{"id":"583505ea-c463-4a4e-895f-4b3ede559419","arxiv_id":"2412.15657","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding an 'overlap' class label for boundary majority points during generative model training improves synthetic minority data quality and downstream classifier accuracy on imbalanced tabular data.","lead":"The paper proposes a preprocessing trick for imbalanced tabular classification: relabel majority-class points that lie on the class boundary as a third 'overlap' class before training a generative model. The trick improves the quality of synthetic minority data and raises classifier accuracy across several deep generative models and real-world datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for an arbitrary third class: ORD's gains may stem from any majority split rather than the true overlap.","rationale":"The reader's weakest assumption—that RF confidence identifies the true overlap—is real, but I think the more fundamental gap is the absence of a control for an arbitrary third class. The paper's positive results could arise simply because a conditional generator benefits from an extra mode for part of the majority, or because the classifier sees a cleaner majority distribution; neither requires the labels to correspond to the Bayes overlap. Table 6 attempts to address this but is methodologically flawed: the 'after synthesis' filter removes an uncontrolled number of points, and its result is far worse than the baseline, so it does not isolate the generator benefit. A random-split/under-sampling control would directly settle whether the overlap concept—the paper's main novelty—is essential. Given the broad empirical support across generators and datasets and the plausible mechanism, the condition is addressable, so a CONDITIONAL verdict remains appropriate.","tokens_in":21273,"tokens_out":12919,"duration_ms":108065,"concrete_test":"On the four main real datasets, run the full ORD pipeline under three alternative majority labelings of the same size as D01: (a) a random subset of majority labeled as class 2, (b) a k-means split of the majority into two clusters labeled 0 and 2, and (c) the RF-based overlap. For each, train the same conditional generator, sample equal class 0 and class 1 instances, and train the same classifiers. Additionally, include a random-majority-under-sampling baseline: train the classifier on the binary generator's synthetic majority after randomly discarding the same fraction of majority points, with no ternary labels. If (a), (b), or the random-under-sampling baseline matches ORD's accuracy, the overlap identification is not load-bearing; if they are significantly worse, the overlap class is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ORD's central claim is that labeling the true overlap region as a third class improves generator quality. The paper never tests whether any three-class split of the majority yields the same improvement. On real data, the overlap is found via random forest confidence (Section 4.1), with no ground-truth validation; on toy data, the overlap is detected with a near-perfect Bayes classifier (Appendix E), not the RF protocol. The only ablation separating generator-side from classifier-side effects (Table 6, 'ORD before vs after synthesis') is confounded: applying the real-data RF to binary-generated data removes an uncontrolled number of majority points, and the 'after' accuracy for Adult (65.32) is far below the binary baseline (78.28), suggesting the filter itself causes the drop. Thus the paper has not shown that the specific overlap labels are the active ingredient; a random split of the majority or simple random under-sampling might match ORD.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ORD, a preprocessing step for imbalanced tabular classification with generative data augmentation. ORD detects a subset of majority points that overlap the minority class using random-forest disagreement in k-fold cross-validation, relabels the training set as ternary (minority, overlap majority, clear majority), trains a class-conditional generator on those labels, and then trains the downstream classifier on balanced synthetic samples from minority and clear majority plus real minority points. The authors report improved classifier accuracy on four datasets across five generators, improved oracle-labeled quality of generated minority points on toy data, and ablations that attribute the gains to both improved generator quality and classifier-side downsampling of overlap points.","tokens_in":21403,"tokens_out":7837,"duration_ms":69896,"significance":"If the central claim is upheld, ORD is a simple, generator-agnostic intervention with practical value for imbalanced tabular data, and the observation that deep generative models degrade on minority classes is useful. Strengths include public code, multiple generators and classifiers, a paired statistical test, and oracle-based quality evaluation on toy data where the true Bayes label is known. However, because the active-ingredient claim is not tested against a placebo split of the majority class, the proposed explanation for the gains is currently unestablished, and a key ablation is confounded.","major_comments":[{"comment":"The central claim is that labeling the true overlap region as a third class improves generator quality, but the paper never compares ORD against a control in which the majority class is split at random (or by a non-overlap rule) into two labels with the same number of points. Without this placebo control, the gains in Table 2 cannot be attributed to the overlap region specifically; they could come from any fine-grained split of the majority class making the conditional generation task easier, or from the classifier-side removal of a subset of majority points. Please add a randomized-split control that matches the size of D01 for each dataset and generator, and report the paired comparison against ORD.","section":"Section 4.1 and Table 2"},{"comment":"The 'ORD before vs after synthesis' ablation is confounded. In the 'after' condition, applying the real-data overlap detector to binary-generated data removes an uncontrolled number of generated majority points; for Adult the resulting XGBoost accuracy (65.32) is far below the binary baseline (78.28 in Table 2), so the drop cannot be attributed solely to the absence of ternary training labels. The authors should control for the number of removed points, for example by also removing the same number of majority points at random from binary-generated data, and should report how many points were filtered in each condition.","section":"Table 6"},{"comment":"The parameterization of the overlap-set size is inconsistent as written. Section 4.1 states that 'the number of overlapped majority points = min(number of minority points, r% of majority points)', with r values 3, 5, 7, 9, but Appendix Table 13 reports overlap counts for Adult and Cardio at thresholds 0.20-0.45 that do not satisfy this formula. For Adult, 3% of the majority count (34,514) is 1,035, yet the largest count in Table 13 is 915 at threshold 0.20; for Cardio, threshold 0.30 yields 700, which equals the minority count if that count is 700, but then the min formula would cap at that value. The relationship between tau and r is therefore not reproducible as written. Please define the exact procedure used to set tau/r and report the resulting overlap counts for the main experiments.","section":"Section 4.1 and Appendix Table 13"},{"comment":"The selection of the r parameter is not fully specified. The text says r is chosen by 'checking the performance on the validation set of a dataset' and then 'use the best r value with all datasets for a given synthesizer,' but it does not state which dataset's validation set is used, whether the same validation split is also used to tune baselines, or whether the reported test results are on data never used for r selection. If r was tuned on the same data used to report the main accuracy numbers, the paired t-test is not a fully held-out comparison. Please specify the exact protocol and, ideally, report the main results for each of the four candidate r values across all datasets and generators, since Appendix Tables 14-15 cover only Adult and Cardio with two synthesizers.","section":"Section 4.1 and Experimental Setup"}],"minor_comments":[{"comment":"The column semantics in Table 1 are unclear: 'Min/Maj' appears to be the ratio of the undersampled minority to the majority, but the Cardio row gives an 'Unb Min' value of 100 while the text and Appendix Table 13 suggest a minority count around 700; please clarify and correct the table.","section":"Table 1"},{"comment":"Figure 1 says 'use k-fold training to identify the overlap in the validation set,' but Algorithm 1 splits only the majority instances D0 into folds within the training set; the terminology should be made consistent.","section":"Figure 1 and Algorithm 1"},{"comment":"The main accuracy tables do not show error bars, and the discussion of gains and losses does not reference the standard deviations in Appendix Table 11. For example, Table 2 shows CTGAN ORD on Heloc dropping from 62.93 to 61.47, and CTAB-GAN+ ORD on Fintech dropping from 57.45 to 54.83; these decreases should be acknowledged when claiming a gain in accuracy.","section":"Main tables"},{"comment":"The abstract and Section 5 refer to an 'Oracle Bayesian model,' but Table 4 uses an XGBoost classifier trained on real balanced data, not a Bayes classifier; the terminology should be adjusted to avoid overclaiming.","section":"Abstract and Section 5"},{"comment":"The paper lists CTabSyn as 'a new SOTA model,' but Table 2 shows that on Fintech the average-of-4 accuracy of CTabSyn without ORD (65.37) is higher than that of CTabSyn ORD (63.58); the SOTA claim should be qualified to the settings where it is actually supported.","section":"Contribution list"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the venue and the empirical method is potentially useful, but the missing placebo control and the confounded ablation in Table 6 are load-bearing for the stated claim. I believe the authors can address these with additional experiments and revised reporting, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful empirical paper with a simple idea—relabel boundary majority examples as a third \"overlap\" class before training deep tabular generators—and the broad evaluation largely backs it up. The missing control is a random split of the majority into two classes; without it the paper cannot support the claim that overlap semantics, rather than any finer-grained majority split, drive the gains. That is the main thing to ask for.\n\nWhat is actually new: to my knowledge nobody has applied this ternary relabeling to deep tabular generators for imbalance. The paper shows consistent gains across CTGAN, CTAB-GAN+, TabDDPM, TabSyn, and ForestFlow on four constructed-imbalance datasets, with a paired t-test p=6.8e-5. The toy experiments with a known Bayes oracle are the strongest evidence: ORD improves minority generation accuracy for every generator in Table 3. The ablations—removing overlap on real data improves the classifier, threshold sensitivity in the appendix—are reasonable. Code is released, and the appendix includes standard deviations over runs.\n\nSoft spots, in order:\n1. No control for an arbitrary third class. If you split the majority randomly into two pseudo-classes during generator training, do you get the same gain? The paper never tests this. Appendix Table 12 addresses classifier boundary shifting, not generator conditioning. This is the key missing experiment for the paper's explanatory claim.\n2. The \"ORD before vs after synthesis\" ablation (Table 6) is confounded, as the stress-test note says: applying the real-data overlap filter to binary-generated data removes an uncontrolled number of points, and Adult drops to 65.32 vs the 78.28 baseline. That table should be redone with a matched number of removed points or a random filter.\n3. The r parameter is tuned per generator on validation; acknowledged, but it means the headline comparison is not fully parameter-free. Minor.\n4. The novelty overclaim—\"first work to harness deep generators for imbalanced data\"—ignores the cited CTGAN-based imbalanced papers. Minor, but easy to fix.\n5. Main tables lack error bars; the appendix std devs are there, but readers should not have to dig.\n\nThe central empirical finding—that adding an overlap class helps generator quality and classifier accuracy—holds up reasonably well. The mechanism is underdetermined, not wrong. This paper is for anyone working on tabular synthetic data or imbalanced classification; it is a practical preprocessing trick that is easy to adopt.\n\nRecommendation: send to peer review. It is empirically rich, reproducible, and the main objection is addressable with one or two control experiments. A referee should push for the random-split control but not desk-reject this.","headline":"A genuinely simple and mostly convincing fix for imbalanced tabular generation, but the missing random-third-class control leaves the mechanism underdetermined.","tokens_in":21923,"tokens_out":2824,"would_cite":true,"duration_ms":25569,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Splitting majority into clear and overlap classes improves synthetic minority data.","keywords":["class imbalance","synthetic tabular data","overlap region","diffusion models","GANs","class-conditional generation","random forest disagreement","machine learning efficacy"],"falsifier":"Compute the true overlap region on a synthetic Gaussian-blob dataset where the Bayes boundary is known, then compare the random forest's overlap set with the truth. If ORD's gains disappear when the true overlap set is used instead of the estimated one—or if the estimated set is far from the truth yet gains remain—then random-forest disagreement is not the mechanism the paper claims.","tokens_in":21067,"feed_emoji":"📊","tokens_out":5610,"duration_ms":45043,"temperature":0.7,"pith_summary":"The paper claims that one preprocessing step—splitting the majority class into a clear part and an overlap part that sits on the boundary with the minority class—markedly improves deep generative models for imbalanced tabular data. It converts binary labels to ternary labels before training the generator, then drops the overlap class when training the downstream classifier. On four real datasets with extreme imbalance (minority around 1.5–2%) and four inherently imbalanced UCI datasets, the method raises classification accuracy for five different classifiers across five different generators, with the largest gains for diffusion models. The paper's own ablations show the gains come from two separate effects: the ternary labels make the generator produce higher-quality minority examples, and excluding overlap majority points from classifier training declutters the decision boundary. A careful reader would care because the method is a cheap, generator-agnostic add-on rather than a new architecture.","feed_headline":"Overlap class boosts synthetic minority data","feed_subtitle":"Splitting majority into clear and boundary labels improves five generators on eight imbalanced datasets.","key_machinery":"The load-bearing mechanism is the overlap class $D_{01}$, defined as the subset of majority instances whose label is uncertain under a k-fold cross-validated random forest. For each fold, a random forest is trained on the other majority folds plus all minority instances; a majority point whose predicted majority probability falls below $1-\\tau$ is declared overlapping and relabeled as class 2. The size of $D_{01}$ is capped by $\\min(\\text{minority count}, r\\% \\text{ of majority count})$, with $r$ tuned on a validation set per generator. This ternary labeling is then fed to any class-conditional generator, and the paper's CTabSyn variant adds the target-class embedding to the TabSyn denoiser input. The overlap class does double work: it teaches the generator where the minority boundary is, and its removal from classifier training acts as a principled under-sampling of boundary points.","core_discovery":"The central discovery is that deep generative models trained on imbalanced tabular data generate minority instances of far lower quality than majority instances, and that this failure can be substantially corrected without changing the generator's architecture. The correction is to label a subset of majority points that lie in the overlap region between the two classes as a third class, train the generator conditionally on these three classes, and then train the final classifier only on generated minority and clear-majority points plus real minority points. The paper demonstrates this through a paired comparison across CTGAN, CTABGAN+, TabDDPM, TabSyn, and ForestFlow, reporting accuracy gains in 26 of 32 (dataset, generator) combinations and a paired t-test p-value of 6.84e-05. It also provides visual and quantitative evidence on toy Gaussian-blob datasets where the true Bayes labels are known, showing that ORD reduces the fraction of wrongly generated minority points.","pith_inferences":["Editorial inference: the overlap-detection idea could extend to multiclass settings by pairwise or one-vs-rest overlap labeling; a natural test is whether per-class ternary splits preserve the same gains for non-binary targets.","Editorial inference: because the method works by making the generator's conditioning signal finer, other forms of auxiliary boundary information—such as density-ratio estimates or calibrated probabilities—might substitute for random forest disagreement and give larger gains.","Editorial inference: a strong testable extension is to replace the random forest with an explicitly calibrated classifier; if gains persist with a calibrated overlap detector, then miscalibration of the random forest is not the limiting factor.","Editorial inference: the method's reliance on validation-tuned $r$ per generator implies a small search cost; a practical extension would be a cheap heuristic for $r$ based on class-overlap statistics rather than a grid search."],"forward_implications":["ORD is a drop-in preprocessing layer: any existing tabular synthesizer gains accuracy when trained on the ternary-labeled data, so future generators can inherit the benefit without architectural change.","Classifiers trained on ORD-augmented data improve in macro-averaged accuracy, with the largest minority-accuracy gains on diffusion models such as TabSyn and ForestFlow.","Augmenting synthetic data with real minority points $D_1$ further improves the average of four classifiers, suggesting that real minority examples carry information the generator still misses.","The overlap region is dataset-specific: the optimal threshold $\\tau$ varies (0.3 for Adult, lower for Cardio), so validation-based selection of $r$ is needed in practice.","ORD is not a simple boundary shift: tuning the decision threshold of a classifier trained without ORD does not match ORD's accuracy in the paper's comparison."],"supporting_citations":[{"why":"Provides the random forest whose cross-validated disagreement identifies the overlap majority set.","marker":"Breiman 2001"},{"why":"SMOTE is the classical augmentation baseline ORD must beat, and its interpolation idea motivates the synthetic-data approach.","marker":"Chawla et al. 2002"},{"why":"CTGAN is one of the five GAN-based generators evaluated with and without ORD.","marker":"Xu et al. 2019"},{"why":"CTABGAN is a GAN-based baseline that ORD is applied to and compared against.","marker":"Zhao et al. 2021"},{"why":"CTABGAN+ is a state-of-the-art GAN baseline in the comparison.","marker":"Zhao et al. 2022"},{"why":"TabDDPM is a diffusion baseline used to show ORD benefits both diffusion and GAN generators.","marker":"Kotelnikov et al. 2022"},{"why":"TabSyn is the latent diffusion baseline that ORD modifies into the conditional CTabSyn variant.","marker":"Zhang et al. 2024"},{"why":"ForestFlow is the gradient-boosted diffusion baseline in the experiments.","marker":"Jolicoeur-Martineau, Fatras, and Kachman 2024"},{"why":"XGBoost is the primary classifier whose accuracy is reported and also serves as an oracle labeler in the synthetic-quality experiment.","marker":"Chen and Guestrin 2016"}],"fun_headline_variants":["Overlap class label boosts minority synthesis in GANs and diffusion","Adding boundary class improves imbalanced synthetic tabular data","Third label for overlap region aids deep generative models","Split majority into overlap class to fix minority generation","Overlap labeling corrects minority quality in synthetic data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"A cross-validated random forest's disagreement score reliably marks where the two classes genuinely overlap, and the overlap size chosen on a validation set carries over to new data.","fun_headline_variants_meta":{"raw":{"variants":["Overlap class label boosts minority synthesis in GANs and diffusion","Adding boundary class improves imbalanced synthetic tabular data","Third label for overlap region aids deep generative models","Split majority into overlap class to fix minority generation","Overlap labeling corrects minority quality in synthetic data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1602,"prompt_tokens":974,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":551}},"tokens_in":590,"tokens_out":628,"duration_ms":6129,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:12:33.739320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the true overlap region on a synthetic Gaussian-blob dataset where the Bayes boundary is known, then compare the random forest's overlap set with the truth. If ORD's gains disappear when the true overlap set is used instead of the estimated one—or if the estimated set is far from the truth yet gains remain—then random-forest disagreement is not the mechanism the paper claims.","supporting_citations":[{"cited_title":"V.; Bowyer, K","cited_arxiv_id":null,"evidence_quote":"SMOTE is the classical augmentation baseline ORD must beat, and its interpolation idea motivates the synthetic-data approach."},{"cited_title":"CTAB-GAN: Effective Table Data Synthesizing","cited_arxiv_id":"2102.08369","evidence_quote":"CTABGAN is a GAN-based baseline that ORD is applied to and compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ForestFlow is the gradient-boosted diffusion baseline in the experiments."}],"review_version":1}