{"id":"b1fcae9d-108b-47e1-b35f-70e521545b70","arxiv_id":"2607.27474","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"NMF-RRR tri-factorizes the non-negative regression coefficient matrix as X1ΘX2, soft-clustering both variable blocks and providing a conditional Wald test for the block-correspondence entries, with size control only in identifiable settings.","lead":"Satoh and Tokuda propose NMF-RRR, a method that splits both the inputs and the outputs of a multivariate regression into soft clusters, then tests which input clusters drive which output clusters. Because the test is performed after the clusters are learned from the same data, the authors call it exploratory and show it is not reliable for measuring link strength.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Path-level tests are not supported in the recommended Q<R regime: per-path nulls are not factorization-invariant, and the 'conservative after re-estimation' claim is conditional on identifiable Q=R rank recovery.","rationale":"I read the paper as a methodological proposal whose central added value is the estimable, testable Θ. The reader's weakest assumption identifies exactly the point that matters: inference for Θ is conditional on estimated bases, not valid selective inference, and it is not decomposition-invariant in the Q<R regime the paper recommends. The paper is unusually honest about these limitations, and the simulations are consistent with the hedged claims, so I do not think the reader's conditional accept should be changed to reject. However, the abstract's wording 'conservative after re-estimation' goes beyond what Table 10 actually shows, because that claim is conditional on rank recovery and on identifiable Q=R designs. My concrete test would settle whether the path-level test has any meaning in Q<R: if significance varies across equally fitting factorizations, then the tested paths are properties of the selected factorization, not of the underlying coefficient matrix. Since the reader already flagged this and conditioned the verdict accordingly, no verdict adjustment is needed.","tokens_in":32589,"tokens_out":7462,"duration_ms":80933,"concrete_test":"In a synthetic Q=2,R=3 design, fix M=X1ΘX2 with response-group anchors (so X1 and G=ΘX2 are identified) and one numerically zero entry in the generating Θ. Run the full pipeline from Section 7 (rank CV, multistart refit, factor alignment) B≥1000 times. For each replication, record the Wald p-value for the target zero path under the selected factorization, and additionally refit the same data from many random starts to enumerate equally fitting factorizations. Report (i) the rejection rate of the zero path at 5% without conditioning on rank recovery, and (ii) the variability of that p-value across equally fitting factorizations. If rejection exceeds 5% or the p-value is not stable across factorizations, the path-level test is an artifact of factorization selection rather than a property of M.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's distinctive inferential contribution is a 'tested matrix of block correspondences' whose entries are estimated and tested (Section 2; abstract). But the calibration evidence for the realistic re-estimated pipeline does not cover the regime the paper recommends. Section 4.3 explicitly recommends Q≤R and takes Q<R as the way to expose cross-structure ('one response group integrates several covariate drivers'). Appendix A, Proposition 5(c) states that for Q<R, Θ and X2 are not separately identified from M even under separability and normalization; Section 5 concedes that for Q<R the null θqr=0 is defined only relative to the selected factorization, so 'presence of a path' is not a decomposition-invariant hypothesis. The simulations supporting 'conservative after re-estimation' (Tables 9–10) are run in identifiable Q=R designs, and Table 10 conditions on recovering the true ranks (74.6% in the synthetic design); unconditional behavior, and any behavior in Q<R, is not reported. Thus the abstract's claim that the conditional Wald test is 'conservative after re-estimation' is established only for a narrow identifiable special case, while the method's primary use case (Q<R cross-structure) lacks a well-defined per-path type-I error. The real-data Q<R findings (nutrimouse, FRANZOSA) are exploratory screens conditional on a chosen factorization, not tested block correspondences in a unique model. This is disclosed in the text, but it undercuts the central selling point as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tri-factorized non-negative reduced-rank regression (NMF-RRR) for two-block data, modeling Y1 ≈ X1 Θ X2 Y2 so that the non-negative regression coefficient matrix M = X1ΘX2 is a soft tri-factorization: X1 clusters response variables, X2 clusters covariate variables, and Θ is a block-correspondence matrix whose entries are estimated and tested. The paper derives multiplicative updates, rank selection by element-wise and sample-wise cross-validation, a conditional Wald test with sandwich standard errors, and identifiability results in Appendix A. It reports simulations on calibration and four applications (Doubs, nutrimouse, FRANZOSA, Wine). The manuscript is unusually candid about limitations: Section 5 states that for Q<R the null θqr=0 is defined only relative to the selected factorization, and Section 7 restricts per-path type-I error simulations to identifiable Q=R designs conditional on rank recovery.","tokens_in":32933,"tokens_out":7115,"duration_ms":74552,"significance":"If the claims were fully supported, the method would fill a genuine gap: a two-sided soft co-clustering of response and covariate variables with an estimable block-correspondence matrix, connecting reduced-rank regression and non-negative matrix factorization. Strengths include the explicit identifiability analysis (Appendix A), transparent simulation reporting, reproducible code and data scripts, and the honest disclosure of the conditional nature of the inference. However, the paper's central inferential selling point—a 'tested matrix of block correspondences'—is narrower than the abstract suggests. The 'conservative after re-estimation' claim is established only for identifiable Q=R designs conditional on rank recovery (Table 10), while the recommended Q<R regime lacks a decomposition-invariant per-path null. The method is still valuable as an exploratory co-clustering and estimation tool, but the path tests in Q<R applications should be described as conditional screens, not confirmatory tests.","major_comments":[{"comment":"The abstract's claim that the conditional Wald test is 'conservative after re-estimation' is not supported in the regime the paper recommends. Section 4.3 recommends Q≤R and uses Q<R to expose cross-structure (nutrimouse Q=2,R=3; FRANZOSA Q=2,R=4), but Section 5 and Appendix A Prop. 5(c) state that for Q<R the split into Θ and X2 is not uniquely identified, so the null θqr=0 is defined only relative to the selected factorization. The simulations in Section 7 restrict per-path size to identifiable Q=R designs, and Table 10 reports sizes conditional on recovering the true ranks (74.6% recovery in the synthetic design). The abstract should be revised to say 'conservative in the identifiable Q=R designs examined, conditional on rank recovery,' and the paper should either provide Q<R simulations that evaluate path tests under the selected factorization or explicitly frame all Q<R path tests a","section":"§4.3, §5, §7, Appendix A Prop. 5(c)"},{"comment":"The masked (weighted) updates for element-wise cross-validation are delegated to the nmfkc package without derivation. Since the rank-selection rule (12) and the full-pipeline simulation (Table 10) depend on these updates, the paper should provide the explicit update formulas and state whether monotonicity/KKT properties analogous to Proposition 4 hold. As written, a central algorithmic component is a black box, which hampers reproducibility and verification.","section":"§4.1"},{"comment":"The 'existence test stays safe' conclusion is based on very limited evidence. Table 9 shows empirical size 0.000 at only two true-zero paths in a high-signal Doubs-calibrated design; this is extreme conservatism, not a demonstration of controlled size. Table 10 conditions on recovering the true ranks and excludes 25.4% of replications in the synthetic design. The unconditional behavior of the full procedure is not reported, and the paper itself notes that an unconditional per-path type-I error is not well defined when Q<R. This limitation should be stated in the abstract and in the conclusions, not only in the simulation section, because the real-data results in Section 6 report significant paths from Q<R analyses.","section":"§7, Tables 9–10"},{"comment":"The wild bootstrap is presented as providing interval estimates, but its calibration is simulated only with fixed bases and correlated errors (Table 11), not with re-estimated bases or rank selection. Given that Table 9 shows severe bias in path magnitudes after basis re-estimation, the bootstrap intervals used in the real-data analyses should either be accompanied by a re-estimation simulation or be labeled as exploratory. Currently the text in Section 6 describes significant Θ entries as 'tested present/absent links' while magnitudes are 'conditional on the estimated bases,' but the absence of bootstrap calibration under re-estimation leaves the interval claims unverified.","section":"§5, §6, Table 11"}],"minor_comments":[{"comment":"The phrase 'conservative after re-estimation' should be qualified as 'conservative in the identifiable Q=R designs examined, conditional on rank recovery.' As written, it overstates the scope of the simulation evidence.","section":"Abstract"},{"comment":"Calling empirical size 0.000 'conservative' is correct but misleading; the test never rejects at the null paths, which is far more conservative than nominal. Consider describing it as 'over-conservative in these settings' and noting that this may reflect low sensitivity rather than ideal control.","section":"§7, Table 9"},{"comment":"The statement that re-fitting the four analyses under the KL objective 'left the significant structure intact' is important but is not accompanied by a table or numeric summary. Please provide the results in an appendix or online supplement.","section":"§8"},{"comment":"The reference to 'Huet's classical longitudinal zonation' is not in the reference list. Please add a citation.","section":"§6.1"},{"comment":"The notation 'σ2IP1N' should be typeset as σ^2 I_{P1N} for consistency with the rest of the paper.","section":"§5, Eq. (14)"},{"comment":"Equation (12) uses σ(Q,R) for RMSE; later in Section 7 the one-standard-error rule is applied to σ². Clarify this distinction in the text.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its limitations, but the abstract and the framing of the 'tested matrix of block correspondences' overstate the inferential contribution relative to the simulation evidence. The heavy reliance on the author's own prior work (Satoh 2023, 2025, 2026a-c) and the delegation of the masked updates to the nmfkc package make independent verification more difficult; I suggest requesting the derivation in the main text or in an appendix. If the authors are willing to reposition the method primarily as an exploratory tool for co-clustering with conditional path screens, the paper could be suitable for publication after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's contribution is real and easy to state: it writes the non-negative coefficient matrix of a multivariate regression as M = X1 Θ X2, with column/row-normalized X1, X2 that act as soft cluster profiles for the response and covariate blocks, and a between-block link matrix Θ that is estimated and tested. That puts it in a different class from nn-PRRR's bi-factorization, signed sparse RRR, and hard fusion clustering. The math core is standard but carefully stated — rank bound, multiplicative updates with monotonicity, Appendix A on identifiability — and the package plus scripts make the work reproducible. The applications (Doubs, nutrimouse, FRANZOSA, Wine) are sensible and the comparisons are fair.\n\nThe authors are also unusually upfront about limitations. They say the Wald test for Θ is conditional on the chosen bases, not a formal selective-inference construction; they note that for Q<R (the regime they recommend for cross-structure) the null θqr=0 is not decomposition-invariant; and they show in Table 9 that with re-estimated bases the magnitude intervals fail to cover truth. That honesty is the paper's best feature.\n\nThe soft spot is in the gap between the abstract and the simulations. The abstract says the test is 'conservative after re-estimation.' What Table 10 establishes is conservativeness conditional on the replication having recovered the true ranks — recovery is 74.6% in the synthetic design. Unconditional behavior isn't reported, and in the Q<R regime the per-path null isn't well defined at all. So the central selling point, a 'tested matrix of block correspondences,' is only a conditional, exploratory screen in the primary use case. Reading the text carefully, the authors mostly disclose this; the abstract overstates it.\n\nMinor issues: the masked weighted updates for element-wise CV are delegated to the nmfkc package without derivation, and the inference framework leans on the first author's earlier GCM work. Neither is disqualifying.\n\nWho gets value: applied statisticians in two-block fields — ecology, nutrigenomics, microbiome–metabolome — who want interpretable soft co-clustering plus a path screen. I'd send it to a serious referee, with the instruction that the abstract and discussion be rewritten so the conditioning and rank-recovery qualifiers are explicit, and that the masked-update derivation be provided or referenced. The method is not a home run, but it is coherent, new, and reproducible.","headline":"A novel tri-factorized non-negative RRR with conditional inference; worth refereeing, but the abstract oversells the 'conservative after re-estimation' claim and per-path tests aren't factorization-invariant in the recommended Q<R regime.","tokens_in":33484,"tokens_out":3466,"would_cite":true,"duration_ms":36029,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62J05","62H25","62F03"],"pacs":[],"model":"deepseek-v4-flash","headline":"A non-negative tri-factorization of the regression coefficient matrix yields soft clusters on both sides of a two-block regression, with significance-tested links between them.","keywords":["Co-clustering","Non-negative matrix tri-factorization","Reduced-rank regression","Block-correspondence inference","Multivariate regression","Two-block data","Conditional Wald test","NMF-RRR"],"falsifier":"Run the full pipeline—rank selection by cross-validation, basis re-estimation, conditional Wald test—on data generated from the paper's identifiable Q=R=3 null design, but count every replication, including those where cross-validation fails to recover the true ranks. The paper reports sizes only after conditioning on correct rank recovery; an unconditional rejection rate above 5% would show that rank selection adds over-rejection that the conditioning hides, while a rate at or below nominal would support the conservative-size claim.","tokens_in":32474,"feed_emoji":"📊","tokens_out":5870,"duration_ms":58825,"temperature":0.7,"pith_summary":"The paper is trying to establish that the relationship between two blocks of variables measured on the same individuals can be read as a co-clustering of the variables themselves. It constrains the multivariate regression Y1 ≈ M Y2 so that the coefficient matrix factors as M = X1 Θ X2, where X1 softly groups response variables, X2 softly groups covariates, and the middle factor Θ records which covariate groups drive which response groups. The paper develops multiplicative updates for fitting, cross-validation for the two group counts, and a conditional Wald test for the entries of Θ. If correct, the method supplies interpretable, tested group-to-group links in settings such as microbiome–metabolome studies, where classical reduced-rank regression and canonical correlation analysis return signed factors or become ill-posed.","feed_headline":"One factorization co-clusters both blocks of a regression","feed_subtitle":"Non-negative factors split responses and covariates into soft groups; a tested middle matrix links them","key_machinery":"The load-bearing object is the tri-factorization M = X1 Θ X2 of the regression coefficient matrix, with X1 and X2 non-negative and normalized so that their columns and rows sum to one, turning them into soft cluster profiles, and with Θ a non-negative Q×R parameter matrix. The middle factor Θ is what the paper identifies as the minimal structure that decouples the response and covariate groupings and carries the tested block correspondences. The machinery also includes multiplicative updates that preserve non-negativity and decrease the squared error, a normalization identity that ties the sum of the entries of Θ to the grand total of the coefficient matrix, and identifiability conditions (a","core_discovery":"The central claim is that the low-rank regression coefficient M = X1 Θ X2, with non-negative and normalized factors, is the natural algebraic object for two-block co-clustering: X1 and X2 are soft cluster profiles of response and covariate variables, and Θ is a Q×R block-correspondence matrix whose entries are estimated and tested. The estimator is called tri-factorized non-negative reduced-rank regression (NMF-RRR); because rank(M) ≤ min(Q,R), it belongs to the same model class as reduced-rank regression, but expresses the shared subspace in non-negative parts rather than signed directions. The paper establishes that fit is capped by min(Q,R), that a square Θ tends toward a near-permutation","pith_inferences":[],"forward_implications":["On microbiome–metabolome data, the method can recover a pronounced cross-structure in which each metabolite module is associated with two microbial groups, all paths significant, a pattern a square permutation Θ cannot express.","Because fit is capped by min(Q,R), the paper recommends taking Q ≤ R: extra response factors beyond the rank are not separately identifiable, while extra covariate factors can expose cross-structure.","When the number of covariates exceeds the sample size, the non-negative, normalized, low-rank parameterization remains well-behaved where unregularized CCA and RRR are ill-posed, at the price of lower in-sample fit.","Re-estimating the bases from the same data keeps the existence test conservative but makes the magnitudes of non-zero paths unreliable, so significant Θ entries should be read as tested presence or absence, not as effect sizes.","When covariate blocks are uncorrelated, unsupervised tri-NMF of the association matrix recovers the same clustering; the supervised factorization mainly buys accuracy when cross-block covariate correlation distorts the association.","Editorial extension: the paper leaves implicit that a formal selective-inference correction, conditioning on the full selection event rather than on the chosen bases, could convert the exploratory screen into confirmatory inference; until that exists, the method is best used for hypothesis generation.","Editorial extension: the one-fifth-of-dominant-path detection threshold observed in simulation suggests that a minimum-effect-size guideline could be pre-registered in future two-block studies, since weak paths are detectable only when dominant paths give up coefficient mass.","Editorial extension: because negative associations can only be encoded through membership in a different non-negative group, a compositional or count-data variant with a log-ratio or Poisson objective would plausibly change which groups emerge; this is a direct consequence of the modelling choice rather than of the factorization itself."],"fun_headline_variants":["Tri-factor regression matrix maps response-covariate groups","Non-negative RRR co-clusters both variable blocks","One trick: factorize coefficient matrix to pair variable groups","Co-cluster responses and covariates via tri-NMF on regression","Low-rank regression, parts-based: groups on both sides"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a significant entry of Θ can be read as a tested group-to-group link even though the test is conditional on bases and ranks estimated from the same data and, when Q < R, the covariate-side split is not uniquely identified.","fun_headline_variants_meta":{"raw":{"variants":["Tri-factor regression matrix maps response-covariate groups","Non-negative RRR co-clusters both variable blocks","One trick: factorize coefficient matrix to pair variable groups","Co-cluster responses and covariates via tri-NMF on regression","Low-rank regression, parts-based: groups on both sides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":2872,"prompt_tokens":900,"completion_tokens":1972,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1902}},"tokens_in":644,"tokens_out":1972,"duration_ms":11675,"temperature":1.0,"reasoning_tokens":1902,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:16:22.517932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full pipeline—rank selection by cross-validation, basis re-estimation, conditional Wald test—on data generated from the paper's identifiable Q=R=3 null design, but count every replication, including those where cross-validation fails to recover the true ranks. The paper reports sizes only after conditioning on correct rank recovery; an unconditional rejection rate above 5% would show that rank selection adds over-rejection that the conditioning hides, while a rate at or below nominal would support the conservative-size claim.","supporting_citations":[],"review_version":1}