{"id":"65ac3918-d77e-4f79-bddd-c7cf4d897702","arxiv_id":"2412.07039","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DAVID combines a rebalanced variational autoencoder with a smoothed bootstrap in latent space to generate synthetic rare target values, improving imbalanced regression on tabular data.","lead":"DAVID is a data augmentation method for imbalanced regression. It trains a variational autoencoder (VAE) with a loss that upweights rare target values, then generates new samples by taking a smoothed bootstrap of the latent representations. Tests on simulations and four tabular datasets show lower weighted error than baseline and several prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-data superiority claim rests solely on a custom wMSE that the method is designed to optimize; without standard MSE/MAE and significance tests, it is not established.","rationale":"The central claim is empirical: DAVID outperforms the initial sample and SOTA on real IR datasets. The evidence is Table 2, which uses only the authors' wMSE. This metric is the exact loss-like weighting used inside DAVID (Eqs. 1 and 2), so winning on it is expected. The simulation (Table 1) shows DAVID also improves standard MSE and MAE, which is supportive, but real-data results omit those standard metrics. Without knowing the standard-metric behavior, the headline claim may overstate the method's general value. The absence of significance tests further weakens the evidence; the 10-fold stds in Table 2 do not allow the reader to judge whether differences are reproducible. The reader's proposed conditions (standard metrics, significance tests, sensitivity analysis) target this gap. The latent-space regularity issue the reader identifies as the weakest assumption is indeed a limitation but is explicitly acknowledged by the authors and applies to any VAE-based generator; it is not the decisive evidence gap. Thus the single most load-bearing concern is the evidential basis for the real-data superiority claim, not the internal geometry of the latent space.","tokens_in":12798,"tokens_out":9831,"duration_ms":101026,"concrete_test":"Rerun the Table 2 evaluation and report, for every dataset and method, standard MSE, MAE, and wMSE, along with paired significance tests (Wilcoxon signed-rank or bootstrap over the 10 folds) comparing DAVID against Baseline and the best ILR baseline. If DAVID's standard MSE or MAE is not better than Baseline, or if the wMSE advantage is not significant, the claimed superiority is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims DAVID ('kBV AEw') gives better results than the initial sample and state-of-the-art on real IR datasets (Section 5, Table 2). That table reports only the self-defined wMSE, whose weights ω_i = 1/f_Y(y_i)^α (Eq. 2) are exactly the inverse-density weights used in the VAE loss (Eq. 1) and in seed drawing for the smoothed bootstrap. Thus wMSE is the very objective the method is built to minimize, so an improvement on wMSE is partly tautological. The simulation (Table 1) does include standard MSE/MAE, but these are omitted for the real datasets, making it impossible to tell whether the wMSE gain reflects genuine rare-value improvement or a trade-off that degrades overall accuracy. No significance tests accompany the 10-fold means; e.g., bank8FM drops from 1.87±0.19 to 1.68±0.13, a difference small relative to the spread. The load-bearing assumption is that the wMSE advantage is not an artifact of the metric choice and is statistically robust. The latent-space regularity concern is real but explicitly acknowledged by the authors in the Discussion; the metric/evidence gap is not acknowledged and directly governs the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DAVID, a data augmentation method for imbalanced regression (IR) on tabular data, which combines a β-VAE with a balanced loss and a smoothed bootstrap over the latent means. The balanced loss weights observations by the inverse of the estimated target density, and the smoothed bootstrap generates new latent points from kernel density estimates centered on the latent means, with seeds drawn according to the same inverse-density weights. The method is evaluated on one simulated dataset and four real IR benchmark datasets, comparing against several baselines and state-of-the-art methods. The authors report that DAVID outperforms the initial training sample and the competitors, especially on the real datasets where only the self-defined wMSE metric is reported.","tokens_in":13042,"tokens_out":4415,"duration_ms":41072,"significance":"If the reported gains hold, the method is a useful contribution to the under-studied area of imbalanced regression, offering a principled way to adapt VAEs to IR without discretizing the target variable. The paper's strengths include a clearly described algorithm, a step-by-step ablation on the simulation, the use of multiple learning algorithms via H2O AutoML, and publicly available code on GitHub. The simulation results do include standard MSE/MAE/MAPE metrics, which is commendable. However, the central claim of superiority on real datasets rests on a custom wMSE metric that is aligned with the training objective, and the real-data evaluation lacks standard error metrics and statistical significance tests. The paper also acknowledges a key limitation—the method fails if the VAE is not functioning properly—but does not provide diagnostics to delineate when it works. These gaps currently prevent full verification of the central claim.","major_comments":[{"comment":"The central claim that DAVID 'gives better results than the initial training sample and the state-of-the-art approaches' is supported in Section 5 only by the self-defined wMSE, whose weights ω_i = 1/f_Y(y_i)^α (Eq. 2) coincide with the inverse-density weights used in the training loss (Eq. 1) and in the smoothed-bootstrap seed drawing (Section 3.4). Because the metric and the training objective share the same weighting, the reported wMSE improvement is partly a direct consequence of the method's design; to establish the claim, the authors should report standard MSE/MAE (as done in the simulation, Table 1) on the real datasets, ideally together with a metric that is not aligned with the training loss.","section":"§5, Table 2"},{"comment":"No statistical significance testing is provided for the 10-fold comparisons. For example, bank8FM baseline wMSE is 1.87 ± 0.19 versus DAVID's 1.68 ± 0.13, and abalone 6.84 ± 0.29 versus 5.52 ± 0.22; these differences are small relative to the reported spreads, so it is unclear whether the improvement is systematic. Paired tests (e.g., Wilcoxon signed-rank across folds) or confidence intervals should be added for both tables; otherwise the 'better results' claim is not robust.","section":"§5, Tables 2 and 1"},{"comment":"Two state-of-the-art methods are missing from the comparison: ADASYN was removed after the authors observed poor results and high computation time, and VIR (Wang and Wang, 2024) was not compared because its code is no longer accessible. The VIR exclusion is particularly relevant because VIR is the closest existing VAE-based imbalanced regression method, and the paper's claim of outperforming state-of-the-art VAE approaches cannot be verified without it. At minimum, the authors should discuss the expected comparison and report any partial results; the ADASYN exclusion should be documented with its quantitative results in an appendix rather than reported only qualitatively.","section":"§4.2"},{"comment":"The paper explicitly states 'the method would not work if the VAE is not functioning properly,' yet no diagnostic of VAE quality (e.g., reconstruction error on rare values, or latent-space uniformity checks) is reported for the simulations or real datasets. Since the method's validity depends on the latent space accurately reflecting the data, the empirical section should include such diagnostics or a sensitivity analysis over β_KL and α to demonstrate the regime in which DAVID is effective. This is not a fatal flaw given the acknowledgment, but it is needed to bound the method's scope.","section":"§6, Discussion"}],"minor_comments":[{"comment":"Table 2 contains a row 'kTrain' that is not defined in the protocol list in §4.2; presumably it refers to smoothed bootstrap applied in the original data space (called CSB in Table 1), but the naming should be consistent.","section":"§5, Table 2"},{"comment":"The label 'OVAE' is used inconsistently for what should be '0VAE' (zero-KL VAE); the text and tables should use the same label.","section":"§4.2, Table 1"},{"comment":"In Eq. (1), the weighting term is written as 'bf (Y )α', but the notation is not formally defined; it should be f̂_Y(y)^α to match Eq. (2).","section":"§3.2, Eq. (1)"},{"comment":"The protocol text states that 70% of the data is allocated to the training set, while Appendix A specifies a 60-40% train-test split; this inconsistency should be reconciled.","section":"§4.2 and Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents a sensible method and the code availability is a plus. The main weakness is the evaluative circularity between the custom wMSE metric and the training objective in the real-data experiments, compounded by the lack of standard metrics and significance tests. If the authors add standard MSE/MAE on the real datasets and some form of statistical testing, the central claim could be established. The omission of VIR and ADASYN should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Sam, here's my read on arXiv:2412.07039. The core idea—train a VAE with an inverse-density-weighted reconstruction loss, then generate rare samples via smoothed bootstrap on the latent means—is coherent and, as far as I can tell, genuinely new. The authors do a careful step-by-step ablation on simulation, they ship code, and the simulation results are reported with standard MSE/MAE/MAPE, where DAVID looks best. That part is solid.\n\nThe weak link is the real-data section. Table 2 reports only wMSE, the self-defined metric whose weights are the same inverse-density weights used in the training loss and bootstrap sampling. So the reported gain is partly a consequence of optimizing and measuring with the same objective. The stress-test note is right: without standard MSE/MAE on real data, or paired significance tests, we can't tell whether the wMSE improvement reflects genuine rare-value gains or a trade-off that hurts overall accuracy. The improvements are also small relative to the spread—bank8FM goes from 1.87±0.19 to 1.68±0.13.\n\nADASYN was excluded post hoc for poor performance and high runtime. That's a practical note, not a fatal flaw, but it weakens the state-of-the-art comparison. Hyperparameters like alpha, beta_y, beta_KL, and the noise multiplier are fixed with no sensitivity analysis; alpha=1 is said to converge to a uniform target, but the relation to wMSE is unexplored. The latent-space regularity assumption is real, but the authors explicitly flag it in the Discussion, so I won't hold that against them.\n\nOverall: the method is sensible, the simulation evidence is decent, but the central real-data claim is not yet established. This paper deserves serious peer review—a good referee can push for standard metrics, significance tests, and sensitivity analysis. I'd cite it if I worked on imbalanced regression, mainly for the method description and the simulation protocol.","headline":"DAVID's method is sensible and the simulation is decent, but the real-data claim rests on a metric the method is built to optimize.","tokens_in":13595,"tokens_out":1379,"would_cite":true,"duration_ms":14709,"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 a variational autoencoder trained with an inverse-density weighted loss and generating new points by smoothed bootstrap in its latent space—the DAVID algorithm—outperforms the initial training sample and existing…","keywords":["imbalanced regression","variational autoencoder","smoothed bootstrap","data augmentation","synthetic data generation","tabular data","balanced loss","rare values"],"falsifier":"On an imbalanced-regression benchmark, train DAVID and a control that uses the same balanced-loss VAE but generates from the native per-observation Gaussians $\\mathcal{N}(\\mu_i, \\sigma_i)$ instead of the smoothed bootstrap, with identical seeds and $\\alpha$; if the control matches DAVID's wMSE, the smoothed bootstrap is not the active ingredient. Alternatively, compute the reconstruction error of DAVID's generated rare-value points through the encoder-decoder: if the error is no better than random latent draws, the generator is producing off-manifold samples.","tokens_in":12579,"feed_emoji":"📈","tokens_out":6293,"duration_ms":56338,"temperature":0.7,"pith_summary":"Imbalanced regression, where a continuous target has few rare extreme values, remains under-served compared with classification. The paper proposes DAVID, which trains a variational autoencoder (VAE) with a loss that weights each observation by the inverse of the estimated density of the target, so rare values contribute more to reconstruction. Instead of sampling new latent points from each observation's inferred Gaussian, DAVID draws seeds from the latent means with a kernel-smoothed bootstrap, using each rare point's neighborhood to generate new rare values. The authors report that on simulated data and four benchmark datasets, DAVID gives lower weighted mean squared error than the initial training sample and than state-of-the-art imbalanced-regression methods such as SMOGN, SMOTE for regression, and Gaussian noise, as well as general-purpose generators like TVAE and CTGAN.","feed_headline":"Smoothing a VAE's latent space fixes rare-value regression","feed_subtitle":"Rebalancing by target density and generating from latent neighborhoods beats SMOTE-style and GAN baselines on four datasets.","key_machinery":"The object that carries the argument is the DAVID generator: a $\\beta$-VAE for regression whose target-reconstruction term is weighted by $\\omega_i = 1/\\hat{f}_Y(y_i)^\\alpha$ (inverse empirical density of the target, estimated by kernel density with Silverman or Scott bandwidth), followed by a smoothed bootstrap over the encoder's latent means $\\mu_i$. The smoothed bootstrap draws a seed $\\mu_i$ with probability $\\omega_i$ and adds Gaussian noise $K_{H_n}(z^* - \\mu_i)$ with $H_n = \\eta \\cdot \\mathrm{Var}(\\mu)$, where $\\eta$ follows Silverman's or Scott's rule of thumb. This replaces the VAE's native sampling $z_i = \\mu_i + \\sigma_i \\varepsilon$, which is unreliable for rare values because $\\sigma_i$ is poorly estimated. The smoothed bootstrap uses the neighborhood structure of the latent space, and the VAE's regularity makes that neighborhood meaningful; this is what allows the non-parametric generation of rare target values.","core_discovery":"The central discovery is that the failure of VAEs on imbalanced regression is not primarily a reconstruction problem but a generation problem: the per-observation latent variance $\\sigma_i$ for rare values is poorly estimated and too large, so sampling from $\\mathcal{N}(\\mu_i, \\sigma_i)$ pushes generated points off the rare-value manifold. DAVID replaces this with a joint generator $g_{z^*}(z^*|\\mu) = \\sum_i \\omega_i K_{H_n}(z^* - \\mu_i)$, a Gaussian-mixture smoothed bootstrap over the latent means $\\mu_i$ with bandwidth $H_n$ from Silverman's or Scott's rule, and seed weights $\\omega_i = 1/\\hat{f}_Y(y_i)^\\alpha$. Because the VAE's latent space is regular (continuous and complete), kernel density estimation is valid there—unlike in the original data space or in a plain autoencoder's latent space. The authors show each component matters: the balanced loss improves over the vanilla $\\beta$-VAE, the smoothed bootstrap improves over natural VAE generation, and combined (kBVAEw, which is DAVID) they beat the baseline and competitors on weighted MSE.","pith_inferences":["A testable extension is to adapt DAVID's weighting scheme to classification by replacing the continuous target density with class-frequency weights, then applying the smoothed bootstrap to generate minority-class latent samples without binning the target.","The success of the smoothed bootstrap suggests that VAE variance estimates are the bottleneck for rare-value generation; one could directly regularize $\\sigma_i$ (e.g., shrink large variances) and test whether the gap between natural VAE generation and smoothed bootstrap closes.","The method's dependence on the VAE's latent regularity implies that the generator's performance can serve as a diagnostic for latent-space quality: if wMSE degrades, the encoder has likely not learned a continuous manifold.","Since the paper averages results over ten models via H2O AutoML, the gains are already model-averaged; a further test would isolate which base learners benefit most from DAVID augmentation."],"forward_implications":["DAVID lowers weighted MSE on the four real benchmark datasets (bank8fm, abalone, boston, NO2), achieving the best mean wMSE among all compared methods.","Standard synthetic generators not designed for imbalanced regression (TVAE, CTGAN, CopulaGAN) degrade performance relative to the initial sample, so they should not be used for this task without adaptation.","Generating in the latent space of a regular VAE is more effective than generating in the original data space (kTrain) or in PCA or kernel-PCA latent spaces.","The balanced loss alone (BVAEw) improves over the vanilla VAE, and the smoothed bootstrap alone (kBVAE) is comparable to natural VAE generation, so the two modifications are complementary.","Because the method is a preprocessing step, it works with any downstream regressor, including the white-box models preferred in insurance and finance."],"supporting_citations":[{"why":"Establishes that standard MSE is inefficient for imbalanced regression, motivating the balanced loss used in DAVID.","marker":"(Ren et al., 2022)"},{"why":"Provides SMOTE for regression, a key state-of-the-art baseline that DAVID must beat, and the reference for UBL-based rebalancing.","marker":"(Torgo et al., 2013)"},{"why":"Supplies the UBL package and weighting framework used to build the oversampling and Gaussian-noise baselines.","marker":"(Branco et al., 2016a)"},{"why":"Introduces SMOGN and Gaussian-noise preprocessing for imbalanced regression, the closest preprocessing competitors and source of benchmark datasets.","marker":"(Branco et al., 2017)"},{"why":"Introduces the smoothed bootstrap, the non-parametric generation mechanism at the heart of DAVID.","marker":"(Silverman and Young, 1987)"},{"why":"Proves consistency of the multivariate smoothed bootstrap, providing theoretical grounding for using it in the latent space.","marker":"(De Martini and Rapallo, 2008)"},{"why":"Supplies the bandwidth selection rule (Silverman's rule of thumb) used for the kernel generator $H_n$.","marker":"(Silverman, 1986)"},{"why":"Documents that neural networks, including VAEs, are inefficient on tabular data, framing the paper's choice of a preprocessing approach.","marker":"(Shwartz-Ziv and Armon, 2022)"},{"why":"Provides the TVAE and CTGAN general-purpose synthetic data generators used as non-IR baselines.","marker":"(Xu et al., 2019)"},{"why":"Presents the VAE-based deep imbalanced regression approach that uses target discretization, which DAVID explicitly avoids to prevent information loss.","marker":"(Wang and Wang, 2024)"}],"fun_headline_variants":["Smoothed bootstrap in VAE latent space fixes rare-value regression","Rare-value regression: sample from smoothed latent means, not VAE noise","VAE's latent-space smoothing beats GANs for imbalanced regression","Fix imbalanced regression by bootstrapping VAE latents, not sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gains depend on the VAE learning a latent space that is regular and faithful enough that kernel-smoothed neighborhoods around the latent means $\\mu_i$ correspond to valid, on-manifold rare-value samples; if the VAE is poorly trained, the generated points are off-manifold and augmentation can worsen regression, as the paper itself notes.","fun_headline_variants_meta":{"raw":{"variants":["Smoothed bootstrap in VAE latent space fixes rare-value regression","Rare-value regression: sample from smoothed latent means, not VAE noise","VAE's latent-space smoothing beats GANs for imbalanced regression","Fix imbalanced regression by bootstrapping VAE latents, not sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1431,"prompt_tokens":943,"completion_tokens":488,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":559,"tokens_out":488,"duration_ms":5300,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:11:44.043080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On an imbalanced-regression benchmark, train DAVID and a control that uses the same balanced-loss VAE but generates from the native per-observation Gaussians $\\mathcal{N}(\\mu_i, \\sigma_i)$ instead of the smoothed bootstrap, with identical seeds and $\\alpha$; if the control matches DAVID's wMSE, the smoothed bootstrap is not the active ingredient. Alternatively, compute the reconstruction error of DAVID's generated rare-value points through the encoder-decoder: if the error is no better than random latent draws, the generator is producing off-manifold samples.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that standard MSE is inefficient for imbalanced regression, motivating the balanced loss used in DAVID."},{"cited_title":"P., Pfahringer, B., and Branco, P","cited_arxiv_id":null,"evidence_quote":"Provides SMOTE for regression, a key state-of-the-art baseline that DAVID must beat, and the reference for UBL-based rebalancing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces SMOGN and Gaussian-noise preprocessing for imbalanced regression, the closest preprocessing competitors and source of benchmark datasets."},{"cited_title":"and Young, G","cited_arxiv_id":null,"evidence_quote":"Introduces the smoothed bootstrap, the non-parametric generation mechanism at the heart of DAVID."},{"cited_title":"and Rapallo, F","cited_arxiv_id":null,"evidence_quote":"Proves consistency of the multivariate smoothed bootstrap, providing theoretical grounding for using it in the latent space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the bandwidth selection rule (Silverman's rule of thumb) used for the kernel generator $H_n$."},{"cited_title":"and Armon, A","cited_arxiv_id":null,"evidence_quote":"Documents that neural networks, including VAEs, are inefficient on tabular data, framing the paper's choice of a preprocessing approach."},{"cited_title":"and Wang, H","cited_arxiv_id":null,"evidence_quote":"Presents the VAE-based deep imbalanced regression approach that uses target discretization, which DAVID explicitly avoids to prevent information loss."}],"review_version":1}