acceptodds
Under review as a conference paper at ICLR 2027

When More Data Is Not Enough: Identifying and Repairing Support Gaps in Mechanism-Resolved Forecasting

Abstract

Accurate marginal prediction does not necessarily determine the predictive law for a specified combination of factors. We quantify this ambiguity with the Support-Induced Composition Radius (SICR), the minimax information radius of target laws compatible with the population training experiment. Positive SICR gives an architecture-independent lower bound on worst-case target excess log loss at every sample size. Within a specified interaction family, we prove that SICR vanishes exactly when the target interaction feature lies in the training row span, under composition-wise context overlap and anchored reference and atomic effects. This characterization yields target-span acquisition (TSA), a greedy rule that selects compositions to reduce the target residual outside the training span. Controlled experiments show a persistent error plateau when sampling from the same support and its removal when the span is expanded. Across 200 six-factor design tasks, three TSA additions reduce the mean normalized squared target residual to 0.0023, compared with 0.5057 for regularized D-optimal design. In a retrospective UCI Bike Sharing audit, TSA reduces target RMSLE averaged across four predictors by 11.4% relative to the best non-targeted policy at 200 added records. An omitted-interaction stress test evaluates the certificate's sensitivity to the chosen interaction basis. Together, these results provide a target-specific workflow for diagnosing support gaps, choosing informative additions, and checking the structural assumptions behind extrapolation.

Primary area:
learning theory
Keywords:
identifiability, compositional generalization, support gaps, forecasting, experimental design, partial identification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YWELOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper is situated in the field of forecasting. Rephrased in ML terminolgy, the paper tackes the issue of OOD prediction; that is, one may query the prediction model for datapoints that lie outside the observed distribution and there the model may be inaccurate. The paper quantifies the ambiguity of the OOD predictions using a measure they call Support Indiced Composition Radius (something like model multiplicity). Then the authors continue to show that when the underlying true function is additive we might be able to extrapolate well when each individual feature has full support (although some combinations of features may not have full support). This can be exploited to query data points that reduce faulty OOD prediction.

Critical strengths

OOD prediction is an important problem, I suppose especially so in forecasting.

Critical weaknesses
  • The main weakness for me is the writing. For example, the paper introduces and uses lots of terminology that I am not aware of and also had trouble understanding when asking LLMs about it. Some examples from the introduction:
    • Many queries "[...] specify a combination of factors, regimes, or control inputs zz". What are factors, regimes, and control outputs? Are factors features as in machine learning? Are control inputs interventional queries do(x)do(x)? What are regimes?
    • What is a "marginal law" and what is a "target law"?
    • What are "atoms" in this context and what are "atomic effects"? Even reading Section 2 did not clarify these terms for me.
  • I am unsure about the novely of the findings in the paper. OOD generalization, model multiplicity, rashomon sets, causal interpretation of purely predictive models are all large literatures that this work directly connects to (although without using their terminology).
  • The application on the bike sharing dataset is pretty limited and fails to convince me that the approach is generally unseful.
Questions for authors

How does the contribution compare to work in Rashomon sets, model multiplicity, OOD generalization, causal interpretation of predictive models? Could the authors try to rephrase the central findings in the machine learning terminology that is common at ICLR?

4 comments

Thank you for the comments on clarity and the connection to familiar machine-learning concepts. We have rewritten the opening around a concrete example and defined the terminology at its first use. The central question is: does the training distribution uniquely determine the predictive distribution at a specified, possibly unseen combination of conditions?

The question in ordinary prediction terms. H is the observed context, Y is the outcome, and Z records a declared combination of factors. A factor is one coordinate of Z; a composition is their joint setting. For example, let z=(year 2012, rush hour, adverse weather), with target 111. These are observed operating conditions. The prediction target is an observational conditional distribution; causal effects would require additional assumptions.

Marginal forecasting predicts p(Y∣H)p(Y\mid H), averaging over the observed mixture of conditions. Mechanism-resolved forecasting predicts p(Y∣H,z⋆)p(Y\mid H,z^\star) for a specified composition. Accurate estimation of the first need not determine the second.

Figure 1 illustrates the distinction in a pairwise interaction family. Training contains 000, 100, 010, 001, 110, and 101, so the interaction z2z3z_2z_3 is never activated. Two models can agree on all these training laws, the baseline at 000, and every single-factor distributional change, yet disagree at 111. Those single-factor changes are the atomic effects. Holding them and the factor meanings fixed defines the atom-preserving comparison. Under the stated reference and overlap assumptions, adding 011 to the training support supplies the missing interaction direction and identifies the target law while 111 remains unobserved.

What the theory contributes. SICR quantifies the disagreement among target distributions compatible with the same population training experiment and semantic anchors. Theorem 1 gives its worst-case excess-log-loss consequence for every training-only learner at every sample size. Appendix A.5 establishes an exact minimax value for a fixed equivalence class. Theorem 2 gives a necessary and sufficient interaction-span condition within the anchored family, and TSA turns the uncovered target direction into a support-selection rule.

The model permits interactions rather than only additive raw-factor effects. The pairwise example uses χ(z)=(z1z2,z1z3,z2z3)\chi(z)=(z_1z_2,z_1z_3,z_2z_3); the theorem also permits higher-order known finite-dimensional maps satisfying its assumptions. Section 5 and Appendix A.6 state the reference, context-overlap, and local-admissibility conditions together.

Relation to model multiplicity, OOD prediction, and prior identification results. Rashomon sets and predictive multiplicity study disagreement among predictors with similar loss (Fisher et al., 2019; Marx et al., 2020). Here, the compared models agree on the entire observable training law, so additional data from that unchanged experiment cannot distinguish them. The question is partial identification of an OOD conditional distribution under a declared model family. This asks what information the training experiment supplies before a prediction architecture is chosen.

Classical estimability provides the row-space algebra. Bravo-Hermsdorff et al. (2023) study full-distribution reconstruction through PR transformations and discuss training-condition design. Our anchored-family construction connects explicit training-equivalent target disagreement to minimax prediction risk and target-specific support repair. Section 8 and Appendix A.8 set out these relationships and the respective modeling assumptions.

The controlled experiments show the predicted error plateau and the equal-budget benefit of support expansion. The 200-task study examines multiple repair choices, and the original Bike comparison reports 11.4% and 10.2% mean target-RMSLE improvements at 200 and 400 added records relative to the strongest non-targeted policy. The accompanying update summarizes these findings and the new precision and context-support analyses.

Thank you for asking about the practical value of the framework. The revised experiments connect the identification question to observable prediction outcomes, using controlled support changes, multiple acquisition choices, and the Bike Sharing evaluation.

Changing the information available to the learner. The controlled experiments show why sample count and support answer different questions. At 200 trajectories per represented composition, the forcing estimator has target regret 0.2247 on ambiguous support and 0.0051 with bridge 011 (Appendix C.1). Marginal regret falls from 2.1675 to 0.0095 while the target law remains unidentified (Appendix C.2). When total observation budgets are matched, targeted support expansion reduces target regret from 0.2286 to 0.0263 at 384 added trajectories, in close agreement with the analytic comparison (Appendix C.3; Figure 2). This isolates the contribution of acquiring a missing interaction direction.

Choosing among multiple repairs and allocating for precision. Across 200 six-factor tasks, three TSA additions reduce mean normalized squared target residual to 0.0023, compared with 0.5057 for D-optimal design, 0.5047 for maximum-residual selection, and 0.6257 for random selection. TSA covers 98% of targets after three additions and all targets after four (Appendix C.4).

The new noisy study compares seven policies at equal observation budgets with the same minimum-norm predictor. TSA again completes target coverage by four observations. At noise standard deviation 0.5 and 48 added observations, following coverage with c-optimal allocation gives target-mean MSE 0.0227, compared with 0.2021 for uniform continuation and 0.0226 for middle-regularization c-optimal alone (Appendix F.1; Table 2). The full curves quantify how estimation variance and allocation affect prediction alongside structural coverage. These MSEs are noise-integrated averages over coefficient draws and tasks; SICR measures worst-case target-law ambiguity.

Predictive gains and practical checks on Bike Sharing. In the original full-feature retrospective comparison, TSA achieves mean target RMSLE 0.619 versus 0.699 for random-pool acquisition at 200 added records, an 11.4% reduction. At 400 added records, the reduction is 10.2% relative to the strongest non-targeted policy. At budget 200, improvements over same-span acquisition are 15.2% for Ridge, 19.6% for random forest, 5.6% for HistGradientBoosting, and 1.3% for MLP. The predictor-specific curves, including the MLP reversal at budget 400, remain reported in Figure 7.

The gain over same-span acquisition also holds at budget 200 under all three rush-hour definitions (Appendix E.3). In the tested three-dimensional setup, the geometric audit takes 57.8 microseconds, compared with 0.012-0.808 seconds for model fitting (Table 1).

A separate common-context audit uses four temperature/humidity cells to examine the overlap issue raised by the full predictor features. Bridge acquisition raises the full design rank from six to seven but does not improve either predictor under this coarser context definition (Appendix F.2; Table 3). This result distinguishes empirical prediction from a structural certificate, whose application requires population overlap and a correctly specified interaction family. The real-data evidence comes from one retrospective dataset; the controlled multi-candidate study and the shared-context audit address complementary questions within that scope.

Thank you for the clarifications which helped me to understand the paper‘s contribution.

Based on my understanding the novelty of the contribution is relatively small, although it is sound. Moreover my main criticism is the writing; I believe that its assessment deserves another round of reviews.

As such, I leave my score unchanged.

Thank you for revisiting our response and for recognizing the technical soundness of the work. We appreciate your feedback on novelty and writing.

Regarding novelty, Section 8 and Appendix A.8 clarify how our contribution differs from classical estimability, sequential c-optimal design, and prior compositional identification results: we connect a necessary-and-sufficient criterion for identifying a specified target predictive law to the Support-Induced Composition Radius (SICR) and its sample-size-independent minimax risk consequence.

As writing was your main concern, we would also like to highlight the changes in the revised PDF uploaded to OpenReview. We rewrote the Introduction around an intuitive three-factor example, clarified key terminology at first use, improved the explanation accompanying Figure 1, organized the theorem assumptions by proof direction, and revised the related work discussion to make the contribution and its relationship to prior work clearer.If any sections remain unclear, we would welcome specific pointers and will address them in the final version.

We appreciate your time and consideration of these revisions.

Reviewer_YPMYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the metric of SICR, which quantifies the ambiguity within target laws induced by a training experiment, with finer resolution. SICR provides lower bound on worst case excess log loss. The paper also proposes TSA as a corresponding policy. Experiments on both synthetic and real-world datasets are conducted to prove soundness.

Critical strengths

The theory part is well written and presented. The proposed metric has promising properties and provides useful bounds, theoretically. Experiments explain the motivation very well.

Critical weaknesses

I’m not very much convinced by its practical uses on the real-world datasets. Since this paper uses only one real-world dataset, the practical significance of TSA is compromised. Some cross validation on multiple kinds of datasets may be helpful.

That said, I wouldn't describe it as ``critical'' weakness since I think this paper appears more of theoretical flavor to me.

Questions for authors

What exactly are the assumptions for Theorem 2 (in a rigorous sense)? Also, currently Theorem 2 still seems quite restrictive to a specified family of interactions. Is it possible to extend it to a broader family? What happens when the interaction model is learned incorrectly in practice?

4 comments

Thank you for the positive assessment of the theory and for asking about the assumptions of Theorem 2, broader interaction families, and misspecification. The revision makes the two proof directions explicit and clarifies how the result applies beyond the pairwise example.

Assumptions and the two proof directions. The submitted Section 5 and Appendix A.4 already stated the fixed-reference, composition-wise overlap, and admissible-perturbation conditions. Appendix A.6 now organizes these conditions by their roles in the proof.

The common setup is a finite composition set with fixed factor semantics and a known finite-dimensional interaction map satisfying χ(0)=χ(ek)=0\chi(0)=\chi(e_k)=0. Compared models belong to the same correctly specified anchored family, with a fixed or population-identified reference law and atomic effects at the evaluation contexts. Their conditional distributions are measurable and normalized, with positive densities on a common outcome support. Log-density effects are taken modulo outcome-independent constants, and the sampling experiment is held fixed.

For sufficiency, the evaluation measure must satisfy ν≪Ptr(H∣Z=z)\nu\ll P_{\mathrm{tr}}(H\mid Z=z) for each training composition needed in the span relation. This transfers equality of observed conditional laws to the evaluation contexts. If S contains the training interaction features, two training-equivalent models have effect difference Δ\Delta satisfying SΔ=0S\Delta=0. Writing the target feature as t=χ(z⋆)=S⊤wt=\chi(z^\star)=S^\top w gives

t⊤Δ=w⊤SΔ=0.t^\top\Delta=w^\top S\Delta=0.

With the common reference law, equality in the normalization quotient gives equality of the full target predictive distributions. This direction requires no local-richness assumption and can identify the target even when the complete interaction parameter remains undetermined.

For necessity, the family must admit sufficiently small perturbations generated by a bounded measurable effect that is nonconstant in the outcome on a set of contexts with positive ν\nu-measure. When t lies outside the training span, a vector a∈Null⁡(S)a\in\operatorname{Null}(S) with t⊤a≠0t^\top a\ne0 defines a normalized exponential tilt with exponent ϵχ(z)⊤a,g(y,h)\epsilon\chi(z)^\top a\\,g(y,h). The tilt preserves the training, reference, and atomic laws while changing the target law. The two-world Jensen-Shannon bound then gives positive SICR. Local admissibility ensures that this witness belongs to the specified family.

Extension beyond pairwise interactions. The argument applies to higher-order and other known finite-dimensional anchored feature maps satisfying the same conditions. Pairwise monomials provide the running example; the theorem's scope is the declared interaction family. The object identified is the entire target conditional distribution.

What happens under an incorrect interaction map? Appendix D provides an explicit answer. A three-way term γz1z2z3\gamma z_1z_2z_3 is zero at every non-target binary composition and active at 111. A pairwise extrapolator then has target excess log loss γ2/2\gamma^2/2 in the unit-variance Gaussian example, even after pairwise coverage. Adding the cubic coordinate to the audit correctly leaves the target outside the observed span. Because the omitted term is invisible on non-target support, validation restricted to that support cannot establish its absence.

This example makes the certificate's interpretation concrete: identification is relative to the specified family, and enriching that family can reveal an additional unresolved target direction. The new noisy-design and common-context experiments examine how this identification analysis connects to finite-sample prediction, as detailed in our update.

Thank you for asking how the theoretical conditions bear on practical use. The completed revision adds a proof-direction assumption table, an explicit connection to sequential c-optimal design, and empirical analyses of finite-sample precision and context support.

Controlled evidence for identification and repair. The original experiments isolate the information gap and the benefit of changing support. Marginal prediction becomes accurate, with regret falling from 2.1675 to 0.0095, while the target conditional law remains unidentified (Appendix C.2). Under equal total budgets, targeted support expansion reduces target regret from 0.2286 to 0.0263 at 384 added trajectories, closely matching the analytic comparison (Appendix C.3; Figure 2). Across 200 multi-candidate tasks, TSA reaches mean normalized squared target residual 0.0023 after three additions, versus 0.5057 for D-optimal design, and covers all targets by four additions (Appendix C.4).

Precision after support coverage. The new noisy study uses those same 200 tasks with a common minimum-norm predictor. TSA again covers all targets by four added observations. The analytic MSE decomposition separates the squared contribution of unobserved coefficient directions from estimation variance (Appendix F.1). At noise standard deviation 0.5 and 48 added observations, TSA followed by c-optimal allocation has MSE 0.0227, compared with 0.2021 for uniform continuation and 0.0226 for middle-regularization c-optimal alone. The full curves also show the advantage of variance-aware allocation at smaller budgets. This comparison quantifies the distinct roles of identifying a target and estimating it precisely.

Context support and interaction specification. The original full-feature Bike comparison retains its 11.4% and 10.2% reductions in mean target RMSLE at 200 and 400 added records relative to the strongest non-targeted policy. A separate, outcome-blind protocol defines H using four shared temperature/humidity cells, with identical acquisition quotas and five seeds. Since reference and atomic terms are estimated, the geometric check uses the full seven-column design. Adding 80 bridge records raises its rank from six to seven, although prediction does not improve for either predictor (Appendix F.2; Table 3).

The restricted study evaluates a coarser conditional prediction problem and establishes finite cell occupancy. Population overlap and correct interaction specification remain the conditions for applying Theorem 2. Appendix D complements this audit by showing exactly how an omitted target-only interaction produces positive target error and how an enriched basis detects the remaining support gap.

The real-data evaluation now includes four predictors, three rush-hour definitions, and the restricted-context study within Bike Sharing. These analyses broaden the checks within that dataset; validation on additional real-world datasets remains a further step.

Thank you for addressing my concerns and the extra efforts. I keep my score of 3 here, and I would advocate for acceptance.

Thank you for carefully considering our revisions and for your support. We sincerely appreciate your thoughtful feedback, which helped us clarify the theoretical assumptions and strengthen the empirical evaluation.

Reviewer_mWnFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies when the population training support fails to identify a predictive law at an unseen factor combination. It introduces the Support-Induced Composition Radius (SICR), derives an all-sample-size lower bound on target excess log loss, and characterizes identification through interaction-feature row-span coverage under explicit structural assumptions. The resulting target-span acquisition (TSA) rule selects compositions that reduce the uncovered target direction. Controlled experiments and a retrospective Bike Sharing audit illustrate the value of expanding support. The framework is clear and technically sound, but the methodological novelty and practical validation are somewhat limited.

Critical strengths
  1. The paper carefully distinguishes structural non-identification from estimation error. The assumptions behind the span certificate are explicit, and the omitted-interaction example clearly illustrates how a misspecified basis can falsely suggest identification.

  2. The controlled experiments include equal-budget comparisons that isolate support expansion from additional sampling. The real-data audit excludes outcome leakage, reports predictor-specific results, and acknowledges that its flexible predictors are not directly certified by the theorem.

Critical weaknesses
  1. The conceptual advance over classical estimability and experimental design is modest. The lower bound follows from averaging predictions under a common training law, and the exact certificate uses standard row-space/null-space duality. Connecting these ideas to normalized predictive laws is useful, but TSA is a greedy projection criterion without a guarantee relating residual reduction to finite-sample prediction risk. A comparison with a target-specific design baseline, such as regularized c-optimal design, would better establish its added value beyond the global D-optimal baseline.

  2. The experiments do not yet convincingly validate TSA in a realistic setting with competing repair choices. The 200-task benchmark measures geometric residuals without outcomes or noise. In Bike Sharing, 011 is the only non-target composition that expands the span, so TSA coincides with a simple rank-gain rule. Furthermore, the full covariates contain the variables defining the compositions, and the required composition-wise context overlap is not established. The authors appropriately acknowledge this gap, but the observed prediction gains therefore provide limited validation of the identification certificate or TSA’s distinctive selection advantage.

Questions for authors
  1. How does TSA compare with a regularized target-specific variance criterion on noisy multi-candidate tasks, measuring both target prediction error and support coverage?

  2. Can the authors provide a nontrivial empirical setting where composition-wise context overlap and the interaction model are justified, and multiple candidate compositions offer different repair choices?

4 comments

Thank you for highlighting the links to estimability and optimal design, and for asking how support repair translates into prediction under noise. The revision develops these connections explicitly and adds an equal-budget, noisy multi-candidate comparison.

Contribution and relation to prior work. Our contribution is a target-specific account of what a training experiment determines: a distribution-level identification criterion, its minimax risk consequence, and a rule for acquiring the missing information. We use classical estimability and information-radius theory to connect these three questions. Within the declared anchored interaction family, a missing target direction yields two admissible models with identical training laws, reference distributions, and atomic effects, but different target laws. Conversely, span coverage identifies the target distribution even when the full interaction parameter remains undetermined (Section 5; Appendices A.4 and A.6).

Bravo-Hermsdorff et al. (2023) establish sufficient and almost-everywhere necessary conditions for full-distribution reconstruction through PR transformations and discuss training-condition design. Our result establishes target disagreement through a training-equivalence construction in the anchored family, connects it to any training-only learner's worst-case risk, and derives target-specific acquisition. Section 8 and Appendix A.8 compare the respective model assumptions and conclusions.

An exact fixed-class risk characterization. For a fixed equivalence class and unrestricted measurable predictive kernels, SICR equals the minimax target excess log loss at every sample size (Appendix A.5). Theorem 1 supplies the lower bound; a class-dependent approximate minimax center supplies the reverse bound without requiring an attained center. This characterizes the fixed-class decision problem, while estimating an unknown population class remains a separate task.

Connection to sequential c-optimal design. Let XX contain observation rows, or composition rows weighted by the square roots of their counts. Write G=X⊤XG=X^\top X, t=χ(z⋆)t=\chi(z^\star), x=χ(c)x=\chi(c), and Aλ=G+λIA_\lambda=G+\lambda I. One additional observation at cc decreases the target-variance proxy vλ=t⊤Aλ−1tv_\lambda=t^\top A_\lambda^{-1}t by

\frac{(t^\top A_\lambda^{-1}x)^2}{1+x^\top A_\lambda^{-1}x}.$$ At a fixed uncovered state, $\Delta v_\lambda(c)/v_\lambda$ converges to TSA's squared residual-cosine score as $\lambda\downarrow0$. In-span candidates and orthogonal residuals have limit zero. This identifies TSA as the support-repair limit of a one-step target-variance criterion. Finite regularization and observation counts also affect allocation, so the limit concerns fixed-state scores rather than entire acquisition sequences (Section 5.1; Appendix A.7). **Evidence for support repair.** The original equal-budget experiment reduces target regret from 0.2286 to 0.0263 through support expansion at 384 added trajectories (Appendix C.3). Across 200 six-factor tasks, TSA reaches mean normalized squared target residual 0.0023 after three additions, compared with 0.5057 for D-optimal design, and covers every target by four additions (Appendix C.4). In the original Bike study, TSA reduces mean target RMSLE by 11.4% and 10.2% at 200 and 400 added records relative to the strongest non-targeted policy. This retrospective comparison uses full predictor features and a unique non-target bridge. The added studies address the two further questions in your review: noisy prediction with multiple repair choices, and prediction under a shared-context Bike protocol. We report those results in the accompanying update (Appendices F.1-F.2).

Thank you for the specific questions about noisy target-specific design and realistic repair choices. We have completed both the 200-task observation-budget comparison and a separate common-context Bike audit.

1. Noisy multi-candidate design under equal observation budgets. The study retains all 200 tasks from Appendix C.4. Each of eight initial compositions has five observations; every acquisition adds one observation from a non-target composition, with repeats allowed. All seven policies share the initial data, 20 unscreened coefficient vectors per task drawn from N(0,I15/15)N(0,I_{15}/15), and paired noise pools at standard deviations 0.1, 0.5, and 1.0.

Every policy uses the same minimum-norm least-squares predictor, which sets unidentified coefficient directions to zero. Primary target-mean MSE integrates observation noise exactly for the sampled coefficients; ten actual noise repeats provide a separate check. Intervals resample tasks after within-task averaging. Acquisition regularization is separate from prediction: the three c-optimal penalties are fixed at s00.001,0.1,1s_0\\{0.001,0.1,1\\}, where G0=X0⊤X0G_0=X_0^\top X_0 is the initial information matrix and s0=tr⁡(G0)/15s_0=\operatorname{tr}(G_0)/15. D-optimal and the hybrid use the middle value.

At noise standard deviation 0.5, the prespecified budgets in Table 2 give:

Acquisition policyB=12B=48
TSA then represented-support uniform0.48570.2021
Sequential c-optimal, low0.18820.0729
Sequential c-optimal, middle0.09290.0226
Sequential c-optimal, high0.09660.0302
Regularized D-optimal0.25520.0581
Uniform non-target sampling1.40110.1314
TSA then middle c-optimal0.15120.0227

TSA and low-regularization c-optimal cover all 200 targets by four added observations; middle and high regularization cover 44.0% and 10.5%, respectively (Figure 8). The prediction comparison shows why allocation matters alongside coverage: middle c-optimal achieves lower MSE at B=12, while switching TSA to c-optimal substantially improves its later precision. At B=48, the hybrid and middle c-optimal have observed MSEs of 0.0227 and 0.0226; the paired interval for their difference includes zero. Figures 8-9 and Table 2 report coverage and prediction separately, with all noise levels and regularization settings retained. These average prediction risks complement SICR's worst-case identification analysis.

2. A shared-context audit on real data. We define H through four temperature/humidity cells using non-target medians, excluding the original features that determine the composition labels. Every non-target composition has at least 76 records in each cell. Five acquisition seeds share 120 initial records, identical cell quotas, and equal observation budgets (Appendix F.2).

At 80 added records, the 011 bridge raises the full design rank, including estimated reference and atomic terms, from six to seven. Target RMSLE is higher than with same-span acquisition: 1.2874 versus 0.9537 for pairwise regression, and 0.8481 versus 0.7891 for HistGradientBoosting. Table 3 includes random acquisition and all budgets. Within each seed, target log-count MSE is averaged equally across cells before taking its square root.

This experiment measures geometric coverage and prediction under a common empirical context definition. Coarsening H changes the prediction problem; finite cell occupancy does not establish population overlap or a correct pairwise model. The original full-feature Bike gains remain a separate predictive result.

The workingday fourth-factor check yields a minimum cell count of 23, below the prescribed capacity, and equal TSA scores for its three repair candidates. Thus the controlled experiment answers the noisy multi-candidate question; a real-data comparison with justified interaction structure and distinct feasible repair paths remains the open part of your second question.

I thank the authors for addressing my comments and adding those new analsyis. I will increase my score to reflect this.

Thank you for taking the time to consider our revisions and additional analyses. We greatly appreciate your constructive feedback and your updated assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.