{"id":"71e78be7-db9c-482e-8a01-08d1e816f463","arxiv_id":"1908.06923","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The epsilon-SVQR loss in Eq (5) equals Park and Kim's epsilon-insensitive pinball loss under a tau-dependent rescaling of epsilon, making the model a reparameterization of prior work.","lead":"This paper proposes a fixed-width asymmetric epsilon-insensitive pinball loss for support vector quantile regression and reports better accuracy with sparsity. The proposed loss is algebraically the same as a prior loss by Park and Kim with rescaled epsilon, so the claimed novelty and gains are not supported.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proposed loss (5) is algebraically identical to Park and Kim's Eq (4) under ε_PK = τ(1−τ)ε, so the claimed novelty and reported improvements rest on a reparameterization, not a new model.","rationale":"Read in good faith, the paper aims to construct an asymmetric ε-insensitive zone of fixed width around the regressor and to divide that zone by τ. Direct algebra from the paper's own equations shows that Eq (5) is identical to Eq (4) when ε_PK = τ(1−τ)ε. Since ε is a free hyperparameter, this is not a new loss family and does not extend the ε-insensitive approach in any way not already available in Park and Kim's model. The empirical comparison uses the same numeric ε grid for both methods, which is systematically unfavorable to Park–Kim at extreme quantiles because the corresponding ε_PK values are much smaller than the grid points. Thus the reported improvements are a reparameterization artifact rather than evidence of a new method. This is exactly the reader's weakest assumption, and the conclusion of REJECT remains appropriate. The paper has no independent support that would rescue the novelty claim: the derivations are standard, Proposition 1 fails at ε = 0, and no reproducible code is provided.","tokens_in":18639,"tokens_out":7241,"duration_ms":73447,"concrete_test":"On the AD1 dataset at τ = 0.1, run the proposed Eq (17) and Park–Kim Eq (4) under matched effective insensitivity: for each proposed ε in {0.1, 0.2, ..., 1}, set ε_PK = τ(1−τ)ε and use identical C, q, kernel, training/test split, and QP solver. If the fitted regressors, support sets, dual objectives, and RMSE values agree up to solver tolerance, the two formulations are the same model family. As a decisive secondary check, extend the Park–Kim ε grid to include the values {τ(1−τ)ε : ε in the paper's grid} and recompute the minima in Tables 3, 6, and 7; if the Park–Kim minima drop to the proposed-model values, the reported improvement is a grid artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq (5) is exactly Park and Kim's loss, Eq (4), read as max(0, Pτ(u) − ε), under the substitution ε_PK = τ(1−τ)ε. For u > (1−τ)ε, L^ε_τ(u) = τ(u − (1−τ)ε) = τu − τ(1−τ)ε = τu − ε_PK. For u < −τε, L^ε_τ(u) = −(1−τ)(u + τε) = −(1−τ)u − τ(1−τ)ε = Pτ(u) − ε_PK, since Pτ(u) = −(1−τ)u on that branch. The flat interval [−τε, (1−τ)ε] equals [−ε_PK/(1−τ), ε_PK/τ]. Because ε_PK is a free user parameter, the two loss families coincide; the only difference is how the tube width is parameterized. The paper's argument that Park and Kim's width varies with τ is an artifact of fixing a numeric ε across τ values, rather than treating ε_PK as a tunable parameter. Consequently, the numerical comparisons in Tables 3, 6, 7, and 11, which give both methods the same ε grid, are not comparisons of different model families. For τ = 0.1, proposed ε = 1 corresponds to Park–Kim ε_PK = 0.09, which the Park–Kim grid was not allowed to use, so the apparent advantage of the proposed method reflects a grid mismatch. The dual problem in Eq (29) is the Park–Kim dual under this reparameterization. Proposition 1 is also invalid at ε = 0: both multipliers can be positive when a residual lies exactly on the boundary. The central claim that this is a true extension of the ε-insensitive idea to SVQR is therefore false.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an asymmetric ε-insensitive pinball loss function, L^ε_τ(u) = max(−(1−τ)(u+τε), 0, τ(u−(1−τ)ε)), for support vector quantile regression, and names the resulting model ε-SVQR. It derives the primal and dual quadratic programs, states a sparsity property via Proposition 1, and reports experiments on artificial and UCI datasets comparing ε-SVQR with standard SVQR, Sparse SVQR (Seok et al.), and Park-Kim SVQR. The abstract claims that the proposed loss is the first to extend the ε-insensitive approach to quantile regression 'in true sense', that it improves prediction significantly, and that it restores sparsity.","tokens_in":18967,"tokens_out":6853,"duration_ms":57517,"significance":"If the proposed loss were genuinely new, it would provide a convenient fixed-width asymmetric tube for quantile regression with sparsity and would constitute a meaningful contribution to the SVQR literature. However, as shown in the major comments, the loss is algebraically identical to the Park-Kim loss under a simple rescaling of the tube-width parameter, so the central novelty claim is not correct. The paper does contain a complete QP derivation and a broad experimental study, but the numerical comparisons with Park-Kim are not comparisons of different model families. The practical value of the paper reduces to a reparameterization of an existing loss, which does not support the claimed significance.","major_comments":[{"comment":"The proposed loss in Eq. (5) is algebraically identical to the Park-Kim loss in Eq. (4) under the substitution ε_PK = τ(1−τ)ε. For u > (1−τ)ε, L^ε_τ(u) = τu − τ(1−τ)ε = P_τ(u) − ε_PK; for u < −τε, L^ε_τ(u) = −(1−τ)u − τ(1−τ)ε = P_τ(u) − ε_PK; and the flat interval [−τε, (1−τ)ε] equals the Park-Kim flat interval [−ε_PK/(1−τ), ε_PK/τ]. Since ε_PK is a free user parameter, the two loss families coincide exactly. This contradicts the paper's central claim (Abstract and Section 1) that the proposed loss extends the ε-insensitive idea to quantile regression 'in true sense' and that Park-Kim's loss has a different, τ-dependent width. The difference between the two formulations is purely a reparameterization of the tube width.","section":"Eq. (5) vs. Eq. (4)"},{"comment":"Proposition 1 states that α_i β_i = 0 for all i for ε ≥ 0, but it is false when ε = 0. At ε = 0, a training point whose residual is exactly zero satisfies both KKT conditions (23) and (24) with ξ_i = ξ*_i = 0, so both α_i and β_i can be positive. The proof adds (30) and (31) to obtain ξ*_i + ξ_i = −ε; at ε = 0 this equation is consistent with ξ_i = ξ*_i = 0, so the alleged contradiction disappears. Since the paper explicitly notes that the model reduces to standard SVQR at ε = 0, the sparsity classification into S1, S2, and S3 does not hold in that case, and the proposition's statement for all ε ≥ 0 is incorrect.","section":"Proposition 1 (Section 3)"},{"comment":"The numerical comparisons with Park-Kim SVQR use the same numerical ε grid for both methods but do not account for the reparameterization ε_PK = τ(1−τ)ε. For example, at τ = 0.1 the proposed ε = 1 corresponds to Park-Kim ε_PK = 0.09, which is not in the stated search set {0, 0.1, 0.2, ...}. Thus the reported improvements of ε-SVQR over Park-Kim SVQR are at least partly artifacts of a mismatched ε grid, not evidence that the proposed model family is superior. The experiments in Section 4.2 and on the UCI datasets therefore do not support the abstract's claim of significant improvement over existing SVQR models.","section":"Tables 3, 6, 11, 12 and Section 4.2"}],"minor_comments":[{"comment":"The Introduction states that Section 5 contains numerical results and Section 6 concludes, but the actual sections are numbered 4 (Experimental Results) and 5 (Conclusion).","section":"Section numbering"},{"comment":"The notation in Eq. (16) is confusing because ξ_i and ξ*_i are reused before and after the rescaling ξ_i := ξ_i/τ and ξ*_i := ξ*_i/(1−τ); the derivation would be clearer with different symbols for the intermediate variables.","section":"Eq. (16)"},{"comment":"There are many typographical errors, including 'Preposition' instead of 'Proposition', 'e-insensitive' instead of 'ε-insensitive', 'artifical' instead of 'artificial', and 'Traizines' instead of 'Triazines' in the UCI dataset name.","section":"Throughout"},{"comment":"The statement that Sparse SVQR 'fails to control the effective width' of the ε-insensitive zone for low τ is a consequence of fixing ε numerically rather than a property of the model family; Seok et al.'s loss also has ε as a free tunable parameter.","section":"Section 4.2, discussion of Sparse SVQR"}],"recommendation":"reject","confidential_remarks":"The central contribution is an algebraic identity with the Park-Kim loss, which the authors cite but do not recognize. This is a rejection-level novelty issue. The experimental comparisons with Park-Kim are unfair because the ε grids are not aligned under the reparameterization. The paper may be salvageable only if reframed as a practical reparameterization study, but the current claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central claim doesn't survive contact with the paper's own equations. The loss in Eq (5) is exactly Park and Kim's loss (Eq 4) with epsilon replaced by tau(1-tau)epsilon. Since epsilon is a free parameter, the model family is identical; the only change is how the tube width is parameterized. So the headline novelty—a fixed-width asymmetric epsilon-tube that Park and Kim supposedly lack—is a reparameterization artifact, not a new model.\n\nWhat the paper does well: the QP derivation is mostly clean, the dual is correct modulo a naming typo in Eq (16) where tau and 1-tau get swapped in the constraint constants, and the experiments are extensive. The sparsity analysis is standard but fine. The authors also correctly identify a real limitation of Seok et al.'s loss, namely that its tube width changes with tau.\n\nWhere it falls apart: the equivalence above makes the numerical comparisons unfair. Both methods are tuned over the same numeric epsilon grid. For tau=0.1, proposed epsilon=1 corresponds to Park–Kim epsilon_PK=0.09, which the Park–Kim grid was not allowed to use. So the apparent improvement is mostly a grid mismatch, not a better model. The paper's own Figure 3 and the discussion around it criticize Park–Kim for variable width, but that criticism only holds if you fix epsilon numerically rather than treat it as tunable. Proposition 1 also fails at epsilon=0: both multipliers can be positive when a residual sits exactly on the regression function; the proof's 'contradiction' relies on getting a negative sum, which doesn't occur when epsilon=0. That is a minor edge case, but it shows the proposition isn't stated carefully.\n\nA related point: the paper claims no existing formulation extends epsilon-insensitivity to quantile regression 'in true sense.' That is simply wrong, since Eq (4) already does so up to scaling.\n\nBottom line: this is a reparameterization of Park and Kim, not a new method. The experiments are thorough and the writing is readable, but the core contribution is not novel. I would not bring it to reading group, and I wouldn't spend a referee slot on it. If the authors reframe as a practical note on tuning Park–Kim with a tau-dependent epsilon, there might be a short note in it, but as a standalone contribution it doesn't hold up.","headline":"The proposed loss is exactly Park and Kim's loss under a tau-dependent rescaling of epsilon, so the novelty claim and the reported improvements are artifacts of the parameterization and the epsilon grid.","tokens_in":19624,"tokens_out":2954,"would_cite":false,"duration_ms":29513,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes an asymmetric epsilon-insensitive pinball loss that keeps a fixed-width quantile tube and restores sparsity in support vector quantile regression.","keywords":["quantile regression","pinball loss","epsilon-insensitive loss","support vector regression","support vector quantile regression","sparsity","asymmetric loss"],"falsifier":"For one of the paper's datasets, solve the earlier epsilon-insensitive SVQR with its epsilon parameter multiplied by $\\tau(1-\\tau)$ and compare the resulting regression functions and error tables with the proposed model across the same epsilon grid; identical outputs would show the two formulations coincide.","tokens_in":18362,"feed_emoji":"📊","tokens_out":8487,"duration_ms":81081,"temperature":0.7,"pith_summary":"The paper proposes a new asymmetric epsilon-insensitive pinball loss and builds an epsilon-Support Vector Quantile Regression model on it. The aim is to bring the epsilon-insensitive tube idea from support vector regression into quantile regression in a way that keeps the tube's total width fixed for all quantiles while splitting it asymmetrically according to tau, so points inside the tube are ignored and the solution becomes sparse. The paper argues that earlier epsilon-insensitive pinball losses either use a symmetric tube, a non-convex loss, or a tube whose effective width changes with tau, and it presents experiments on artificial and real-world benchmark datasets claiming improved prediction and sparsity. If correct, this would give practitioners a sparse quantile regression model with a single interpretable epsilon parameter tied to noise variability.","feed_headline":"Fixed-width epsilon-tube loss restores sparsity to quantile regression","feed_subtitle":"The proposed epsilon-SVQR keeps the tube width fixed across quantiles and ignores inside points, improving prediction.","key_machinery":"The central object is the proposed asymmetric epsilon-insensitive pinball loss in Eq. (5): $L^\\epsilon_\\tau(u)=\\max(-(1-\\tau)(u+\\tau\\epsilon),0,\\tau(u-(1-\\tau)\\epsilon))$, where $u$ is the signed residual $y-(w^\\top\\varphi(x)+b)$ and $\\tau\\in(0,1)$ is the target quantile. It combines a one-sided upper arm with slope $\\tau$ that activates above $(1-\\tau)\\epsilon$ and a one-sided lower arm with slope $-(1-\\tau)$ that activates below $-\\tau\\epsilon$, joined by a flat zero-cost tube of total width $\\epsilon$. This machinery produces sparsity: inside the tube the dual variables $\\alpha_i,\\beta_i$ vanish, boundary points determine the bias, and the final quantile regressor is built only from boundary and outside points, with contributions weighted in the ratio $(1-\\tau)$ to $\\tau$.","core_discovery":"The central claim is that the loss $L^\\epsilon_\\tau(u)=\\max(-(1-\\tau)(u+\\tau\\epsilon),0,\\tau(u-(1-\\tau)\\epsilon))$ is the first asymmetric epsilon-insensitive pinball loss that extends the epsilon-insensitive idea faithfully: the insensitive zone has total width $\\epsilon$ regardless of $\\tau$, and the $\\tau$ value only decides how that width is divided between the region above and below the regressor. Training points whose residuals fall inside this asymmetric tube contribute zero loss and drop out of the dual solution, restoring sparsity to support vector quantile regression, which the standard pinball loss lacks. With $\\epsilon=0$ the model reduces to the standard support vector quantile regression model; with nonzero $\\epsilon$ the paper reports lower root-mean-square error and coverage error than the standard, sparse, and earlier epsilon-insensitive models on the datasets tested.","pith_inferences":["Algebraically, the proposed loss is $\\max(0, P_\\tau(u)-\\tau(1-\\tau)\\epsilon)$, which is exactly the loss of the earlier epsilon-insensitive quantile regression model with its own parameter set to $\\tau(1-\\tau)\\epsilon$; the two models therefore form the same one-parameter family, so the claimed new family is a reparameterization.","If that equivalence holds, the reported improvement over the earlier model likely reflects the effective tube widths the search grid happened to select rather than an intrinsic property; a matched-scale comparison would test this directly.","A practical extension the paper leaves open is a data-driven rule for setting $\\epsilon$, since the optimal value in the experiments grows with the noise variance in the response.","The fixed-width property suggests interpreting $\\epsilon$ as a noise-scale parameter shared across quantiles, which could simplify tuning when several quantile curves are estimated jointly."],"forward_implications":["The epsilon-SVQR dual solution has many zero coefficients, so predictions use only boundary and outside points rather than the entire training set.","For a given quantile, the epsilon tube can be widened or narrowed without changing the quantile's asymmetry, and the paper's experiments tie the best epsilon value to the noise variance in the response.","With epsilon set to zero, the model coincides with the standard SVQR, so the new model generalizes rather than replaces the existing formulation.","Coverage error and sparsity both improve as epsilon grows in most of the paper's benchmark experiments, suggesting a practical trade-off between ignoring more points and maintaining quantile accuracy."],"supporting_citations":[{"why":"Introduces pinball loss and the quantile regression problem the paper builds on.","marker":"[1]"},{"why":"Supplies the standard SVQR formulation that the proposed model extends and reduces to when the tube width is zero.","marker":"[6]"},{"why":"Brings the epsilon-insensitive tube idea from support vector regression that the paper adapts to quantiles.","marker":"[7]"},{"why":"Describes the Sparse SVQR model with an asymmetric epsilon-insensitive loss; a primary experimental baseline.","marker":"[13]"},{"why":"Proposes an epsilon-insensitive pinball loss whose effective tube width varies with the quantile; the key comparison baseline.","marker":"[14]"},{"why":"Presents an earlier symmetric epsilon-tube pinball loss that the paper argues is non-convex and is not a true epsilon extension.","marker":"[11]"},{"why":"Provides the exhaustive grid-search tuning procedure used for kernel width, cost, and epsilon.","marker":"[15]"},{"why":"Supplies the evaluation criteria, including RMSE, MAE, Theil U, coverage error, and sparsity, used in the experiments.","marker":"[16]"}],"fun_headline_variants":["Asymmetric epsilon tube makes quantile regression sparse again","New loss brings sparsity back to support vector quantile regression","Epsilon-insensitive pinball loss improves quantile prediction","Fixed-width tube loss revives sparsity in quantile regression","Novel epsilon tube boosts quantile regression accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the proposed loss is genuinely different from a previously proposed epsilon-insensitive quantile regression loss, so that using the same numeric epsilon for both is a fair test; if the two losses are the same family after rescaling epsilon, the reported advantage is an artifact of the parameter grid.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetric epsilon tube makes quantile regression sparse again","New loss brings sparsity back to support vector quantile regression","Epsilon-insensitive pinball loss improves quantile prediction","Fixed-width tube loss revives sparsity in quantile regression","Novel epsilon tube boosts quantile regression accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000523,"raw_usage":{"total_tokens":2511,"prompt_tokens":909,"completion_tokens":1602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1522}},"tokens_in":525,"tokens_out":1602,"duration_ms":12239,"temperature":1.0,"reasoning_tokens":1522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:32:23.446876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For one of the paper's datasets, solve the earlier epsilon-insensitive SVQR with its epsilon parameter multiplied by $\\tau(1-\\tau)$ and compare the resulting regression functions and error tables with the proposed model across the same epsilon grid; identical outputs would show the two formulations coincide.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces pinball loss and the quantile regression problem the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard SVQR formulation that the proposed model extends and reduces to when the tube width is zero."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Brings the epsilon-insensitive tube idea from support vector regression that the paper adapts to quantiles."},{"cited_title":"1 (2010), pp V1-438","cited_arxiv_id":null,"evidence_quote":"Describes the Sparse SVQR model with an asymmetric epsilon-insensitive loss; a primary experimental baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes an epsilon-insensitive pinball loss whose effective tube width varies with the quantile; the key comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents an earlier symmetric epsilon-tube pinball loss that the paper argues is non-convex and is not a true epsilon extension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the exhaustive grid-search tuning procedure used for kernel width, cost, and epsilon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the evaluation criteria, including RMSE, MAE, Theil U, coverage error, and sparsity, used in the experiments."}],"review_version":1}