{"id":"0ee88bd0-d2b7-4a49-986b-e7ec65248dc1","arxiv_id":"2608.09162","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A preprocessing framework that reallocates transformed-space width to minimize the target function's Dirichlet energy shows aggregate gains on tabular benchmarks, with the supervised variant best in regression.","lead":"The paper introduces \"stretch transformations\", a new way to preprocess numeric features for tabular deep learning by explicitly making the target function smoother. Experiments across 38 benchmark datasets suggest the supervised variant improves regression performance, though the claim that it \"consistently outperforms all baselines\" is stronger than the paper's own tables show.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Supervised stretch does not minimize its stated Dirichlet energy objective: the Cauchy-Schwarz surrogate (Eqs. 13-15) minimizes a different functional, so the theoretical justification for w_t ∝ S_t is unsupported.","rationale":"The reader's weakest assumption targeted exactly this surrogate gap, and my analysis confirms it is the most load-bearing weakness. The paper's headline claim has two components: (1) supervised stretch consistently outperforms baselines, and (2) this success stems from minimizing the target function's Dirichlet energy. The empirical component is also overstated—Table 1 shows PLE beats supervised stretch on MLP (0.6835 vs 0.6343) and unsupervised stretch beats it on MLP-PLR (0.7347 vs 0.6238)—but that could be softened by rewording. The theoretical component is more fundamental: the derivation in Section 3.4 replaces the exact objective (Eq. 29, derived from the piecewise-linear map in Eq. 2) with a lower-bound surrogate (Eq. 14) and then presents the minimizer of the surrogate as the solution. Since equality in the Cauchy-Schwarz step requires Δy_i ∝ ||Δf_i||, while the piecewise-linear map forces Δy_i ∝ Δx_i, the surrogate is not tight under the transformation family. Minimizing a lower bound does not minimize the original energy, so the widths w_t ∝ S_t are not optimal for the stated objective. The appendix acknowledges the surrogate status but the abstract and main text claim Dirichlet energy minimization without this caveat. This is an internal inconsistency, not a disagreement with consensus: the math is checkable and shows the claim as stated is false. The paper could be rehabilitated by recharacterizing supervised stretch as a heuristic and providing direct evidence that the surrogate correlates with downstream performance, but in its current form the central theoretical contribution is unsupported, and the empirical overclaim compounds the problem. Therefore the verdict should move from CONDITIONAL to REJECT, or be made conditional on a substantial revision that either proves the surrogate tracks the true optimum or abandons the Dirichlet-minimization interpretation.","tokens_in":24444,"tokens_out":5912,"duration_ms":56369,"concrete_test":"Construct a synthetic 1D regression problem with a known, non-linear f(x). Compute three width vectors: (i) w_t ∝ S_t as in Eq. 15; (ii) the exact minimizer of the true Dirichlet energy under the piecewise-linear map, w_t ∝ sqrt((b_t-b_{t-1}) · Σ_{i∈I_t} ||Δf_i||^2/Δx_i); (iii) uniform widths. For each, compute the actual discrete Dirichlet energy (Eq. 7) and the resulting neural network test error. If widths (i) do not yield the lowest energy and are not clearly better than uniform, the surrogate is not tracking the stated objective. Report the ratio of energy under (i) to energy under (ii) across datasets; if it exceeds about 1.1, the theoretical justification fails and the empirical gains must be attributed to another mechanism.","verdict_should_be":"REJECT","load_bearing_attack":"Under the piecewise-linear stretch map, within-bin spacings are Δy_i = (w_t/(b_t-b_{t-1})) Δx_i, so the exact bin-level contribution to the discrete Dirichlet energy (Eq. 7) is E_t = (b_t-b_{t-1})/w_t · Σ_{i∈I_t} ||Δf_i||^2/Δx_i (Eq. 29). Minimizing Σ_t E_t subject to Σ w_t = 1 gives w_t ∝ sqrt((b_t-b_{t-1}) · Σ_{i∈I_t} ||Δf_i||^2/Δx_i), not w_t ∝ S_t. The paper instead minimizes the Cauchy-Schwarz lower bound Σ_t S_t^2/w_t (Eq. 14), which yields w_t ∝ S_t. Equality in the bound would require Δy_i ∝ ||Δf_i||, but the piecewise-linear geometry forces Δy_i ∝ Δx_i, so the bound is generally loose and the constraints are incompatible unless the feature spacings align with target variation. Consequently, the widths delivered by Eq. 15 are not minimizers of the paper's own objective (Eq. 6/7) under the stated transformation family. The central theoretical claim that supervised stretch optimizes target-function smoothness by minimizing Dirichlet energy is therefore not established; the method minimizes a heuristic surrogate that can diverge substantially from the true energy (e.g., two bins with equal S_t but different within-bin sample density receive equal widths under Eq. 15, whereas the exact optimizer allocates more width to the denser bin). The theoretical contribution collapses to an unsupported heuristic, and the empirical gains cannot be attributed to the proposed mechanism without direct evidence that the surrogate tracks the true objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces \"stretch transformations\" for numeric tabular features. Unsupervised stretch is a piecewise-linear monotone map with uniform bin widths, derived from a minimax worst-case Dirichlet energy argument, and is claimed to be a memory-efficient alternative to PLE and an approximation to the empirical CDF. Supervised stretch allocates bin widths proportionally to the bin-wise total variation of an out-of-fold estimate of the target function, motivated by minimizing a lower bound on the target's discrete Dirichlet energy in the transformed space. The authors prove formal connections to PLE, CDF transformation, and target encoding in limiting cases, and report experiments on 38 TALENT datasets with five neural architectures, concluding that supervised stretch consistently outperforms all baselines, especially in regression.","tokens_in":24872,"tokens_out":8232,"duration_ms":75193,"significance":"The empirical study is extensive (38 datasets, 5 models, 100 Optuna trials, 15 seeds) and the out-of-fold estimation protocol is a thoughtful attempt to avoid label leakage. If the empirical results survive correction of the overclaim in the abstract, supervised stretch would be a useful addition to the tabular preprocessing toolbox. The theoretical connections to PLE and CDF (unsupervised) are elegant and largely correct. However, the central mechanistic claim—that supervised stretch minimizes the Dirichlet energy of the target function—is not established, because the method optimizes a surrogate lower bound that can deviate from the exact objective. The paper's contribution is therefore better described as an empirically driven heuristic with a plausible smoothness rationale, rather than a method derived from its stated optimization problem.","major_comments":[{"comment":"The paper claims to minimize the discrete Dirichlet energy (Eq. 7) under the piecewise-linear stretch map, but the width allocation w_t ∝ S_t (Eq. 15) is derived from the Cauchy-Schwarz lower bound (Eq. 14), not from the exact objective. Under the piecewise-linear map, the exact within-bin energy is E_t = ((b_t-b_{t-1})/w_t) Σ_{i∈I_t} ||Δf_i||^2/Δx_i (Eq. 29), and minimizing Σ_t E_t over w_t gives w_t ∝ sqrt((b_t-b_{t-1}) Σ_{i∈I_t} ||Δf_i||^2/Δx_i). These two objectives can disagree: two bins with identical S_t but different within-bin sample densities receive equal widths under Eq. 15, while the exact optimizer gives more width to the denser bin. Since Appendix A.3 explicitly calls Eq. (14) a \"robust lower-bound surrogate,\" the abstract and Section 4.2 (Finding 1) overstate the theoretical contribution when they attribute the empirical gains to \"minimizing the Dirichlet energy.\" The paper should either provide empirical evidence that the surrogate tracks the exact energy on representative data, or reframe the theoretical claim as a heuristic width allocation.","section":"Section 3.4 / Appendix A.3"},{"comment":"The claim that supervised stretch \"consistently outperforms all baselines\" is not supported by Table 1. For MLP, the Sup. Overall Score is 0.6343, below PLE (0.6835), RS-SC (0.6653), and Standard (0.6423); for MLP-PLR, Sup. (0.6238) is below Unsup. (0.7347) and Standard (0.6443); and for ResNet Avg. R^2, Sup. (0.651) is below Minmax (0.655). The paper should replace \"consistently outperforms\" with a precise, qualified statement (e.g., strongest aggregate performance in regression, top-tier in classification) or provide a formal meta-analysis across the 190 dataset-model combinations that quantifies how often supervised stretch wins.","section":"Abstract / Section 4.2 / Table 1"},{"comment":"The claim that unsupervised stretch requires only O(1) memory per feature instead of O(T) is incorrect: a piecewise-linear map with T bins stores T+1 boundary values and T widths (or equivalently T slopes), which is O(T) memory per feature. The O(1) quantity is the output dimensionality (a scalar) compared to PLE's T-dimensional encoding, which is a different statement. This should be corrected in the contribution list and in Section 4.2, Finding 2.","section":"Sections 1, 3.3, 4.2"},{"comment":"The transformation definition uses the dataset range [x_min, x_max] and quantile binning to set boundaries, and the evaluation protocol says the test split is used only after configuration selection. It should be stated explicitly that all transformation parameters (bin boundaries, bin widths, and the out-of-fold regression estimates used for supervised stretch) are fit on the training split only and then applied unchanged to validation and test splits. Without this clarification, it is difficult to rule out feature/target leakage in the reported results.","section":"Section 3.1 / Section 4.1"}],"minor_comments":[{"comment":"The phrase \"thestretch transformation framework\" is a typo; it should read \"the stretch transformation framework.\"","section":"Abstract"},{"comment":"The sentence beginning \"A feature transformations\" should be singular (\"A feature transformation\"), and the following sentence beginning \"but assumes\" should start with a capital letter.","section":"Section 2"},{"comment":"The minimax derivation introduces the assumption ||Δf_i||^2 ≤ C without discussing how C is chosen or why the resulting allocation is independent of C; a sentence stating that C cancels out would clarify the argument.","section":"Section 3.3"},{"comment":"The connection to target encoding via Eq. (16) is informal; the \"remarkably similar\" relationship to min-max scaled target encoding should be made precise, since for non-monotonic target functions the mapped coordinates are cumulative target variation, not the target values themselves.","section":"Section 3.4"},{"comment":"The integration-by-parts step in Eq. (18) assumes Neumann boundary conditions g'(0)=g'(1)=0; these boundary conditions should be stated in the main text where the Dirichlet energy is introduced.","section":"Appendix A.1"},{"comment":"Some dataset names appear to contain typos (e.g., \"E-CommereShippingData\", \"Biascorrectionr\", \"polreg\"), which should be checked against the source benchmark.","section":"Table 4"},{"comment":"The PLE fallback scheme replaces PLE with standardization on some dataset-model combinations; the paper should list which combinations are affected, because those cells no longer evaluate PLE.","section":"Appendix B.2"},{"comment":"The caption and text refer to \"the adaptive threshold used in Table 1\" but the figure sweeps fixed thresholds; the relationship would be clearer if the adaptive threshold value were marked on the plot.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely salvageable through major revision. The central theoretical gap (surrogate vs. exact Dirichlet energy) and the empirical overclaim in the abstract both need to be addressed; the authors should either provide experiments measuring the gap between the surrogate and the exact objective, or reframe the contribution as an empirical heuristic. The O(1) memory claim must be corrected, and the training/test fitting procedure should be stated precisely. The experimental protocol is otherwise strong, and the concept is interesting enough to warrant a revised submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the stretch transformation paper. The short version: the empirical study is a real contribution, and the arc-length equivalence between unsupervised stretch and PLE is genuinely nice. But the theoretical story is softer than the abstract makes it sound, and the headline claim of \"consistently outperforming all baselines\" is not supported by the paper's own Table 1.\n\nWhat's new and good: framing numeric preprocessing as width allocation over a piecewise monotone map is a clean idea. The supervised variant uses target information in a principled way, with out-of-fold Nadaraya-Watson estimates to avoid direct leakage. The experiments are thorough: 38 datasets, 5 architectures, 100 Optuna trials, 15 seeds, with a careful significance-filtering protocol. The regression gains are consistent and non-trivial, and the limiting connection to target encoding is a useful intuition. Unsupervised stretch with O(1) memory per feature and its arc-length equivalence to PLE is a solid practical result.\n\nThe soft spots: (1) The central derivation. Under the piecewise-linear map, the exact bin-level Dirichlet contribution is E_t = (b_t - b_{t-1})/w_t * sum ||Δf_i||^2/Δx_i, and minimizing the sum gives w_t ∝ sqrt((b_t-b_{t-1}) * sum ||Δf_i||^2/Δx_i), not w_t ∝ S_t. The paper instead minimizes a Cauchy-Schwarz lower bound, and while Appendix A.3 is transparent about this, the main text presents the formula as if it directly minimized the Dirichlet energy. So the theoretical justification for the specific width rule is not established; at best it's a motivated heuristic. That doesn't kill the paper—the heuristic can still work—but it should be framed that way. (2) The \"consistently outperforms all baselines\" claim is not true. In Table 1, unsupervised stretch scores higher than supervised on MLP-PLR, and on MLP the standard and RS-SC baselines are competitive or better in some rows. Aggregate win rates are favorable, especially for regression, but the absolute claim needs toning down. (3) A minor gap: the unsupervised minimax solution gives equal Δy_i, which the piecewise-linear map with quantile bins achieves only in the T→n limit; the paper hand-waves the realizability step.\n\nWho this is for: people working on tabular preprocessing will get value from the method and the empirical comparison. The stress-test note you shared is largely correct, but I don't think it's a load-bearing flaw because the paper itself concedes the surrogate. The bigger problem is the overclaiming in the abstract.\n\nRecommendation: send it to peer review. A serious referee can push for a major revision on framing and theory, but the experiments alone justify referee time.","headline":"A useful empirical study and a clean geometric observation, but the Dirichlet-energy theory is really a lower-bound surrogate in disguise and the abstract's 'consistently outperforms' claim overshoots the data.","tokens_in":25356,"tokens_out":4954,"would_cite":true,"duration_ms":46380,"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 paper claims that numeric tabular features should be preprocessed to make the target function smoother, and that the resulting supervised stretch transformation consistently outperforms all baselines across 38 benchmark datasets.","keywords":["tabular data","feature transformation","Dirichlet energy","supervised preprocessing","piecewise linear encoding","target encoding","spectral bias","regression"],"falsifier":"On a real dataset, compute the exact discrete Dirichlet energy E_disc(w) = sum_t (b_t-b_{t-1})/w_t * sum_{i in I_t} ||Δf_i||$_2^{2}$/Δx_i under the piecewise-linear stretch map for a grid of width vectors w, and compare its minimizer to the surrogate-optimal w_t* = S_t/sum S_u. If for many features the exact minimizer differs substantially from w_t* while the surrogate energy is lower at w_t*, then the lower-bound surrogate is not faithfully optimizing the paper's own smoothness objective, even if the empirical gains remain.","tokens_in":24268,"feed_emoji":"📈","tokens_out":5303,"duration_ms":50024,"temperature":0.7,"pith_summary":"Most numeric preprocessing for tabular data is unsupervised: it reshapes feature distributions without looking at the target. This paper argues that the right goal is instead to make the target function easier for a neural network to learn, and that this reduces to allocating transformed-space width across bins of each feature. Its supervised stretch method minimizes a discrete Dirichlet energy of the target in the transformed space, giving a closed-form width per bin proportional to the bin's total target variation. The paper reports that on 38 TALENT datasets, across five architectures, supervised stretch consistently outperforms all seven baselines, with the largest gains in regression. If true, this establishes target-function smoothness as a design principle for tabular preprocessing rather than an incidental byproduct of distribution matching.","feed_headline":"Supervised stretch beats all baselines on 38 tabular datasets","feed_subtitle":"Reshaping numeric features to flatten the target's Dirichlet energy, with the biggest gains in regression.","key_machinery":"The central object is the stretch transformation, a strictly increasing, piecewise-linear map s(x)=c_{t-1}+(x-b_{t-1})/(b_t-b_{t-1}) w_t that sends each numeric feature into [0,1] while preserving order; its design problem is the width vector {w_t}. The argument is carried by the discrete Dirichlet energy E_disc = sum_i ||Δf_i||$_2^{2}$ / Δy_i, which measures how much the target changes per unit of transformed space. Because substituting the piecewise-linear map into this energy is numerically unstable, the paper replaces each bin's within-bin contribution by the Cauchy-Schwarz lower bound $S_t^{2}$/w_t, turning the width optimization into the convex problem min sum $S_t^{2}$/w_t with sum w_t=1, whose closed-form solution is w_t* = S_t / sum_u S_u. This identity, width proportional to bin-wise total target variation, is what makes the method tractable and is the mechanism behind all theoretical connections and empirical gains.","core_discovery":"The central claim is that the learnability of a numeric feature for a neural network is governed by the smoothness of the target function in the transformed feature space, and that smoothness can be optimized directly by a monotone piecewise-linear 'stretch' map. For a feature partitioned into T quantile bins, the map assigns each bin a width w_t. Supervised stretch chooses widths to minimize the discrete Dirichlet energy of the target's conditional expectation; because the exact piecewise-linear objective is numerically unstable, the paper minimizes a Cauchy-Schwarz lower bound and obtains w_t* = S_t / sum_u S_u, where S_t is the total variation of the estimated target within bin t. This concentrates resolution where the target varies rapidly, equalizing slope magnitudes across bins. In the unsupervised case, minimax reasoning over bounded-variation targets yields uniform widths, recovering a piecewise-linear approximation of the empirical CDF and, by arc-length parameterization, an equivalent of Piecewise Linear Encoding. The paper further shows that in the fine-binning limit supervised stretch approaches target encoding, giving a theoretical explanation for why target-based transformations work.","pith_inferences":["Testable extension: learn the widths w_t jointly with the network parameters instead of fixing them before training; the smoothness argument suggests differentiable width optimization could improve on the closed-form allocation, since the current method optimizes preprocessing independently of the learned function.","The spectral-bias explanation implies an interaction with network scale: supervised stretch's gains should be larger for small or undertrained networks that have difficulty with high-frequency components, and shrink as capacity and training time grow; this is measurable with the paper's protocol.","The method is marginal, so a joint version that allocates width per feature while accounting for other features' target variation could address cases where the target changes mostly through interactions; the paper explicitly leaves interaction-awareness out of scope.","The Dirichlet-energy objective applies to any continuous target, so supervised stretch should transfer to multi-output regression, survival targets, or quantile targets without changing the formula; this is a direct corollary of the vector-valued target treatment noted in the paper."],"forward_implications":["If supervised stretch is right, target-aware preprocessing is a repeatable performance lever: across 38 datasets and five architectures it beats all seven baselines, and on regression tasks the margin is decisive.","Unsupervised stretch gives a target-free method that matches or exceeds piecewise linear encoding while using O(1) memory per feature and no dimensional expansion, so the benefit of density reshaping does not require PLE's computational cost.","The framework provides explanations for existing empirical findings: uniform stretch approximates empirical CDF transformation as T grows, explaining why CDF preprocessing reduces high-frequency content; supervised stretch approaches target encoding in the fine-binning limit, explaining why target-based transforms help.","The bin count T acts as a regularization parameter: small T preserves distributional structure, large T approaches full CDF (unsupervised) or target-encoding-like behavior (supervised), giving a controlled interpolation between extremes.","Because bin widths are proportional to bin-wise total target variation, supervised stretch tends to equalize the target function's slope magnitudes across bins, which is the concrete sense in which the learned function becomes smoother."],"supporting_citations":[{"why":"Supplies Piecewise Linear Encoding and MLP-PLR, the main unsupervised baseline and the shared piecewise-linear geometry that unsupervised stretch is compared against.","marker":"[10]"},{"why":"The frequency-principle result that neural networks learn low-frequency functions first is the justification for why smoothing the target function should improve learnability.","marker":"[34]"},{"why":"Documents spectral bias in neural networks, which the paper invokes to motivate Dirichlet energy minimization as a preprocessing objective.","marker":"[28]"},{"why":"The observed claim that CDF transformation reduces frequency content in tabular data is the empirical phenomenon that unsupervised stretch's theory explains.","marker":"[3]"},{"why":"Provides the per-(dataset, model) min-max normalization used in the aggregation protocol and the discussion of tabular feature heterogeneity that motivates the framework.","marker":"[12]"},{"why":"The TALENT benchmark supplies the 38 datasets, train-validation-test partitions, and model hyperparameter search protocol that ground the empirical comparison.","marker":"[35, 23]"},{"why":"Nadaraya-Watson kernel regression is the out-of-fold estimator used to obtain target estimates for supervised stretch's bin-width computation.","marker":"[25, 32]"},{"why":"RealMLP and its RobustScale+SmoothClip preprocessing serve as one of the five architectures and a baseline transformation in the controlled comparison.","marker":"[17]"}],"fun_headline_variants":["Supervised stretch: smooth targets beat tabular baselines on 38 sets","Stretch for smoothness: Dirichlet energy wins tabular learning","Optimize feature stretch, flatten target energy, dominate 38 benchmarks","Smoothness via stretch: new tabular preprocessing beats all baselines","Target-smoothness stretch: top 38 tabular datasets with one trick"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Cauchy-Schwarz lower bound is a faithful stand-in for the true Dirichlet energy of the piecewise-linear transformation, even though the true objective is numerically unstable and the surrogate ignores the within-bin constant-slope constraint.","fun_headline_variants_meta":{"raw":{"variants":["Supervised stretch: smooth targets beat tabular baselines on 38 sets","Stretch for smoothness: Dirichlet energy wins tabular learning","Optimize feature stretch, flatten target energy, dominate 38 benchmarks","Smoothness via stretch: new tabular preprocessing beats all baselines","Target-smoothness stretch: top 38 tabular datasets with one trick"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3030,"prompt_tokens":991,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1943}},"tokens_in":607,"tokens_out":2039,"duration_ms":15168,"temperature":1.0,"reasoning_tokens":1943,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:18:14.877717+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a real dataset, compute the exact discrete Dirichlet energy E_disc(w) = sum_t (b_t-b_{t-1})/w_t * sum_{i in I_t} ||Δf_i||$_2^{2}$/Δx_i under the piecewise-linear stretch map for a grid of width vectors w, and compare its minimizer to the surrogate-optimal w_t* = S_t/sum S_u. If for many features the exact minimizer differs substantially from w_t* while the surrogate energy is lower at w_t*, then the lower-bound surrogate is not faithfully optimizing the paper's own smoothness objective, even if the empirical gains remain.","supporting_citations":[{"cited_title":"On embeddings for numerical features in tabular deep learning","cited_arxiv_id":null,"evidence_quote":"Supplies Piecewise Linear Encoding and MLP-PLR, the main unsupervised baseline and the shared piecewise-linear geometry that unsupervised stretch is compared against."},{"cited_title":"An inductive bias for tabular deep learning","cited_arxiv_id":null,"evidence_quote":"The observed claim that CDF transformation reduces frequency content in tabular data is the empirical phenomenon that unsupervised stretch's theory explains."},{"cited_title":"Why do tree-based models still out- perform deep learning on typical tabular data?Advances in neural information processing systems, 35:507–520, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the per-(dataset, model) min-max normalization used in the aggregation protocol and the discussion of tabular feature heterogeneity that motivates the framework."},{"cited_title":"Better by default: Strong pre-tuned MLPs and boosted trees on tabular data","cited_arxiv_id":null,"evidence_quote":"RealMLP and its RobustScale+SmoothClip preprocessing serve as one of the five architectures and a baseline transformation in the controlled comparison."}],"review_version":1}