine learning. In: Globerson, A., et al. (eds.) Advances in Neural Information Processing Systems. vol. 37, pp. 130324–130357. Curran Associates, Inc. (2024)

    2024

  16. [24]

    Journal of Machine Learning Research20(53), 1– 32 (2019)

    Probst, P., Boulesteix, A.L., Bischl, B.: Tunability: Importance of hyperparameters of machine learning algorithms. Journal of Machine Learning Research20(53), 1– 32 (2019)

    2019

  17. [25]

    Data Science and Manage- ment4, 19–29 (2021)

    Pugliese, R., Regondi, S., Marini, R.: Machine learning-based approach: global trends, research directions, and regulatory standpoints. Data Science and Manage- ment4, 19–29 (2021)

    2021

  18. [26]

    why should i trust you?

    Ribeiro, M.T., Singh, S., Guestrin, C.: "why should i trust you?": Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Data Mining. pp. 1135–1144. KDD ’16, Association for Computing Machiner...

    2016

  19. [27]

    In: 6th International Renew- able and Sustainable Energy Conference

    Salam,A.,Hibaoui,A.E.:Comparisonofmachinelearningalgorithmsforthepower consumption prediction: Case study of Tetouan city. In: 6th International Renew- able and Sustainable Energy Conference. pp. 1–5. IEEE, Rabat, Morocco (2018)

    2018

  20. [28]

    In: Proceedings of the 25th international con- ference on Machine learning

    Sorokina, D., Caruana, R., Riedewald, M., Fink, D.: Detecting statistical interac- tions with additive groves of trees. In: Proceedings of the 25th international con- ference on Machine learning. pp. 1000–1007. ICML ’08, Association for Computing Machinery, New York, NY, USA (2008)

    2008

  21. [29]

    The Annals of Statistics13(2), 689–705 (1985)

    Stone, C.J.: Additive regression and other nonparametric models. The Annals of Statistics13(2), 689–705 (1985)

    1985

  22. [30]

    The Annals of Statistics14(2), 590–606 (1986)

    Stone, C.J.: The dimensionality reduction principle for Generalized Additive Mod- els. The Annals of Statistics14(2), 590–606 (1986)

    1986

  23. [31]

    Wilson, D.J.: The harmonic mean p-value for combining dependent tests. Proceed- ings of the National Academy of Sciences of the United States of America116(4), 1195–1200 (2019) Sails: Surrogate-based Analysis of Interactions via Local Effect Smooths 17

    2019

  24. [32]

    Information Processing & Management60(4), 103373 (2023)

    Zhang,H.,Zhang,X.,Zhang,T.,Zhu,J.:Capturingtheformoffeatureinteractions in black-box models. Information Processing & Management60(4), 103373 (2023)

    2023

  25. [33]

    doubly center

    Zhang, X., Zhang, H., Zhu, J., Li, Z.: Revealing the structure of prediction mod- els through feature interaction detection. Knowledge-Based Systems236, 107737 (2022) 18 T. Heiß et al. A Appendix A.1 Theoretical Evidence..............................................19 Proof of...

    2022

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.