{"id":"667e1462-383e-4166-9f10-ec49caf1c85f","arxiv_id":"2412.17152","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified framework classifies feature-based explanation methods by the distribution used to remove features (baseline, marginal, conditional) and by how much higher-order interaction they attribute.","lead":"This paper presents a framework that organizes many feature-attribution methods for machine learning models into a single map, using functional ANOVA decompositions and cooperative game theory. A practitioner can read off how a method handles feature dependencies and higher-order interactions before choosing one.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IG is catalogued as partial b-fANOVA, but its weights depend on Taylor monomial degree rather than subset size; an analytic counterexample shows IG differs from the Shapley partial effect, so the gradient-method unification is not exact.","rationale":"The paper's core framework, built from fANOVA decompositions and game-theoretic summaries, is sound; the definitions are precise and the proofs of Theorems 1-4 and Corollary 1 are correct. The Reader's identified weakness, the analyticity assumption in Theorem 1, is real, but it is not the most load-bearing gap. Even for analytic functions, Integrated Gradients does not equal the Shapley partial effect because IG weights individual Taylor monomials by kappa_i / sum(kappa), whereas the framework's partial effects weight an entire subset interaction f_S by 1/s. The x1^2*x2 example demonstrates this exactly: the b-fANOVA interaction effect is 1, the Shapley value splits it equally, and IG splits it as 2/3 and 1/3. Consequently, the claim that every catalogued method is characterized by the pair (imputation family, interaction family) is not exact for gradient-based methods such as IG and Integrated Hessians. If the authors intend 'partial' only as a loose qualitative category meaning 'some higher-order interactions are included', then the two-component claim is weakened because the precise degree-dependent weights are not specified by the pair. A conditional acceptance is appropriate: the framework and its exact results for perturbation-based methods stand, but the paper should either revise the categorization of IG and Integrated Hessians to 'approximate partial' with explicit weighting formulas, or extend the interaction-family axis to include degree-dependent weights. This is a fixable presentation issue, not a rejection of the underlying theory. The proposed counterexample provides a decisive, reproducible test of whether the exact partial-effect claim holds for IG.","tokens_in":41218,"tokens_out":13409,"duration_ms":133529,"concrete_test":"Implement the two-feature analytic counterexample F(x1,x2) = x1^2*x2 with baseline b = (0,0) and instance x0 = (1,1). Compute the local b-fANOVA game values nu(S) = F(x_S, b_{-S}), then compute both the Shapley value (partial effect) and Integrated Gradients. If IG returns (2/3, 1/3) while the Shapley value returns (1/2, 1/2), the categorization of IG as a partial b-fANOVA effect in Table 4 is not exact. Optionally repeat with a non-analytic model such as F(x) = ReLU(x1 + x2 - 0.5) to quantify the additional divergence caused by non-analyticity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires every catalogued method to be characterized by an imputation family and an interaction family (pure/partial/full), where 'partial' is defined in Section 3.4 as the Shapley value (SV) for individuals. For Integrated Gradients (IG), Appendix E.1.2, Eq. (5) gives phi_IG(i) = psi(i) + sum_{S containing i, s>1} sum_{kappa in Omega_S} [kappa_i / sum(kappa)] I(kappa). The weight on a b-fANOVA interaction f_S^{(b)} is not the uniform Shapley weight 1/s; it depends on the degree vector kappa inside that subset. Hence IG is not, in general, the partial b-fANOVA effect. This failure is independent of the analyticity issue raised by the Reader. Concrete counterexample: F(x1,x2) = x1^2*x2, baseline b = (0,0), instance x0 = (1,1). The local b-fANOVA game has nu(empty)=0, nu({1})=0, nu({2})=0, nu({1,2})=1, so the Shapley partial effect is (0.5, 0.5). Integrated Gradients gives (2/3, 1/3) because the single Taylor interaction x1^2*x2 has kappa=(2,1) and IG assigns feature 1 the weight 2/3. Thus Table 4's placement of IG under 'partial individual, b-fANOVA' is not exact. The same degree-dependence affects Integrated Hessians. The Limitations section states that some gradient methods summarize derivatives on a more fine-grained level than b-fANOVA effects, which is precisely this issue, but the summary tables and Figure 1 present the placement without this caveat. For non-analytic models such as ReLU networks the gap can be even larger, but the analytic counterexample is enough to show that the claimed exact characterization of gradient methods by the two components is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a unified framework for feature-based explanations by combining functional ANOVA (fANOVA) with cooperative game theory. Three fANOVA decompositions (baseline, marginal, conditional) capture the influence of the feature distribution, while three game-theoretic summary types (pure, partial, full) capture increasing influence of higher-order interactions. The framework is instantiated on local, sensitivity, and risk explanation games, and a catalog of existing XAI methods (SHAP variants, IG, DeepSHAP, SAGE, PFI, Sobol indices, etc.) is mapped to pairs (fANOVA type, interaction summary). The paper includes proofs of the main theorems in the appendix, a practitioner guide, and experiments on synthetic and real-world data.","tokens_in":41593,"tokens_out":2714,"duration_ms":27103,"significance":"If the categorization is accurate, the paper provides a valuable two-dimensional taxonomy that clarifies why explanation methods disagree: the choice of imputation distribution and the choice of higher-order interaction summary. The core mathematical machinery (Möbius inversion relating local games to fANOVA effects, Theorem 4's summary of the Möbius transform) is correct and well-presented. Concrete strengths are the machine-checkable-style appendix proofs, the explicit treatment of pure/partial/full effects in Tables 2 and 3, the reproducible code link, and the clear, falsifiable claim that each method is characterized by an imputation family and an interaction family. The empirical sections usefully illustrate the framework, especially the synthetic experiments varying feature correlation. The main weakness is that the catalog places Integrated Gradients and Integrated Hessians as exact instances of the partial b-fANOVA effect, but their weights depend on Taylor monomial degree rather than subset size, so that placement is not exact.","major_comments":[{"comment":"Integrated Gradients is catalogued as the partial individual b-fANOVA effect, but Eq. (5) shows that its weight on a b-fANOVA interaction I(κ) is kappa_i / sum(kappa), which depends on the degree vector, not the uniform Shapley weight 1/s. A concrete counterexample makes the gap concrete: for F(x1,x2)=x1^2*x2 with baseline b=(0,0) and instance x0=(1,1), the local b-fANOVA game has ν(empty)=0, ν({1})=0, ν({2})=0, ν({1,2})=1, so the Shapley partial effect is (0.5, 0.5). Integrated Gradients, however, assigns weights 2/3 and 1/3 because the only Taylor interaction has κ=(2,1). Thus IG is not the partial b-fANOVA effect, and Table 4's placement is inaccurate unless it is explicitly qualified as a degree-weighted generalization. The same issue affects Integrated Hessians, whose weights are product-weighted by degrees. The Limitations section (§7) states that some gradient methods summarize derivatives on a more fine-grained level, but the main text and Figure 1 present the placements without this caveat. I request that the authors either correct the classification (e.g., mark IG/IH as 'degree-weighted partial b-fANOVA') or add a prominent qualification in Section 4.1 and Figure 1.","section":"Section 3.1, Theorem 1 and Appendix A.1"},{"comment":"Theorem 1 identifies b-fANOVA effects with Taylor interactions only when F is represented by its Taylor series expanded around the baseline b. This analyticity assumption is load-bearing: deep networks with ReLU activations and tree ensembles are not globally Taylor-representable, so for such models the mappings of DeepSHAP, Integrated Gradients, and Integrated Hessians to b-fANOVA effects are at best approximate. The paper should state this assumption explicitly wherever the gradient-method taxonomy is used (Section 4.1 and Appendix E.1.2), rather than only in the Limitations section. The authors should also indicate whether the degree-weight issue in the previous comment is independent of this analyticity concern; the counterexample above uses a polynomial, so it is not an artifact of non-analyticity.","section":"Section 3.1, Theorem 1 and Appendix A.1"}],"minor_comments":[{"comment":"In the bullet 'Interpretations for three types of explanations ...', the word 'undermined' appears where 'underlined' or 'supported' is presumably intended; please correct the typo.","section":"Section 1, Contributions"},{"comment":"The title and abstract contain 'ANOV A' with an extra space; this appears to be a formatting artifact and should be fixed to 'ANOVA'.","section":"Title and abstract"},{"comment":"Corollary 1 is essentially the Möbius inversion theorem applied to the inclusion ordering, as the authors themselves note in Remark 1. This is fine, but the framing as a 'corollary' of the framework should be softened if the reader might otherwise over-interpret it as a new prediction; the current text already handles this, so this is only a presentation suggestion.","section":"Section 3.2, Corollary 1 and Remark 1"},{"comment":"Since Integrated Gradients and Integrated Hessians are not exact Shapley partial effects, Figure 1 and Table 4 would benefit from a footnote or asterisk indicating the degree-weighted nature of these methods, to avoid misleading practitioners who read the table as an exact correspondence.","section":"Figure 1 and Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid framework contribution with correct core theorems and a useful taxonomy. The load-bearing issue is the exactness of the gradient-method placement: Integrated Gradients and Integrated Hessians do not coincide with the Shapley partial effect, and the analyticity assumption of Theorem 1 limits the applicability to ReLU networks and trees. These are fixable within the paper's scope by reclassifying or explicitly qualifying these methods in the tables and main text. The paper should not be rejected, but the camera-ready version must address the degree-weight counterexample and the Taylor-representability caveat directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one, but with the gradient-method caveat in mind from the start. The paper's real contribution is the clean two-axis organization: three fANOVA imputations (baseline/marginal/conditional) crossed with pure/partial/full game-theoretic summaries. For perturbation-based methods the taxonomy is mostly right, Table 2 is a handy reference, and Corollary 1 correctly identifies the Möbius transform with the fANOVA effects. The proofs are standard inclusion-exclusion, and they're done carefully. The code and data are public. That is a solid piece of work for a large part of the XAI toolbox.\n\nThe soft spot is the gradient-method bucket, and it is worse than the reader's Taylor-representability concern. The stress-test counterexample is correct: take F(x1,x2)=x1^2*x2, baseline b=(0,0), instance x0=(1,1). The local b-fANOVA game has Shapley values (0.5, 0.5), but Integrated Gradients returns (2/3, 1/3). The reason is that IG weights the single Taylor interaction by the degree vector (2,1), i.e. by kappa_i/sum(kappa), not by Shapley's uniform 1/s. So IG is not the partial b-fANOVA effect, even for perfectly analytic functions. The same degree-dependence affects Integrated Hessians. Table 4 and Figure 1 list IG under 'partial, b-fANOVA' without this caveat. The Limitations section does say some gradient methods summarize derivatives 'on a more fine-grained level than the related b-fANOVA effects,' which is exactly this issue, but the text doesn't connect that to the tables. A serious referee should ask the authors to either move IG to an 'approximate partial' category or state explicitly in the main text that gradient-method placements are not exact equalities.\n\nThat said, this is not a load-bearing flaw for the whole framework. The perturbation-based taxonomy — SHAP variants, ICE/Occlusion, PDP/M-plots, Sobol indices, SAGE/PFI — holds up as stated. The analyticity assumption for Theorem 1 is a known limitation, not a hidden one. The real-world experiments are illustrative rather than conclusive, and the lack of uncertainty quantification is minor.\n\nMy bottom line: the paper deserves peer review and likely acceptance after a revision that fixes the gradient-method claim. It will be cited, because the two-axis framing is actually useful for teaching and for comparing methods, even if the gradient corner needs an asterisk.","headline":"A genuinely useful two-axis framework for perturbation-based explanations, but the exact placement of Integrated Gradients and other gradient methods under 'partial b-fANOVA' is wrong and needs a caveat.","tokens_in":42176,"tokens_out":1737,"would_cite":true,"duration_ms":18516,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Feature-based explanations reduce to two choices: the feature distribution used for imputation and the aggregation of higher-order interactions.","keywords":["feature-based explanations","functional ANOVA","Shapley value","Shapley interactions","Möbius transform","sensitivity analysis","explainable AI","unified framework"],"falsifier":"Take a one-dimensional ReLU network F(x)=\\max(0,x-c) with a baseline b<c and an instance x_0>c. The exact b-fANOVA effect is $f^{{(b)}}$_1(x_0)=F(x_0)-F(b)=x_0-c, but the Taylor series of F around b is identically zero in a neighborhood of b, so Theorem 1's equality fails. This is a concrete model where the framework's gradient-method categorization is not exact.","tokens_in":41014,"feed_emoji":"🧩","tokens_out":6896,"duration_ms":57112,"temperature":0.7,"pith_summary":"The paper aims to show that most feature-based explanations—local and global, perturbation- and gradient-based—are not separate algorithms but instances of a single construction. That construction pairs a functional ANOVA (fANOVA) decomposition, which decides how strongly the feature distribution enters through baseline, marginal, or conditional imputation, with a game-theoretic aggregation, which decides how strongly higher-order interactions enter through pure, partial, or full effects. The paper proves identities that place SHAP variants, SAGE, PFI, PDP, integrated gradients, DeepSHAP, and others on one grid, and it demonstrates the grid on synthetic and real-world models. If the framework is right, apparent disagreements between explanation methods trace back to these two choices rather than to arbitrary implementation details, giving practitioners a principled way to choose, compare, and interpret methods.","feed_headline":"All feature explanations boil down to two choices","feed_subtitle":"Why it matters: method disagreements now trace to two dials.","key_machinery":"The load-bearing object is the fANOVA decomposition generated by a value function F_S(x)=\\int F(x)\\,dP(x_{-S}), together with the Möbius transform of the resulting explanation game. The fANOVA component recursively defines effects f_S(x)=\\int F(x)\\,dP(x_{-S})-\\sum_{T\\subset S}f_T(x), so the choice of P completely determines one dimension of the framework; the Möbius transform m(S)=\\sum_{T\\subseteq S}(-1)^{s-t}\\nu(T) is the pure additive contribution of a set in cooperative game theory, and Theorem 4 shows that pure, partial, and full influence measures are precisely different summaries of this transform. Theorem 1 supplies the bridge to gradient methods by proving that b-fANOVA effects coincide with Taylor interactions when F is represented by its Taylor series around the baseline.","core_discovery":"The central claim is that every feature-based explanation in the framework's catalog is determined by two components. First, the choice of distribution P in the fANOVA value function F_S(x) = \\int F(x)\\,dP(x_{-S}) specifies how the feature distribution enters the explanation: baseline (b-fANOVA) sets P to Dirac masses at a point b, marginal (m-fANOVA) integrates over the joint marginal distribution, and conditional (c-fANOVA) integrates over p(x_{-S} \\mid x_S). Second, the game-theoretic aggregation of the Möbius transform m(S)=\\sum_{T\\subseteq S}(-1)^{s-t}\\nu(T) yields three regimes: pure effects keep only the isolated effect m(S), partial effects (Shapley value, generalized values, Shapley interactions) spread higher-order terms with index-specific weights, and full effects include all higher-order terms involving the set. Theorem 1 identifies b-fANOVA effects with Taylor interactions, which is what places gradient-based methods in the framework; Theorem 2 states when the three decompositions coincide; Theorem 3 identifies the Möbius transform of the sensitivity game with the variance of c-fANOVA effects for independent features; and Theorem 4 gives the summary rules that make pure, partial, and full effects increasingly sensitive to higher-order interactions.","pith_inferences":["The paper does not make this explicit, but the framework suggests that comparing explanation methods is a matter of comparing two dials; a practitioner could use it to pre-register exactly which distributional and interaction assumptions a method encodes.","An implication beyond the paper is that benchmark disagreements between attribution methods can be diagnosed by holding one dial fixed and varying the other, rather than by treating methods as incommensurable.","One testable extension is to quantify how far gradient-based attributions deviate from exact b-fANOVA effects on non-analytic models such as ReLU networks; Theorem 1 guarantees equality only when the Taylor representation exists.","The appendix already points toward a broader grid: replacing the three imputations with retraining-based value functions would add model fitting as a third dimension, a direction the paper identifies as future work."],"forward_implications":["Baseline, interventional, and observational SHAP are all partial individual effects that differ only in whether the imputation is b-, m-, or c-fANOVA.","PDP and M-plots compute pure joint effects of m- and c-fANOVA, so their centering choices align directly with fANOVA components.","SAGE and PFI/CFI are the partial and full individual effects of the risk game, meaning they answer different questions about performance loss rather than competing versions of the same question.","For independent features, the Möbius transform of the sensitivity game is the variance of the corresponding c-fANOVA effect, linking global sensitivity indices for dependent data to the same grid.","Gradient-based methods such as Integrated Gradients, DeepSHAP, and Integrated Hessians summarize b-fANOVA effects, subject to the Taylor-representability condition in Theorem 1."],"supporting_citations":[{"why":"Supplies the Taylor interaction decomposition that Theorem 1 equates with b-fANOVA effects and that carries the gradient-method part of the unification.","marker":"(Deng et al., 2024)"},{"why":"Defines SHAP and its baseline, interventional, and observational variants, which the framework categorizes as partial individual effects.","marker":"(Lundberg and Lee, 2017)"},{"why":"Provides the Shapley value and its axiomatic characterization, the basis for all partial individual effects in Theorem 4.","marker":"(Shapley, 1953)"},{"why":"Provides the Möbius transform and co-Möbius transform representations that underlie pure and full effects.","marker":"(Grabisch et al., 2000)"},{"why":"Introduces the superset measure and generalized fANOVA diagnostics, which the framework classifies as full interaction effects.","marker":"(Hooker, 2004)"},{"why":"Supplies the variance decomposition and Sobol-index background that the global sensitivity game builds on.","marker":"(Owen, 2013)"},{"why":"Unifies Shapley interactions and Shapley-GAM across marginal and conditional imputations, anchoring the partial interaction effects.","marker":"(Bordt and von Luxburg, 2023)"},{"why":"Defines the Shapley Taylor interaction index, categorized here as a partial interaction effect of b-fANOVA.","marker":"(Sundararajan et al., 2020)"},{"why":"Introduces M-plots and ALE, which the framework connects to c-fANOVA pure effects and derivative-level explanations.","marker":"(Apley and Zhu, 2020)"}],"fun_headline_variants":["Two dials explain all feature attribution methods","Every explanation method is two choices away","Unified theory: fANOVA + game theory explain them all","Two dials: distribution and interaction shape all explanations","One framework unifies attribution: distribution + interaction choice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The identification of b-fANOVA effects with Taylor interactions in Theorem 1 assumes that the model F is represented by its infinite Taylor series around the baseline b; for ReLU networks and tree ensembles this analyticity fails, so the framework's placement of gradient-based methods is only approximate for those model classes.","fun_headline_variants_meta":{"raw":{"variants":["Two dials explain all feature attribution methods","Every explanation method is two choices away","Unified theory: fANOVA + game theory explain them all","Two dials: distribution and interaction shape all explanations","One framework unifies attribution: distribution + interaction choice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001012,"raw_usage":{"total_tokens":4276,"prompt_tokens":946,"completion_tokens":3330,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":3256}},"tokens_in":562,"tokens_out":3330,"duration_ms":21993,"temperature":1.0,"reasoning_tokens":3256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:45:00.375696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a one-dimensional ReLU network F(x)=\\max(0,x-c) with a baseline b<c and an instance x_0>c. The exact b-fANOVA effect is $f^{{(b)}}$_1(x_0)=F(x_0)-F(b)=x_0-c, but the Taylor series of F around b is identically zero in a neighborhood of b, so Theorem 1's equality fails. This is a concrete model where the framework's gradient-method categorization is not exact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Taylor interaction decomposition that Theorem 1 equates with b-fANOVA effects and that carries the gradient-method part of the unification."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines SHAP and its baseline, interventional, and observational variants, which the framework categorizes as partial individual effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Shapley value and its axiomatic characterization, the basis for all partial individual effects in Theorem 4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Möbius transform and co-Möbius transform representations that underlie pure and full effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the superset measure and generalized fANOVA diagnostics, which the framework classifies as full interaction effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variance decomposition and Sobol-index background that the global sensitivity game builds on."},{"cited_title":"and von Luxburg, U","cited_arxiv_id":null,"evidence_quote":"Unifies Shapley interactions and Shapley-GAM across marginal and conditional imputations, anchoring the partial interaction effects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces M-plots and ALE, which the framework connects to c-fANOVA pure effects and derivative-level explanations."}],"review_version":1}