Pith. sign in

REVIEW 3 major objections 4 minor 40 references

The Quadrilateral Loss: Additivity as a Measurable Behavior of Dense Neural Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The quadrilateral residual — a four-point swap difference — vanishes exactly when a coordinate carries no interaction, making additivity measurable, trainable, and certifiable in dense neural networks.

desk verdict A useful interaction penalty and surrender-curve instrument, with an over-sold abstract and a real on-manifold gap. read the letter →

arxiv 2607.20201 v1 pith:4TDZN6FU submitted 2026-07-22 cs.LG cs.AI

Antonio Di Cecco This is my paper · ORCID
classification cs.LGcs.AI MSC 68T0762G08
keywords additivemodelsquadrilaterallossfeatureinteractionsneuralnetworksShapley-GAMinterpretabilityregularizationfinitedifferences
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

Additive models typically buy interpretability by forbidding feature interactions architecturally. This paper claims that additivity can instead be measured and controlled as a behavior of any dense network, using a four-corner finite difference called the quadrilateral residual: swap one coordinate between two inputs and form the alternating sum; on a product domain the residual is identically zero across all pairs exactly when the network is additive. Because the residual's expectation equals the total interaction mass of the interventional Shapley-GAM decomposition involving that coordinate, penalizing its square gives a differentiable dial over interaction strength, and a near-zero value doubles as a certificate that ordinary SHAP explanations coincide with shape functions. The empirical sweep shows the trade-off is strongly asymmetric — most learned interactions are removable at small accuracy cost — and per-feature surrender curves reveal that pre-regularization interaction magnitude is a weak, seed-dependent predictor of what survives. Comparing routes to exact additivity, the paper finds that behavior-first regularization can be crystallized into exact structure and that several independent procedures converge to the same shape functions.

What carries the argument

The quadrilateral residual of Eq. (1) is the second-order mixed difference of f along coordinate i against all remaining coordinates, evaluated at the four corners of an axis-aligned quadrilateral; a single-coordinate swap between two points yields the four terms. On product domains its vanishing across all pairs characterizes additivity exactly. Its expectation with the partner point drawn from the data distribution equals the sum of all interventional Shapley-GAM interaction components that contain coordinate i, which gives the penalty its meaning as a coordinate-wise interaction-mass term. The same residual serves as diagnostic Q², as training penalty, and as the criterion that turns one-

What would settle it

On a small synthetic function with known Shapley-GAM components, compute E[Q_i] and E[Q_i²] per feature and rank both against the true interaction masses; if the squared statistic's ranking diverges from the expectation's on any feature, the loss is not measuring interaction mass as advertised.

Watch

Extended reading notes

Core claim

The paper's central claim is Proposition 1: for a product domain and any function f, the quadrilateral residual Q_i(x, x̃) = f(x) − f(x^(i)) − f(x̃^(i)) + f(x̃) is zero for all pairs and all coordinates if and only if f is additive. Tying this statistic to the interventional Shapley-GAM, the expectation of Q_i over a data-distributed partner point equals the sum of all interaction components containing coordinate i, so the training penalty in Eq. (2) — task loss plus λ times the average squared residual — penalizes interaction mass coordinate by coordinate. This makes the loss meaningful for piecewise-linear networks, where mixed partial derivatives vanish almost everywhere, and cheap to eva

Load-bearing premise

The exact additivity characterization holds for all pairs on a product domain, but the training penalty samples swap partners from the observed data, so a low Q² certifies near-additivity only where the data live; separately, the equality to Shapley-GAM interaction mass is proven for E[Q_i], not for the squared statistic the loss actually penalizes.

Editorial extensions

If this is right

  • A dense network trained with the quadrilateral loss can operate at a chosen near-additivity point, with the interaction residual measured on the same scale as the task error; λ is dimensionless and can be set by a balance heuristic rather than per-dataset tuning.
  • Near-zero Q² implies the interventional Shapley-GAM of the model is first order, so ordinary SHAP values coincide with the model's shape functions even under dependent features — making the loss a faithfulness dial for explanations.
  • Surrender curves provide an online, per-feature interaction monitor during training; post-hoc interaction magnitude rankings are unreliable guides to which interactions a regularized model will keep.
  • Behavior-first regularization can be crystallized into an exactly additive masked network at no accuracy cost relative to a from-scratch structurally additive model, and it dominates weight-space constraints like winner-take-all decay.
  • On small datasets the additive class can be more accurate than the dense network, so the penalty can act as a regularizer that improves accuracy and additivity simultaneously, and every exact-additivity route beats the dense baseline on the clinical cohorts.

Reading between the lines

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

  • Because the proof equates E[Q_i], not E[Q_i²], to Shapley-GAM interaction mass, a natural follow-up would be a loss on |E[Q_i]| or a debiased estimator; whether the squared surrogate changes the selected function class is testable on synthetic data.
  • Restricting the sum in (2) to swaps across prescribed feature blocks would turn the dial into an interaction-order prior, potentially replacing explicit pairwise screening in GA2M-style models; the paper names this variant but does not run it.
  • The surrender-curve finding suggests a stronger, testable claim: if stubbornness is a stable dataset-level property, an early training probe could predict final retained interactions on a single run; the paper's weak pooled correlation implies the opposite, so the question is open.
  • The shared-section model's advantage is attributed to implicit basis sharing, but no ablation isolates it; comparing shared versus disjoint sections under matched capacity would settle whether the benefit is architectural or purely capacity-related.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes the quadrilateral loss, a training penalty based on the second-order mixed difference Qi(x,x̃) = f(x) − f(x⁽ⁱ⁾) − f(x̃⁽ⁱ⁾) + f(x̃). Proposition 1 shows that on a product domain Qi vanishes for all pairs and all coordinates iff f is additive. Equation (2) penalizes the empirical second moment of Qi, with x̃ obtained by minibatch permutation, and the paper claims that this statistic equals, in expectation, the per-coordinate interaction mass of the interventional Shapley-GAM, making the loss a dial for additivity and a certificate for faithful SHAP explanations. The paper also contributes surrender curves (per-feature interaction tracked during training), a λ-balance heuristic, a crystallization procedure that converts a penalized dense network into an exactly additive masked model, a shared-section read-out that is exactly additive, and a six-dataset comparison of routes to additivity (NAM, backfitting, TEAM, weight decay, etc.). The empirical sections are multi-seed and the paper explicitly corrects some of its own earlier single-seed conclusions.

Significance. If the two load-bearing claims — that the loss (2) penalizes the Shapley-GAM interaction mass, and that a low value of (2) provides an off-distribution certificate of additivity and SHAP faithfulness — were established as stated, this would be a practically useful instrument. The finite-difference statistic is simple, differentiable, piecewise-linear-friendly, and four forward passes are cheap; the surrender-curve analysis is a genuinely useful diagnostic; and the route comparison is carefully executed with a reproducible public implementation. The paper is also unusually honest about its limitations, including the on-manifold penalty in Section 10 and the identifiability issues in Section 8. However, as written, two central claims need either derivation or substantial qualification before the results can be accepted as stated.

major comments (3)
  1. [§3, Eq. (2), §10, Abstract] The population characterization (Proposition 1) requires Qi=0 for all pairs of inputs in a product domain. The training objective (2), however, samples x̃ from the data distribution via minibatch permutation, so the penalty is enforced only on the empirical data manifold. Section 10 states this explicitly: 'the network may retain interactions in regions unsupported by data' and 'no claim is made about interactions off-distribution.' Yet the abstract and contribution (i) claim that driving the loss down 'certifies the faithfulness of ordinary SHAP explanations' and that near-additivity is guaranteed, without the on-manifold qualification. Section 5.3's shape-function plots at extreme standardized values are exactly in the unenforced regions. This is load-bearing: a user who interprets low Q² as proof of additivity at an arbitrary query point is applying a guarantee whose preconditions are
  2. [§3, paragraph 'Connection to the Shapley-GAM'] The paper derives E_x̃[Qi(x,x̃)] = Σ_{L∋i, |L|≥2} f_L(x_L), i.e., the first moment of Qi equals the sum of the Shapley-GAM interaction components containing i. It then says: 'The quadrilateral loss therefore penalizes the ℓ2 mass of the Shapley-GAM's interaction components, coordinate by coordinate.' But the loss (2) penalizes E[(Qi)²], not E[Qi]. The squared statistic satisfies E[Qi²] = E[(E[Qi|x])²] + E[Var(Qi|x)], which contains the variance of Qi as well as the squared conditional interaction mass, and the square of an expectation is not the ℓ2 norm of the corresponding components (cross terms appear). At nonzero λ, minimizing E[Qi²] is therefore not the same as minimizing the claimed interaction mass; the exact-zero equivalence is a trivial consequence of Proposition 1 and does not justify the surrogate at finite λ. This is load-bearing for the paper's central interpretation of the
  3. [§3, last paragraph; §5.3] The SHAP-faithfulness statement is made as a qualitative implication of the Bordt–von Luxburg recovery theorem: 'once a model is (near-)additive its ordinary interventional SHAP values coincide with its shape functions.' The paper does not state a quantitative version: how small must Q² be, and in what norm, for the SHAP values to be close to the shape functions? Without such a bound, the certificate is only 'drive Q² to zero exactly,' which the stochastic estimator cannot do (the paper itself notes the ~10⁻⁴ noise floor in Section 5.1). Given the paper's emphasis on measured near-additivity as a contract, a quantitative statement (even conservative) would be needed to support the claim that a low but nonzero Q² certifies SHAP faithfulness.
minor comments (4)
  1. [Throughout] Typesetting issues: 'T able 1' and 'T able 2' in the text, 'Q2/Q2dense' in Table 3's header, and 'ℓ2' should be 'ℓ²' consistently. The subscript in 'train = 404' (Section 7) is missing in the text.
  2. [§5.2, Table 1] The table lists a 'Rank at convergence' column but the ranking is discussed with caveats about transient spikes. A supplementary table separating the smoothed-tail rank from the peak rank would make the point cleaner.
  3. [§6.3, Eq. (3)] The notation gθ(xi, x̄−i) is clear, but it would help to define explicitly that x̄ is the vector of training-means and that the section is evaluated with all other coordinates held at x̄.
  4. [§8] The agreement figure is described as showing 'five estimates of an under-determined quantity' under concurvity. This is an important caveat and should appear in the main text before the strong sentence 'a single solution reached from five directions,' not only as a subsequent diagnosis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central theorem and loss derivation are self-contained; self-citations are non-load-bearing baselines; acknowledged limitations narrow but do not invert the derivational chain.

full rationale

The paper's central claim—that the quadrilateral residual Qi (Eq. 1) characterizes additivity (Proposition 1)—is a proven population theorem on product domains, independent of the training loss. The loss (Eq. 2) penalizes a Monte Carlo second moment of the same residual, and the Shapley-GAM connection relies on the external Möbius/Bordt–von-Luxburg theorem, not on the authors' own prior work. The paper's self-citations (TEAM, book chapters) are used only as comparison baselines or descriptions of the author's earlier software, never as the load-bearing justification for the main claim. The two genuine shortcomings—E[Qi] vs. E[Qi²] and the on-manifold enforcement of the penalty—are explicitly acknowledged in Section 10 and elsewhere; they are correctness/scope limitations, not cases where a prediction reduces by construction to a fitted input or where an ansatz is smuggled in via self-citation. The balance-heuristic 'prediction' for Bike is an out-of-sample check, not a fit to the predicted quantity. No circular step can be exhibited from the paper's equations. Therefore the derivation chain is self-contained against external benchmarks and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are postulated; the 'shared-section model' and 'crystallization' are procedures, not entities. The free parameters are hyperparameters of the training protocol; the central math relies on standard product-domain assumptions and cited theorems.

free parameters (3)
  • λ (quadrilateral penalty coefficient) = λ=1 default; λ*=0.06–2.6 (tuned); λbal=0.18–4.4; λ=20 for crystallization
    Controls trade-off between task loss and interaction penalty; chosen by balance heuristic or validation, not derived. Sections 5.1, 7.
  • |S| (coordinates sampled per step) = 4
    Hand-picked subset size in Eq. (2) to limit cost; not justified by theory.
  • Warmup and annealing schedule = 15-epoch warmup, linear anneal
    Training-protocol choice; affects surrender-curve readings and final interaction residuals.
assumptions (4)
  • domain assumption X is a product domain X1×...×Xd, and Proposition 1 quantifies over all x,x̃
    The iff characterization holds for all pairs, but training samples only data pairs.
  • domain assumption Bordt & von Luxburg's Möbius transform and recovery theorem for the interventional SHAP value function
    The Shapley-GAM equivalence and SHAP faithfulness claim rest on this cited theorem (ref [33]).
  • domain assumption The background distribution for do-interventions is the empirical data distribution D
    E[Qi] is computed over x̃∼D; the value function v(x,S)=E_z[f(z)|do(x_S)] uses D; under dependent features this is a modeling choice, not a guarantee.
  • domain assumption Piecewise-linear interactions live in gradient discontinuities, so finite differences are informative
    Motivates the penalty for ReLU networks; no formal theorem is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Quadrilateral Loss: Additivity as a Measurable Behavior of Dense Neural Networks." pith.science (2026). https://pith.science/paper/4TDZN6FU

@misc{pith2026260720201,
  author       = {Pith},
  title        = {Pith review of: The Quadrilateral Loss: Additivity as a Measurable Behavior of Dense Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TDZN6FU}},
  note         = {Machine review of arXiv:2607.20201}
}
read the original abstract

Additive models buy interpretability by forbidding feature interactions, a constraint that neural instantiations enforce architecturally. We introduce the quadrilateral loss, a differentiable penalty that treats additivity as a measurable behavior instead: a second-order mixed difference on pairs of training points swapping one coordinate, which vanishes if and only if the coordinate carries no interaction, remains informative for piecewise-linear networks, and equals in expectation the per-coordinate interaction mass of the interventional Shapley-GAM. The loss turns additivity into a dial - most learned interactions prove removable almost for free, and on small datasets a moderate penalty improves accuracy and additivity simultaneously - and into an online observable: its per-feature surrender curves show, across seeds and datasets, that pre-regularization interaction magnitude barely predicts what a regularized model retains, undermining post-hoc interaction rankings. Against this instrument we compare routes to exact additivity, spanning structural masks, behavioral penalties (optionally crystallized into exact structure), weight decay, backfitting, the shared-section model, and bagged boosted stumps: constraining behavior before structure dominates weight-space constraints, rankings reverse between data regimes, and converging routes agree on the shape functions themselves. Three silent failure modes we document share one anatomy: guarantees imported into settings that quietly void their preconditions.

Figures

Figures reproduced from arXiv: 2607.20201 by the authors.

Figure 1
Figure 1. Ontology of the routes compared, organized by where the additivity constraint acts: on the structure (the architecture cannot express interactions), on the behavior (a dense model is penalized, then optionally crystallized), on the fitting algorithm (components are estimated additively), or on the read-out (an unconstrained network is queried through its one-feature sections). The quadrilateral loss additionally ser… view at source ↗
Figure 2
Figure 2. Accuracy–interaction trade-off under the quadrilateral penalty [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Win–win regime of the quadrilateral penalty [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Seed replication of the magnitude–stubbornness relation. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Surrender curves on four datasets (three most stubborn features highlighted per panel; shaded band = penalty annealing). Panel titles report ρs between warmup magnitude and reduction factor; panel curves are shown for seed 0; the seed-replicated statistics are in [PIT…
Figure 6
Figure 6. Figure 6: One-dimensional response sections as a function of [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Agreement and convergence in one picture [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 2 linked inside Pith

  1. [1]

    Hastie and R

    T. Hastie and R. Tibshirani. Generalized additive models.Statistical Science, 1(3):297–310, 1986

    1986

  2. [2]

    Agarwal, L

    R. Agarwal, L. Melnick, N. Frosst, X. Zhang, B. Lengerich, R. Caruana, and G. Hinton. Neural additive models: Interpretable machine learning with neural nets. InAdvances in Neural Information Processing Systems, 2021

    2021

  3. [3]

    Y. Lou, R. Caruana, and J. Gehrke. Intelligible models for classification and regression. In KDD, 2012

    2012

  4. [4]

    Y. Lou, R. Caruana, J. Gehrke, and G. Hooker. Accurate intelligible models with pairwise interactions. InKDD, 2013

    2013

  5. [5]

    Chang, R

    C.-H. Chang, R. Caruana, and A. Goldenberg. NODE-GAM: Neural generalized additive model for interpretable deep learning. InInternational Conference on Learning Representa- tions, 2022

    2022

  6. [6]

    Radenovic, A

    F. Radenovic, A. Dubey, and D. Mahajan. Neural basis models for interpretability. InAdvances in Neural Information Processing Systems, 2022

    2022

  7. [7]

    I. M. Sobol’. Global sensitivity indices for nonlinear mathematical models and their Monte Carlo estimates.Mathematics and Computers in Simulation, 55(1–3):271–280, 2001

    2001

  8. [8]

    S. M. Lundberg and S.-I. Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, 2017

    2017

Show all 40 references
  1. [9]

    J. D. Janizek, P. Sturmfels, and S.-I. Lee. Explaining explanations: Axiomatic feature inter- actions for deep networks.Journal of Machine Learning Research, 22(104):1–54, 2021

    2021

  2. [10]

    Tsang, D

    M. Tsang, D. Cheng, and Y. Liu. Detecting statistical interactions from neural network weights. InInternational Conference on Learning Representations, 2018

    2018

  3. [11]

    Tsang, H

    M. Tsang, H. Liu, S. Purushotham, P. Murali, and Y. Liu. Neural interaction transparency (NIT): Disentangling learned interactions for improved interpretability. InAdvances in Neural Information Processing Systems, 2018

    2018

  4. [12]

    G. Hooker. Generalized functional ANOV A diagnostics for high-dimensional functions of de- pendent variables.Journal of Computational and Graphical Statistics, 16(3):709–732, 2007

    2007

  5. [13]

    Molchanov, S

    P. Molchanov, S. Tyree, T. Karras, T. Aila, and J. Kautz. Pruning convolutional neural net- works for resource efficient inference. InInternational Conference on Learning Representations, 2017

    2017

  6. [14]

    W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li. Learning structured sparsity in deep neural networks. InAdvances in Neural Information Processing Systems, 2016

    2016

  7. [15]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations, 2015

    2015

  8. [16]

    A. Buja, T. Hastie, and R. Tibshirani. Linear smoothers and additive models.The Annals of Statistics, 17(2):453–510, 1989. 23

    1989

  9. [17]

    J. H. Friedman. Greedy function approximation: A gradient boosting machine.The Annals of Statistics, 29(5):1189–1232, 2001

    2001

  10. [18]

    Grinsztajn, E

    L. Grinsztajn, E. Oyallon, and G. Varoquaux. Why do tree-based models still outperform deep learning on typical tabular data? InAdvances in Neural Information Processing Systems, 2022

    2022

  11. [19]

    Di Cecco

    A. Di Cecco. TEAM: Tree ensemble additive models. GitHub repository,https://github. com/AntonioDiCecco/TEAM, 2024

    2024

  12. [20]

    Di Cecco and L

    A. Di Cecco and L. Gianfagna.Explainable AI with Python. Springer, Cham, 2025. ISBN 978-3-031-92229-9

    2025

  13. [21]

    Di Cecco and L

    A. Di Cecco and L. Gianfagna. Additive models for interpretability. InExplainable AI with Python, chapter 6. Springer, Cham, 2025

    2025

  14. [22]

    H. Nori, S. Jenkins, P. Koch, and R. Caruana. InterpretML: A unified framework for machine learning interpretability.arXiv preprint arXiv:1909.09223, 2019

    1909 arXiv

  15. [23]

    Dubey, F

    A. Dubey, F. Radenovic, and D. Mahajan. Scalable interpretability via polynomials. In Advances in Neural Information Processing Systems, 2022

    2022

  16. [24]

    Z. Yang, A. Zhang, and A. Sudjianto. GAMI-Net: An explainable neural network based on generalized additive models with structured interactions.Pattern Recognition, 120:108192, 2021

    2021

  17. [25]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljaˇ ci´ c, T. Y. Hou, and M. Tegmark. KAN: Kolmogorov–Arnold networks.arXiv preprint arXiv:2404.19756, 2024

    2024 arXiv

  18. [26]

    Siems, K

    J. Siems, K. Ditschuneit, W. Ripken, A. Lindborg, M. Schambach, J. S. Otterbach, and M. Genzel. Curve your enthusiasm: Concurvity regularization in differentiable generalized additive models. InAdvances in Neural Information Processing Systems, 2023

    2023

  19. [27]

    V. I. Arnold. Proof of a theorem of A. N. Kolmogorov on the preservation of conditionally pe- riodic motions under a small perturbation of the Hamiltonian.Russian Mathematical Surveys, 18(5):9–36, 1963

    1963

  20. [28]

    A. S. Ross, M. C. Hughes, and F. Doshi-Velez. Right for the right reasons: Training dif- ferentiable models by constraining their explanations. InInternational Joint Conference on Artificial Intelligence, 2017

    2017

  21. [29]

    Erion, J

    G. Erion, J. D. Janizek, P. Sturmfels, S. M. Lundberg, and S.-I. Lee. Improving performance of deep learning models with axiomatic attribution priors and expected gradients.Nature Machine Intelligence, 3(7):620–631, 2021

    2021

  22. [30]

    Rieger, C

    L. Rieger, C. Singh, W. J. Murdoch, and B. Yu. Interpretations are useful: Penalizing explana- tions to align neural networks with prior knowledge. InInternational Conference on Machine Learning, 2020

    2020

  23. [31]

    D. W. Apley and J. Zhu. Visualizing the effects of predictor variables in black box supervised learning models.Journal of the Royal Statistical Society: Series B, 82(4):1059–1086, 2020

    2020

  24. [32]

    Lengerich, S

    B. Lengerich, S. Tan, C.-H. Chang, G. Hooker, and R. Caruana. Purifying interaction effects with the functional ANOV A: An efficient algorithm for recovering identifiable additive models. InInternational Conference on Artificial Intelligence and Statistics, 2020. 24

    2020

  25. [33]

    Bordt and U

    S. Bordt and U. von Luxburg. From Shapley values to generalized additive models and back. InInternational Conference on Artificial Intelligence and Statistics, pages 709–745. PMLR, 2023

    2023

  26. [34]

    A. N. Kolmogorov. On the representation of continuous functions of many variables by super- position of continuous functions of one variable and addition.Doklady Akademii Nauk SSSR, 114:953–956, 1957

    1957

  27. [35]

    A. G. Vitushkin. On Hilbert’s thirteenth problem.Doklady Akademii Nauk SSSR, 95:701–704, 1954

    1954

  28. [36]

    Caruana, Y

    R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. InACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2015

    2015

  29. [37]

    Bouchiat, A

    K. Bouchiat, A. Immer, H. Y` eche, G. R¨ atsch, and V. Fortuin. Improving neural additive models with Bayesian principles. InInternational Conference on Machine Learning, pages 4416–4443. PMLR, 2024

    2024

  30. [38]

    J. H. Friedman and B. E. Popescu. Predictive learning via rule ensembles.The Annals of Applied Statistics, 2(3):916–954, 2008

    2008

  31. [39]

    Sorokina, R

    D. Sorokina, R. Caruana, M. Riedewald, and D. Fink. Detecting statistical interactions with additive groves of trees. InInternational Conference on Machine Learning, 2008

    2008

  32. [40]

    R. K. Pace and R. Barry. Sparse spatial autoregressions.Statistics & Probability Letters, 33(3):291–297, 1997. 25

    1997

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.