{"id":"a627da04-5412-4c86-bb0a-2c48b358ceb2","arxiv_id":"1908.05978","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A trained MLP can be rewritten as a sparse additive model of one- and two-variable effects, called a Partial Response Network, with no loss in classification accuracy on benchmark tabular data.","lead":"This paper introduces the Partial Response Network (PRN), a method that rewrites a trained neural network as a sparse additive model of one- and two-variable effects. On five benchmark datasets, the resulting nomograms match or beat the original black-box network's accuracy, suggesting interpretability does not have to cost performance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unverified truncation hypothesis: the MLP's logit is assumed to be well approximated by univariate and bivariate ANOVA terms, but no evidence is given for datasets where higher-order interactions matter.","rationale":"Agreement with reader: the weakest assumption is the truncation hypothesis. This stress-test sharpens it: the anchored partial responses can be zero for every first- and second-order term even when the MLP relies on a third-order interaction, so the proposed test would directly falsify the 'fully interpretable representation' claim. I do not think this moves the verdict: the paper already receives CONDITIONAL, and the concern specifies the condition (demonstrate truncation holds or characterize when it fails). The check is decisive and easy to run once code is released. I also note the orthogonality claim for the anchored components is not generally true, but it is secondary; even if the components were non-orthogonal, the method might still work, whereas the truncation is necessary for the method to work at all. Hence no verdict change.","tokens_in":14413,"tokens_out":11910,"duration_ms":113635,"concrete_test":"Run the PRN pipeline on a synthetic binary classification problem whose Bayes logit is a pure third-order interaction, e.g., y ~ Bernoulli(sigmoid(3*x1*x2*x3)) with independent Gaussian inputs and n=2000 train / 1000 test. Train an MLP with a single hidden layer and ARD as in §2.1, then apply steps ii–v. Because φ_i, φ_ij are identically zero at the anchor for a product interaction, the Lasso will select no univariate or bivariate terms; if the final PRN's test AUROC is near 0.5 while the original MLP's is substantially higher, the truncation hypothesis fails. A complementary check is to compute the L2 norm of the third-order ANOVA term of the trained MLP on each benchmark dataset relative to total logit variance; if that ratio is above 10% for any dataset, omitted interactions are not negligible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—that the PRN is a fully interpretable representation of the MLP with comparable or superior performance—depends on the Section 2 'truncation hypothesis': identity (1) is exact, but the method keeps only univariate and bivariate terms and discards all higher-order interactions as noise. This is an empirical assumption, not derived from the MLP or from theory. The benchmark evidence (Tables 1 and 2) is from five UCI datasets plus Shuttle, all of which are known to be well fitted by low-order additive or linear models; the cited review [2] even argues that machine learning offers no benefit over logistic regression for clinical prediction. Replacing the MLP's logit by a truncated ANOVA is therefore untested in regimes where the MLP's accuracy relies on a genuine third-order or higher interaction. In such a regime the anchored partial responses defined by (14)–(16) can be almost identically zero for every univariate and bivariate term, so the Lasso would select no features and the PRN would reduce to a constant, losing all discriminative power. Additionally, the paper claims the ANOVA terms are 'orthogonal in a functional sense' and thus independent, but the recursively computed components (14)–(16) are anchored at the median, not the orthogonal projection components of the functional ANOVA [19]; orthogonality does not generally hold, so the 'independent effects' interpretation is also not established. The truncation assumption is the load-bearing one, and it is currently unsupported in general.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Partial Response Network (PRN), a method that starts from a trained multilayer perceptron (MLP), decomposes its logit using a functional ANOVA representation anchored at the median, truncates the decomposition to univariate and bivariate terms, and applies a logistic Lasso for feature selection. The selected partial responses are then replicated as a generalized additive neural network / self-explaining neural network, which can be retrained and optionally re-pruned. The authors claim that the PRN is as transparent as logistic regression and achieves comparable or superior classification performance to the original MLP on five UCI datasets plus the Statlog Shuttle dataset.","tokens_in":14668,"tokens_out":6374,"duration_ms":69436,"significance":"If the central claim is validated, the PRN would be a practically useful bridge between flexible neural-network classifiers and globally interpretable additive models, offering sparse feature selection and a nomogram-like visualization of univariate and bivariate effects. The paper has several genuine strengths: identity (1) and the recursive construction (14)-(17) are exact by design; the weight-replication equations (18)-(23) are explicit and testable; the empirical section compares against strong baselines including SAM, GBM, SVM, and RF; and the reported stability across random initializations is encouraging. The main reservations are that the truncation hypothesis is an empirical assumption supported only by low-dimensional benchmark data, and that the stated orthogonality of the anchored decomposition is not generally correct.","major_comments":[{"comment":"The claim that the decomposition terms are 'orthogonal in a functional sense' and can be regarded as independent effects is not correct for the components defined in Eqs. (14)-(16). These are point-anchored (cut-HDMR) components obtained by setting all other variables to zero, not the measure-based orthogonal ANOVA components discussed in [19]. In the standard functional ANOVA, components are typically defined through conditional expectations and satisfy zero-integral conditions, which do not generally hold for the anchored recursion. The Lasso is applied to these components, so the selection and the interpretation of each partial response as an independent effect depend on the actual correlations among components. I recommend either replacing the anchored components with measure-based orthogonal ANOVA components or explicitly reframing the method as an anchored decomposition and removing the orthogonality/independence interpretation.","section":"Section 2, Eqs. (14)-(17)"},{"comment":"The central claim that the decomposition can be truncated to univariate and bivariate terms 'without compromising performance' rests on an untested empirical assumption. The paper's own wording calls this a hypothesis, but the validation is limited to five UCI datasets plus Shuttle, all comparatively low-dimensional and known to be well approximated by low-order structure; there is no synthetic or real dataset in which higher-order interactions are known to be essential. In a regime where a third-order or higher interaction is required, the anchored univariate and bivariate components can be arbitrarily close to zero, and the PRN could collapse to a constant while the MLP retains discriminative power. I ask the authors to add a diagnostic for the magnitude of the discarded higher-order terms and to include at least one experiment (e.g., synthetic data with a controlled interaction order) that characterizes when truncation fails, or to soften the 'fully interpretable representation' claim accordingly.","section":"Section 2, truncation hypothesis; Section 3, Tables 1-2"},{"comment":"The sentence stating that performance differences are not significant at the 5% level via the McNemar test is unsupported by any reported test statistic, p-value, or description of how the test was applied across the ten random initializations. The reported standard deviations for the PRN are very small relative to the MLP, but it is unclear whether the ten runs vary the MLP initialization on a single fixed split or also vary the split; the text describes a fixed split for Pima but not consistently for the other datasets. This information is needed to assess the 'comparable or superior performance' claim.","section":"Section 3, Table 1 and McNemar statement"}],"minor_comments":[{"comment":"The citation for Generalized Additive Models appears to be incorrect: the text says 'Generalized Additive Models (GAMs) [11]', but reference [11] is the Sparse Additive Model paper by Ravikumar et al.; the GAM citation should likely be [10] or another appropriate source.","section":"Section 1.1 and References"},{"comment":"The term 'logistic Lasso' is cited to [21], which is the group lasso for logistic regression; the text later refers to 'group Lasso' in the Discussion. Please clarify whether a standard Lasso or a group Lasso is used for the partial responses, since this affects the theoretical justification and implementation.","section":"Section 2.2"},{"comment":"The notation in Eq. (17) is ambiguous: the summation over subsets '{i1,...,in-1}' should specify that the sum runs over all proper subsets of the index set, and the notation in Eq. (1) for pairs should be i<j rather than i≠j to avoid double-counting.","section":"Eq. (17) and surrounding text"},{"comment":"No confidence intervals or standard errors of the AUROC differences are reported, and the claim that PRN 'often selects fewer variables than SAM' is not quantified in a table; please add the selected-feature counts for each dataset and method.","section":"Section 3, Tables 1-2"},{"comment":"There are numerous typographical errors, including 'univatiate', 'wtih', 'explitic', 'trunctated', 'ANOV A', 'inonosphere', and 'witb'; a thorough proofreading pass is needed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable contribution to interpretable ML, but the central claim is currently stronger than the evidence. The orthogonality issue is a technical error that should be fixed; the truncation hypothesis needs explicit stress-testing or a more careful statement of scope. The empirical evaluation also lacks the statistical detail normally expected for a claim of 'comparable or superior performance'."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives a practical pipeline: train an MLP, decompose its logit with an anchored functional ANOVA, select univariate and bivariate terms with the logistic Lasso, then re-code the selected terms as a GANN and retrain. The result is a sparse, additive nomogram for tabular data. That's genuinely useful, and the empirical comparison across six UCI datasets shows the PRN is competitive with the original MLP and with GBM/SVM/RF. On Ionosphere it actually beats the MLP by a clear margin (98.1 vs 94.6 AUROC), which is a nice counterexample to the usual accuracy-interpretability trade-off.\n\nThe novelty is modest but real. The pieces are known—ANOVA decompositions, Lasso, GANNs—but the specific combination and the weight-replication formulas are new as far as the cited literature goes. The paper is also honest about its main assumption: the truncation hypothesis in Section 2. It states outright that retaining only univariate and bivariate terms is a hypothesis, not a theorem. That honesty counts.\n\nNow the soft spots. The \"fully interpretable representation of the MLP\" claim is too strong. The PRN is a truncated representation, and the truncation is tested only on datasets that are known to be reasonably well fitted by additive or low-interaction models. If a problem's logit is dominated by a genuine third-order or higher interaction, the anchored univariate and bivariate terms can be near zero everywhere, the Lasso selects nothing, and the PRN collapses to a constant. The paper does not address this failure mode.\n\nThe orthogonality claim is also overstated. The recursive terms are anchored at the median, not the orthogonal projection components of functional ANOVA in Hooker's sense. So calling them \"orthogonal in a functional sense\" and therefore \"independent inputs\" is not accurate. It's a moderate issue: the method still works, but the statistical interpretation is muddier than the text implies.\n\nFinally, no code is provided, and the benchmark details are thin: no architecture specifics, no Lasso implementation, no confidence intervals for the McNemar tests. That makes reproduction harder than it should be for a methods paper.\n\nWho is this for? Researchers working on interpretable ML for tabular medical or financial data. A serious referee should engage with it—the core idea deserves scrutiny and the empirical claims need independent reproduction. I would not desk-reject.\n\nRecommendation: send to peer review, but expect heavy revision on the truncation generalizability and the orthogonality claim, and require code or a detailed reproducibility appendix.","headline":"A useful but not fully verified recipe for converting an MLP into a sparse additive nomogram; the main claim outruns the evidence.","tokens_in":15238,"tokens_out":4626,"would_cite":true,"duration_ms":42894,"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 Partial Response Network rewrites a trained multilayer perceptron as a sparse additive model of one- and two-variable effects, matching black-box accuracy while making every effect visible in a nomogram.","keywords":["interpretability","generalised additive neural networks","self-explaining neural networks","functional ANOVA","partial responses","nomogram","logistic Lasso","tabular data"],"falsifier":"Construct or find a tabular classification problem with an essential three-way interaction, train an MLP, and apply the PRN pipeline; if the PRN's holdout AUROC falls clearly below the original MLP's, the truncation hypothesis is false for that setting. A broader version is a benchmark sweep measuring the PRN-minus-MLP AUROC gap across many datasets to estimate how often the truncation holds.","tokens_in":14182,"feed_emoji":"📊","tokens_out":8833,"duration_ms":81808,"temperature":0.7,"pith_summary":"The Partial Response Network (PRN) turns a trained multilayer perceptron into a globally interpretable additive model. The paper shows that the MLP's logit, the log-odds of class membership, can be decomposed exactly into a constant plus functions of single variables and pairs of variables, and that keeping just these low-order components is enough to retain, and sometimes improve, classification performance. Applying a logistic Lasso to the component functions selects a sparse set of partial responses, which are then rebuilt as a self-explaining network with the same transparent structure as a traditional logistic regression. On six tabular benchmark problems the PRN matches or beats the original MLP, gradient boosting, support vector machines, and random forests in AUROC while using fewer variables. If the claim holds, practitioners in high-stakes fields can have an accurate model and read its reasoning directly from a nomogram.","feed_headline":"Neural net becomes a nomogram without losing accuracy","feed_subtitle":"Partial Response Network keeps a black-box MLP's performance while exposing every one- and two-variable effect.","key_machinery":"The machinery is the functional ANOVA decomposition of the logit, anchored at the median of the data. Each partial response is computed by holding all other inputs at the anchor: the univariate term for $x_i$ is the logit at $(0,\\ldots,x_i,\\ldots,0)$ minus the anchor logit, and the bivariate term for $(x_i,x_j)$ subtracts the two univariate terms and the anchor. The paper emphasises that these terms add back up to the exact logit, so interpretability is not bought by a local Taylor expansion. The logistic Lasso then selects which partial responses are statistically meaningful, and the selected terms are re-expressed as a GANN/SENN by copying the hidden weights of the original MLP and rescaling the output weights and biases according to equations (18)--(23), so the structured network's initial output exactly matches the Lasso fit. A final pass of gradient descent recalibrates the component functions without the removed variables and interactions, which the paper argues further smooths the curves and can improve accuracy.","core_discovery":"The paper's central claim is that a fully connected MLP for binary classification does not have to be treated as a black box. The functional ANOVA decomposition of its logit is an identity, not a local approximation: it holds at every point in input space and decomposes the logit into a sum of a constant, univariate terms, bivariate terms, and higher-order interactions. The authors hypothesise that the higher-order interactions can be truncated without compromising performance, and the experiments support that hypothesis on the benchmarks considered. What remains is a sparse additive model whose components, the partial responses, are selected by a logistic Lasso and then assembled into a Generalised Additive Neural Network, a self-explaining network whose output-layer weights are rescaled copies of the original MLP weights. After a short gradient-descent recalibration, the resulting PRN is claimed to be exactly as transparent as logistic regression while matching or exceeding the original MLP's AUROC; on the Shuttle data it reaches AUROC 1.000 from just two variables and one two-way interaction.","pith_inferences":["The same ANOVA-plus-Lasso recipe could be applied to other probabilistic base classifiers, for instance probabilistic support vector machines or gradient-boosted probability models, to give them the same nomogram-style explanations.","If the stability across random seeds extends beyond the benchmarks, the PRN could be used as a diagnostic for whether a black-box MLP has learned spurious high-order structure: effects that disappear under truncation are likely noise.","The Lasso coefficients on partial responses may be interpretable as odds-ratio weights, which would let clinicians quote adjusted odds ratios for continuous variables directly from the network."],"forward_implications":["Users of tabular classifiers can obtain a global nomogram showing how each variable and pair of variables shifts the log-odds, not just a local explanation around one test point.","GANN architecture is inferred from data: the logistic Lasso over partial responses decides which univariate and bivariate terms to keep, removing the need for predefined structure or prior feature selection.","The PRN can outperform the original MLP on the same data, because dropping noisy higher-order interactions acts like signal denoising rather than a loss of information.","Bivariate terms are sufficient to capture real interactions, as in the Shuttle example where two variables plus one two-way interaction reach perfect AUROC.","The method is stable: random restarts of the MLP produce almost identical partial responses and consistent selected structure."],"supporting_citations":[{"why":"Provides the functional ANOVA decomposition used to represent the MLP logit as a global sum of low-order component functions.","marker":"[19]"},{"why":"Supplies the Bayesian Automatic Relevance Determination framework used to train the initial MLP with soft feature selection.","marker":"[20]"},{"why":"Defines the Lasso penalty used to select a sparse set of partial responses.","marker":"[13]"},{"why":"Extends Lasso-type selection to grouped logistic regression, matching the grouped partial-response structure.","marker":"[21]"},{"why":"Defines Self-Explaining Neural Networks, the additive architecture class the PRN instantiates.","marker":"[9]"},{"why":"Earlier SVM nomogram method whose partial-response idea the paper generalises to MLP logits.","marker":"[18]"},{"why":"Defines Sparse Additive Models, the spline-based benchmark against which the PRN is compared.","marker":"[11]"},{"why":"Provides the backfitting GAM estimation used for the Sparse Additive Model baseline.","marker":"[10]"}],"fun_headline_variants":["Neural net black box opens into a nomogram","Partial Response Network: interpretable MLP, same accuracy","MLP's black box becomes a transparent nomogram","PRN: nomogram from MLP without loss of performance","From deep net to interpretable nomogram with equal AUROC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ANOVA decomposition of a trained MLP's logit can be truncated to univariate and bivariate terms without a meaningful loss of classification performance.","fun_headline_variants_meta":{"raw":{"variants":["Neural net black box opens into a nomogram","Partial Response Network: interpretable MLP, same accuracy","MLP's black box becomes a transparent nomogram","PRN: nomogram from MLP without loss of performance","From deep net to interpretable nomogram with equal AUROC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1877,"prompt_tokens":1093,"completion_tokens":784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":709,"completion_tokens_details":{"reasoning_tokens":703}},"tokens_in":709,"tokens_out":784,"duration_ms":7716,"temperature":1.0,"reasoning_tokens":703,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:58:37.056273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or find a tabular classification problem with an essential three-way interaction, train an MLP, and apply the PRN pipeline; if the PRN's holdout AUROC falls clearly below the original MLP's, the truncation hypothesis is false for that setting. A broader version is a benchmark sweep measuring the PRN-minus-MLP AUROC gap across many datasets to estimate how often the truncation holds.","supporting_citations":[{"cited_title":"(10) This uses the fact that when the MLP has converged, the cost function is at an extremum hence the first derivative vanishes","cited_arxiv_id":null,"evidence_quote":"Provides the functional ANOVA decomposition used to represent the MLP logit as a global sum of low-order component functions."},{"cited_title":"de Waal, J","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian Automatic Relevance Determination framework used to train the initial MLP with soft feature selection."},{"cited_title":"Alvarez-Melis, T.S","cited_arxiv_id":null,"evidence_quote":"Defines the Lasso penalty used to select a sparse set of partial responses."},{"cited_title":"Cadarso-Suárez, J","cited_arxiv_id":null,"evidence_quote":"Extends Lasso-type selection to grouped logistic regression, matching the grouped partial-response structure."},{"cited_title":"Lundberg, S.-I","cited_arxiv_id":null,"evidence_quote":"Defines Self-Explaining Neural Networks, the additive architecture class the PRN instantiates."},{"cited_title":"Sarle, Neural Networks and Statistical Models, SAS USers Gr","cited_arxiv_id":null,"evidence_quote":"Earlier SVM nomogram method whose partial-response idea the paper generalises to MLP logits."}],"review_version":1}