{"id":"e4cf3ebb-0d95-49c0-822f-302d172233ba","arxiv_id":"2506.08338","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The midr package builds a global additive surrogate for any black-box model by least-squares projection with strict centering constraints, yielding interpretable main effects, interactions, and SHAP-style attributions.","lead":"This paper introduces midr, an R package that approximates a black-box model's predictions with a simple additive formula, splitting them into main effects and pairwise interactions. The package is meant to make machine learning models easier to understand, audit, and explain in practice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'most accurate' claim is about the population MID in Eq. (30), but the package solves the finite-basis, grid-binned problem in Eqs.","rationale":"The reader's weakest assumption targeted the finite basis and grid centering; I agree with that diagnosis and focus it further on the second-order constraint matrix. In the population setting, the projection argument supports the paper: the set of all order-m additive functions equals the set of sums of strictly centered effects, so the theoretical MID is indeed the best order-m additive model. The gap is the leap from Eq. (30) to the package solver. The Appendix's M'_{pq} enforces bin-averaged conditions, not pointwise conditional means, and the paper provides no error analysis. This is not a disagreement with the intended method; it is a missing link between the stated definition and the shipped implementation. The concern is addressable by deriving the KKT conditions or by a convergence experiment, so the reader's CONDITIONAL verdict is appropriate and I would not change it.","tokens_in":19357,"tokens_out":14498,"duration_ms":192588,"concrete_test":"Independently derive the KKT system for the constrained least-squares problem (65) with the second-order constraint matrix M'_{pq} in (63) and compare it with the population strict-centering condition E[f_{pq}(X_p,X_q)|X_p=x_p]=0 from Eq. (41). If M'beta'=0 does not imply zero conditional mean at every grid point for piecewise-linear encoding, the package is not computing the MID defined in Section 3 and the Section 6 accuracy claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim in Section 6 asserts that the MID model of order m is the most accurate additive representation of order m. This is true for the population minimizer in Eq. (30), because any order-m additive function can be re-expressed with strictly centered components. The package, however, solves the finite-dimensional problem (65), where the strict centering condition (34) is replaced by bin-level constraints. For continuous features, the text says equality x_{iJ'}=v_{J'} is relaxed using 'weights distributed over a discrete grid,' but the Appendix's second-order constraint blocks M'_{pq} in (63) only enforce sum_i chi^a_p(x_{ip}) f_{pq}(x_{ip}, x_{iq}) = 0 for each encoding bin a. This is a necessary but not sufficient condition for E[f_{pq}(X_p,X_q)|X_p=x_p]=0, so the feasible set differs from the defined MID class. No grid-placement rule, consistency theorem, or approximation-error bound is given. Without such a result, the representation returned by interpret() is not shown to be the most accurate order-m additive representation, and the attribution guarantees in Eq. (41) need not hold for the fitted object.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents the R package midr, which implements Maximum Interpretation Decomposition (MID), a functional-decomposition method that approximates a black-box prediction function f by a low-order additive function minimizing expected squared error subject to strict centering of all effects. The manuscript reviews related work on PDP, ALE, SHAP, and ICE; defines MID and its numerical estimation by constrained least squares with piecewise-constant or piecewise-linear basis functions; and demonstrates the package on a Friedman regression benchmark, including main-effect and interaction plots, importance measures, ICE/breakdown/SHAP visualizations, a pragmatic-stability simulation with correlated features, and runtime comparisons for least-squares solvers.","tokens_in":19577,"tokens_out":7900,"duration_ms":94357,"significance":"If the theoretical and numerical claims hold, MID is an attractive global surrogate: it combines least-squares optimality with exactly centered components, which yields a closed-form SHAP decomposition (44) and a unified local-global interpretation. The paper's strengths are its clear connection to existing IML methods, a complete and reproducible R implementation with extensive code examples, and a useful numerical comparison of linear algebra backends. The central limitation is that the package solves a finite-dimensional grid-based version of the problem, and the manuscript does not supply the approximation theory needed to transfer the attractive population-level properties to the fitted object.","major_comments":[{"comment":"Existence and uniqueness of the MID minimizer are asserted without proof: the text states that the decomposition is 'guaranteed to exist' and that uniqueness is obtained by selecting the minimum-norm solution among candidates, but the only reference is Iwasawa and Matsumori (2025), which is listed as 'Forthcoming'. Because the paper's headline optimality claim and the solver's tie-breaking procedure both depend on these properties, this is a load-bearing point. Please include a proof or a precise statement of the conditions under which existence and uniqueness hold, or clearly state them as assumptions imported from the forthcoming paper.","section":"Section 3, Definition, Eqs. (30)-(31)"},{"comment":"The empirical strict centering constraint (34) is not equivalent to the population constraint for continuous features. The appendix replaces E[f_pq(X_p,X_q)|X_p=x_p]=0 by the bin-level constraints sum_i chi^a_p(x_ip) f_pq(x_ip,x_iq)=0 for a=1,...,k_p, and symmetrically for q. These constraints are necessary but not sufficient for the population conditional expectation to vanish: a fitted interaction can satisfy every bin constraint while E[f_pq|X_p=x_p] is nonzero within a bin. The manuscript gives no grid-placement rule, no consistency theorem, and no approximation-error bound. Consequently, the object returned by interpret() is not shown to be the most accurate order-m additive representation, and the attribution property (41), used for breakdown plots and SHAP values, need not hold exactly for the fitted object. This gap should be addressed by a theorem connecting the grid solution to the population MID as n and k_j grow, or by clearly labeling the output as an approximate MID with appropriate diagnostics.","section":"Section 3, Numerical estimation; Appendix, Eq. (63)"},{"comment":"The closed-form MID-Shapley values and the 'most accurate' claim are valid for the population MID defined by Eq. (30) with exact strict centering. The package, however, solves the finite-dimensional problem (65), whose feasible set differs from the MID class, as explained in the previous comment. The sentence in Section 6 that 'the MID model of order m is the most accurate global additive representation' should therefore be restricted to the population quantity, and the finite-basis solution should be presented as an approximation whose error is not yet quantified. In the same vein, the characteristic function in Eq. (43) drops all J not contained in S only because of strict centering; this equality may fail for the bin-level solution, so the 'exact' MID-derived SHAP values require either a proof of exact centering or a revision to describe them as approximations.","section":"Section 3, Visualization; Section 6"},{"comment":"The tie-breaking criterion used in the solver is the weighted coefficient norm ||Delta^{1/2} beta||^2, where Delta is a diagonal matrix with entries sum_i chi^s_j(x_ij). This does not coincide with the criterion stated in Eq. (35), which minimizes sum_J sum_i f_J(x_iJ)^2. The two agree for the piecewise-constant encoding (50)-(51), where the basis functions have disjoint supports, but not for the piecewise-linear encoding (52), where X^T X has nonzero off-diagonal entries. Please clarify whether the weighted coefficient norm is an intentional solver choice; if so, it should be disclosed as a modification of Eq. (35), and if not, the solver should minimize the empirical norm actually specified there.","section":"Appendix, Eqs. (57)-(59)"}],"minor_comments":[{"comment":"The abstract and Section 4 mention classification models, but all demonstrations in the paper are regression examples; either add a small classification example or make the wording explicitly refer users to the vignettes for classification.","section":"Section 4 and Abstract"},{"comment":"The caption contains the typo 'newral network models'; it should read 'neural network models'.","section":"Figure 9 caption"},{"comment":"The bibliography entry for the package lists 'R. Asasihba', but the author name in the paper is 'Ryoichi Asashiba'; the spelling should be corrected.","section":"Reference list"},{"comment":"The boundary encoding functions are defined using ratios such as (x_j - x_j^(0))/(x_j^(1) - x_j^(0)) with x_j^(0) = -infinity, and the text says the ratio is taken to be 1; this should be phrased as a limit rather than an ordinary ratio.","section":"Appendix, Eq. (52)"},{"comment":"The runtime benchmark reports average evaluation times but does not state the number of replicates or provide standard errors; adding this information would make the comparison fully reproducible.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's theoretical core is deferred to a self-cited 'Forthcoming' manuscript (Iwasawa and Matsumori, 2025). If that manuscript is not available to the editorial process, the editor may wish to request it, because the existence, uniqueness, and strict-centering properties of MID are load-bearing for the package's advertised guarantees. The package itself appears well structured and useful; the revision should focus on closing the gap between the population MID definition and the finite-basis grid implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read for you. The paper presents midr, an R package implementing Maximum Interpretation Decomposition (MID). MID is a clean idea: project a black-box prediction function onto a low-order additive model with strict centering constraints. The package does this with a least-squares fit and gives exact SHAP values for the fitted surrogate. That last bit is useful — the closed-form Shapley identity for strictly centered additive decompositions is correct, and the package makes it convenient.\n\nWhat's actually new is the implementation and the numerical illustrations. The Friedman example shows the surrogate tracks the true function well, and the pragmatic-stability simulation is a useful point: on correlated features, MID behaves like ALE rather than PDP, which is a real advantage. The computational benchmark is fine for a software paper. The writing is clear, and the authors explicitly flag the overfitting risk when assessing the surrogate on training data.\n\nThe soft spots are real but not fatal. First, the theoretical guarantees (existence, uniqueness, the optimality of MID) are not proven here; they're deferred to a self-cited forthcoming paper. That's acceptable for a software paper only if the claims are non-essential, but here the headline 'most accurate' claim is load-bearing. Second, and more troubling, the implementation of strict centering for continuous features is underspecified. The text says values are 'represented as a set of weights distributed over a discrete grid,' but the Appendix gives only the empirical constraints in the encoding basis. As far as I can tell from what's written, those constraints enforce that sums over observations in each bin vanish, not that the conditional expectation E[f_pq(X_p,X_q)|X_p=x_p] is zero for each x_p. So the fitted object is not necessarily a member of the population MID class, and the attribution guarantees in Eq. (41) and the SHAP formula hold exactly only for the population minimizer, not for the object interpret() returns. The paper doesn't provide a grid-placement rule or any approximation-error analysis. That gap should be closed, or the claims softened.\n\nThe empirical sections also lack error bars and full code/data for reproduction, though the package is on CRAN, which helps.\n\nBottom line: this is a useful package and a reasonable software paper, but the central optimality claim currently applies to the definition, not the implementation. I'd send it to a serious referee with a request for a technical appendix on the continuous-feature centering approximation and a consistency argument. The paper deserves review, not a desk reject.","headline":"midr is a workmanlike package for a reasonable idea, but the fitted object may not satisfy the strict centering that powers the paper's headline claims.","tokens_in":20120,"tokens_out":5434,"would_cite":true,"duration_ms":54737,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The midr package builds the most accurate low-order additive surrogate of any black-box prediction function, with strict centering so every effect can be read as a contribution.","keywords":["Maximum Interpretation Decomposition","functional decomposition","interpretable machine learning","global surrogate model","model-agnostic explanation","strict centering","SHAP values","R package"],"falsifier":"On a synthetic model with correlated continuous features and a known closed-form MID, run midr at increasing grid resolutions; if the fitted surrogate's uninterpreted variation ratio does not converge to the theoretical optimum, or if some other strictly centered additive order-2 fit attains lower squared error on a large held-out sample, the implementation does not realize the paper's central claim.","tokens_in":19142,"feed_emoji":"🧩","tokens_out":10150,"duration_ms":112104,"temperature":0.7,"pith_summary":"This paper introduces midr, an R package that implements Maximum Interpretation Decomposition (MID), a way to open up a black-box prediction function by replacing it with a low-order additive surrogate. The central claim is that among all additive representations that use terms up to a chosen order, the MID model of that order is the most accurate in squared error, and a strict centering constraint gives each extracted main effect and interaction a clean meaning as a contribution. The payoff for a curious reader is a single model-agnostic tool that covers both global summaries, such as effect plots and importance rankings, and local explanations, such as prediction breakdowns and SHAP-style attributions, without falling apart when features are correlated. The package makes the method usable in practice and supplies visualizations for all of these views.","feed_headline":"Black-box models break into their best additive parts","feed_subtitle":"midr's MID surrogate minimizes squared error while keeping every main effect and interaction readable.","key_machinery":"The machinery has three parts. The first is the functional decomposition itself: effects $f_J$ indexed by feature subsets $J$, with order $|J|$, whose sum up to order $m$ defines the surrogate $F_m$. The second is the optimization that defines MID: minimize the uninterpreted variation ratio $U(F_m)$ subject to strict centering, $\\mathbb{E}[f_J(X_J)\\mid X_{J'}]=0$ for all proper $J'\\subset J$, and zero higher-order effects, with uniqueness restored by minimizing the expected squared norm among all minimizers. The third is the computational engine: each effect is expanded in a piecewise constant or piecewise linear encoding basis with $k_j$ knots per feature, centering is imposed as linear constraints on a discrete grid for continuous features, and the whole problem becomes a constrained least squares problem solved with a choice of fast linear algebra methods. This same machinery yields a closed-form expression for Shapley values of the surrogate, so local attribution is a byproduct rather than a separate expensive computation.","core_discovery":"Maximum Interpretation Decomposition is a functional decomposition: the fitted prediction function $f$ is approximated by $\\sum_{|J|\\le m} f_J(x_J)$, an intercept plus main effects and interaction effects up to order $m$. The defining claim is that among all additive representations built from terms of order at most $m$ that also satisfy strict centering, meaning $\\mathbb{E}[f_J(X_J)\\mid X_{J'}]=0$ for every proper subset $J'\\subset J$, the MID model of order $m$ minimizes the expected squared error, measured by the uninterpreted variation ratio $U(F_m)=\\mathbb{E}[(f(X)-F_m(X))^2]/\\mathrm{Var}(f(X))$. Strict centering is what makes each effect readable as a standalone contribution, and the same structure yields a closed-form Shapley attribution, $\\phi^{ms}_{ij}=\\sum_{J\\ni j, |J|\\le m}\\frac{1}{|J|}f^{\\mathrm{mid}}_J(x_{iJ})$. The package realizes this by encoding each feature in piecewise constant or piecewise linear basis functions, imposing centering as linear constraints on a discrete grid for continuous features, and solving a constrained least squares problem, with a minimum-norm rule restoring uniqueness when the solution is not unique. A regression-benchmark demonstration shows a near-zero uninterpreted variation ratio, and a correlated-features experiment shows MID main-effect plots tracking the true function instead of following the extrapolation artifacts that affect partial dependence plots.","pith_inferences":["An implication the paper leaves implicit: the optimality claim is about the population problem, and the empirical solution depends on the chosen encoding basis and grid; finer grids should converge to the population MID, but the paper does not quantify that convergence.","Another extension: the same formulation could be pushed to higher-order interactions or made sparse, selecting only the feature pairs that most reduce the uninterpreted variation ratio, which would help when many interactions are negligible.","The 'most accurate' claim means among additive representations of the same order; if the black box is strongly non-additive, even the best order-2 MID will miss structure, so the package's diagnostic ratio, not the decomposition alone, should decide whether the surrogate is trustworthy.","Because the closed-form Shapley values are Shapley values of the surrogate rather than of the original black box, their faithfulness to the original prediction is bounded by the surrogate's accuracy. This point deserves to be stated explicitly."],"forward_implications":["Any black-box model can be summarized by main-effect and interaction plots that are, by construction, the best low-order additive approximation in squared error.","The same surrogate supports local explanations: per-observation prediction breakdowns, conditional curves, and SHAP values computed exactly in closed form rather than by sampling.","For correlated features, MID avoids the extrapolation distortion that affects partial dependence plots, because centering is imposed on observed conditional structure rather than on marginal averages.","Users can check surrogate quality by the uninterpreted variation ratio on held-out predictions, so the method carries its own accuracy diagnostic.","The package's computation cost is dominated by the constrained least squares solve and grows with the number of parameters, with faster solvers reducing runtime substantially."],"supporting_citations":[{"why":"Defines MID and its optimality of order-m among additive representations; this paper implements it.","marker":"Iwasawa and Matsumori (2025)"},{"why":"Source of the correlated-features simulation and the ALE comparison used to show pragmatic stability.","marker":"Apley and Zhu (2020)"},{"why":"Provides the Friedman 1 regression benchmark on which the package demonstration is built.","marker":"Friedman (1991)"},{"why":"Supplies the additive feature attribution and Shapley value framework that yields the closed-form MID-derived attributions.","marker":"Lundberg and Lee (2017)"},{"why":"Motivates low-order additive representations with pairwise interactions, the class MID optimizes over.","marker":"Lou et al. (2013)"},{"why":"Supports the paper's statement that additive representations of predictive models perform well across interpretation tasks.","marker":"Tan et al. (2023)"},{"why":"Provides the fast linear algebra routines used for the constrained least squares solves.","marker":"Bates and Eddelbuettel (2013)"}],"fun_headline_variants":["MID decomposes black-box models into readable parts","Crack open any black-box model with MID","MID: parse black-box predictions into clean additive pieces","Learn from any model via maximum interpretation decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim depends on the estimated, grid-discretized version of the decomposition being a faithful stand-in for the mathematically defined optimum, even when features are continuous and correlated.","fun_headline_variants_meta":{"raw":{"variants":["MID decomposes black-box models into readable parts","Crack open any black-box model with MID","MID: parse black-box predictions into clean additive pieces","Learn from any model via maximum interpretation decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000796,"raw_usage":{"total_tokens":3516,"prompt_tokens":971,"completion_tokens":2545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":2498}},"tokens_in":587,"tokens_out":2545,"duration_ms":22178,"temperature":1.0,"reasoning_tokens":2498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:14:12.030555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic model with correlated continuous features and a known closed-form MID, run midr at increasing grid resolutions; if the fitted surrogate's uninterpreted variation ratio does not converge to the theoretical optimum, or if some other strictly centered additive order-2 fit attains lower squared error on a large held-out sample, the implementation does not realize the paper's central claim.","supporting_citations":[],"review_version":1}