acceptodds
Withdrawn from ICLR 2027

TabMaDo: Dual-Guided Tabulamba Diffusion Oversampling for Imbalanced Tabular Data Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tabular data, as a prevalent form of structured data, often suffers from class imbalance that limits classification performance. To address this issue, diffusion models have been employed for oversampling due to their ability to capture complex feature dependencies. However, when minority class samples are extremely scarce, they struggle to learn the genuine distribution of the minority class, consequently yielding generated samples of poor quality and diversity. To mitigate this challenge, this study proposes TabMaDo, a dual-guided diffusion framework tailored for tabular data. The framework consists of two components, a Mamba-based noise predictor named Tabulamba incorporates Class-Aware Mamba Blocks that dynamically highlights minority class information, preventing the generation process from being overwhelmed by the majority class. This enables the diffusion model to learn the overall data distribution, moving beyond the constraints imposed by scarce minority class samples. Complementing this, a lightweight Guider is introduced to provide dual guidance on both class label and distribution. This ensures the reverse diffusion process is steered toward minority class characteristics while preserving sample diversity, effectively overcoming the limitation of minority class sample scarcity. Together, they form a cohesive framework that enables high-quality sample generation under extreme class imbalance. Extensive experiments on 10 binary classification datasets demonstrate that TabMaDo outperforms six SOTA baseline methods across F1-score, G-mean and AUC metrics, achieving maximum improvements of 12% in F1-score, 16.5% in G-mean and 0.45% in AUC. These results confirm that TabMaDo provides an effective solution for imbalanced tabular data generation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
tabular data, class imbalance, oversampling, diffusion models, Mamba

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_egVCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TabMaDo, a diffusion-based oversampling method for imbalanced binary tabular classification. Motivated by the observation that the extreme scarcity of minority samples can prevent a diffusion model from learning their distribution directly, TabMaDo learns the overall data distribution from all samples and steers the reverse process toward the minority class. It has two components: Tabulamba, a Mamba-based noise predictor, and a light Guider that combines a minority-class classifier.

Critical strengths

Generative oversampling under extreme imbalance is a practical and valuable problem.

Critical weaknesses
  • Both components of TabMaDo have close counterparts in prior work. The idea of learning the distribution from all data and then steering toward the minority class is also what class-conditional diffusion models such as TabDDPM already do. The paper does not clearly explain how TabMaDo fundamentally differs from these methods, nor does it compare against the most direct alternatives.

  • Many citations are misattributed: the references given for TabDiT, ACWGAN, random oversampling, and GBDT point to unrelated papers, while some important foundational works such as DDPM, Mamba, and SMOTE are not cited. The related work section also lacks a substantive comparison between existing tabular oversampling and generation methods and the proposed method.

  • The most critical issue is that in Table 4, the rows for wilt and vehicle0 are identical for all methods and all metrics. Given how different these two datasets are, this is not plausible.

Questions for authors
  • Could you please provide corrected and real Table 4 rows for wilt and vehicle0 results and confirm whether other tables are also affected?

  • What is the fundamental difference between the Guider and RelDDPM’s controller, or standard classifier guidance?

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

This paper proposes TabMaDo, a diffusion based oversampling framework for imbalanced tabular classification, targeting the difficulty of generating minority-class samples when training examples are scarce. Its noise predictor, Tabulamba, learns the overall data distribution using bidirectional Mamba blocks and class-aware state-update gains that emphasize minority-class information. A separate Guider is trained with classification loss and prototype-based contrastive losse, using the empirical mean of minority class samples as the prototype. During sampling, the Guider provides the class label and prototype guidance to steer the reverse diffusion process toward the minority class. The method is evaluated against six oversampling baselines on 10 binary classification datasets using Random Forest, GBDT, and TabPFN classifiers. The manuscript reports experimental improvements of TabMaDo over baselines in metrics of the downstream classifiers(F1, G-mean, AUC).

Critical strengths
  1. Extreme class imbalance in the tabular domain is an actually significant problem to tackle, especially in actual industry situations.
  2. The reported experimental gains are promising - the proposed method consistently outperforms baseline for the three metrics for most datasets.
Critical weaknesses
  1. No experiments that control the imbalance ratio within the same dataset, so the comparisons across datasets do not isolate robustness to imbalance severity.
  2. The evaluation is limited to binary classification, with no multiclass experiments. It remains unclear whether the proposed method remain effective when multiple minority classes have different frequencies and overlapping distributions.
  3. Steering generation toward a single minority prototype may induce mode collapse or degenerate samples, particularly when the minority distribution is multimodal.
Questions for authors
  1. (minor) Is it a broken citation in L150?
Reviewer_vqNdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper deals with the problem of oversampling for imbalanced binary tabular data with a diffusion model, and the pitch is that when you only have a handful of minority rows you should not try to fit a generator to them directly. Instead, learn the whole data distribution and then steer sampling toward the minority region at generation time.

Two pieces make up the method. The first is Tabulamba, a Mamba-based noise predictor that treats each row as a sequence of feature tokens. Each block has a learnable feature-interaction matrix, forward and backward SSM branches, and a "class-aware gain" that multiplies the input term by 1 + alpha for minority rows and 1 for majority rows (Eq. 6), followed by a gated residual. The second is the Guider, a small two-encoder network trained with a BCE minority/majority loss plus a prototype loss that pulls noisy minority embeddings toward the minority mean embedding and pushes majority embeddings away. At sampling time, the Guider's input gradient is added to the reverse-process mean, mu* = mu_t + s g_t, which is in the form of classifier guidance. Section 3.6 adds three monotonicity lemmas and a convergence theorem for the Guider's loss.

The validation is a fairly standard grid for this subarea: ten KEEL/UCI-style datasets with imbalance ratios from 1.84 to 41.4, six oversampling baselines (stable SMOTE, TabDDPM, CoDi, TabDiT, RelDDPM, GOIO), three downstream classifiers (RF, GBDT, TabPFN), stratified 10-fold CV, and F1 / G-mean / AUC. On top of that, there is a Friedman plus Nemenyi analysis, a four-variant ablation (MLP backbone, label-only Guider, no Guider, MLP and no Guider), eight distribution-similarity metrics for the generated minority samples, and a small feature-order sensitivity study on yeast6. The headline claims are that TabMaDo has the best average performance on all three metrics, is statistically significantly better than most baselines, and that both Tabulamba and the Guider are necessary.

Critical strengths
  1. The problem decomposition is right. Sec. 3.1 argues that under extreme scarcity one should not fit a generator to 17 or 20 minority rows but instead learn the joint distribution on all the data and steer sampling toward the minority region with a lightweight guide. That is the correct way to think about this regime, it is stated clearly, and it is the reason I think there is a publishable paper here once the empirical section is repaired. The ablation, read correctly, actually supports this part of the design: removing the Guider (V3) or reducing it to labels only (V2) costs a great deal on every classifier and metric, so the guidance mechanism is doing real work.

  2. The experimental grid is the right shape, and the underlying runs appear to be traceable. Three downstream classifiers, three metrics, stratified 10-fold CV with standard deviations for every cell, a rank-based omnibus test, a four-arm ablation, intrinsic distribution-similarity metrics, and a feature-order check is more than most papers in this subarea provide. Full pseudocode is given for all four procedures. Table 2 reconciles exactly, and most of Table 6 recomputes exactly from Tables 1/3/4 once the CoDi/TabDiT swap is accounted for, which tells me the per-dataset numbers are real and the problems are in the summary layer (statistics, captions, one duplicated row) rather than in the experiments themselves. That is consequential because it makes the required revision a matter of rerunning analyses and a few controls rather than redoing the study, and it is why my recommendation is a reject that could become an accept at a later venue rather than a reject on the merits of the idea.

Critical weaknesses
  1. The statistical analysis and the summary tables do not describe the experiment in the paper. Appendix A.4.3 reports a Nemenyi CD of 2.4078 and a Friedman critical value of 2.2172; with k = 7 methods both constants invert uniquely to N = 14 datasets, whereas Sec. 4.1, the abstract and Table 2 say ten (for N = 10 the CD is 2.8483, 18% wider). The captions of Tables 1, 9 and 10 say fourteen datasets and Sec. A.4.4 says twelve. On top of that, Table 6's GBDT F1 row sums to 28.071 (mean ranks over seven methods must sum to 28), its CoDi and TabDiT columns are transposed relative to Tables 1/3/4 in eight of nine rows, Table 7 is inconsistent with Table 6 and the stated CD in four of nine rows, Table 5's Coverage column sums to 18.10 instead of 21, and Table 4 has two structurally different datasets (wilt, N = 4839, F = 5; vehicle0, N = 846, F = 17) with identical values in all 21 cells. The consequence is that every significance statement in Sec. 4.3 and Table 7, and a third of the rank analysis (the TabPFN block), cannot be trusted as printed. I read this as carry-over from an earlier 14-dataset draft plus a copy-paste error rather than intent, but it means the empirical section cannot currently be checked against itself, and repairing it requires rerunning the statistics and part of the TabPFN grid. That is beyond what a rebuttal can deliver and is the main reason for my recommendation.

  2. The paper's central claims are not supported by its own tables, and the ablation points away from the titular contribution. "TabMaDo achieves the highest average performance on all three metrics" is false for AUC under RF (RelDDPM 0.8810 vs 0.8724) and GBDT (0.8933 vs 0.8908). "On datasets with high IR ... other methods degrade" is contradicted on abalone9-18, where TabMaDo's RF AUC ranks 4th of 7. The three headline improvements in the abstract are single cells on glass-0-1-4-6-vs-2 (17 minority rows, 1.7 per test fold) and yeast6 with standard deviations of 0.33, 0.44 and 0.11; a Welch test on the paper's own means and sds gives p = 0.39, 0.38 and 0.92. The honest effect size against the strongest baseline is a consistent but sub-one-point median gain (TabMaDo beats RelDDPM in 75/90 cells, median delta +0.0045, |delta| < 0.01 in 55/90). Meanwhile the clean backbone contrast in the ablation, V3 (Tabulamba, no Guider) vs V4 (MLP, no Guider), goes 39-48 against Tabulamba with sign-test p = 0.39; Sec. 4.4 supports RQ-1 by quoting one cell of that 90-cell comparison. Removing the Guider costs far more than swapping the backbone. So the method that actually works is the Guider, the Mamba backbone is close to neutral once the Guider is present, and the paper's framing, title and Sec. 5 conclusion ("applying Mamba to tabular data is necessary") are not what the evidence shows. This matters because even after the arithmetic is fixed, the contribution as claimed would still not be established without new controls (alpha = 0, matched-capacity backbones, paired per-fold tests), which makes this a new experimental section rather than a revision.

Questions for authors
  1. How many datasets were actually used for the statistical analysis? The reported CD = 2.4078 and F_alpha = 2.2172 both correspond to N = 14 with k = 7, not to the ten datasets in Table 2. If a 14-dataset experiment exists, please report it in full; if not, please recompute both statistics for N = 10 and regenerate Tables 6 and 7.

  2. Which of the wilt or vehicle0 rows in Table 4 is correct? The two rows are identical in all 21 cells under TabPFN while the datasets differ in size, dimensionality and imbalance ratio. Please check the raw fold-level logs and state which row was overwritten, and whether the same error reaches Table 10's TabMaDo column.

  3. What value of g_class is used at sampling time? Algorithm 4 line 5 calls Diffuser(x_t, t) with no label. If g_class = 1 + alpha, the Diffuser is class-conditional and V3 (Guider removed) is not the control it is described as; if g_class = 1, the network is run in a regime it was never trained in for minority generation. Please also report the alpha = 0 run, same architecture and Guider, everything else fixed. This single control decides whether the class-aware element of the backbone contributes anything.

  4. Does the implementation sample x_{t-1} = mu* + sigma_t z or mu* + sigma_t^2 z, as Algorithm 4 line 11 is written? Eq. 10 says the former. If the code uses sigma_t^2, all results come from a modified sampler.

  5. Are the Diffuser, the Guider and the prototype p_min = E[x_min] re-fitted on the nine training folds only, inside each CV split? p_min is a sufficient statistic of the minority class; on glass-0-1-4-6-vs-2 the test fold holds 1.7 minority rows on average, so a prototype computed over the full dataset leaks the held-out rows into generation on exactly the dataset where the headline gains are claimed.

  6. How do the authors reconcile the claim "highest average performance on all three metrics" with Table 1 and Table 3, where RelDDPM has the higher mean AUC (0.8810 vs 0.8724 under RF, 0.8933 vs 0.8908 under GBDT)? And with abalone9-18 under RF, where TabMaDo's AUC ranks 4th of 7, against the statement that other methods degrade on high-IR datasets?

  7. In the ablation, why is RQ-1 supported by a single cell (abalone9-18 G-mean, V3 0.3572 vs V4 0.3098) when the full V3 vs V4 contrast across Tables 8-10 is 39 wins for Tabulamba, 48 for the MLP, 3 ties? Do the authors have per-fold scores that would allow a paired test of this contrast, and of TabMaDo against RelDDPM per dataset?

  8. What is the scalar objective whose input gradient is g_t in Algorithm 4 line 7 (log P(x_t), the cosine similarity, or L_class + gamma L_proto at y = y_min), and what value of s was used? s is the single most consequential sampling hyperparameter and is never given.

  9. Were the baselines run at author-recommended settings, retuned per dataset, or at defaults? GOIO reaching F1 = 0.1533 on wilt and 0.2356 on phoneme under GBDT, against 0.85 and 0.76 for plain stable SMOTE, looks like a misconfiguration rather than a property of the method, and a broken baseline inflates both the Friedman statistic and TabMaDo's rank advantage.

  10. What is the relationship to SOS (Kim et al., KDD 2022)? It is a score-based diffusion oversampler for imbalanced tabular data, it is in the bibliography, and it is neither discussed as a competitor nor run as a baseline.

  11. A matched-capacity comparison of Tabulamba vs MLP vs Transformer noise predictors with parameter counts reported and the Guider held fixed, plus a held-out diffusion loss, would make the Sec. 4.5 argument much stronger than the training-MSE comparison in Fig. 5.

  12. A direct diversity measurement of the generated minority rows (nearest-neighbour distances to real minority rows, coverage/density in the sense of Naeem et al.) as a function of s would address the collapse risk that comes from guiding every trajectory toward one prototype vector, and would substantiate the abstract's diversity claim.

  13. A no-oversampling row and a class-weighted RF/GBDT row would let the reader see whether any generative method beats the cheapest alternative.

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

This paper proposes TabMaDo, a dual-guided diffusion framework for imbalanced tabular data oversampling. It combines a Mamba-based noise predictor, Tabulamba, with a Guider that uses class labels and minority-class prototypes to guide the reverse diffusion process. The goal is to improve the quality and diversity of generated minority-class samples under extreme class imbalance. Experiments are conducted on 10 imbalanced binary classification datasets with multiple downstream classifiers.

Critical strengths

1 The paper addresses a practically relevant problem. Combining global distribution learning with targeted minority-class generation is a reasonable approach to handling extreme class imbalance. 2 The experimental evaluation covers multiple datasets and downstream classifiers, with ablation studies, synthetic data quality evaluations, and statistical analyses.

Critical weaknesses

1 The single-prototype guidance mechanism has a fundamental limitation. Minority-class distributions may contain multiple distinct modes, which cannot be adequately represented by a single mean prototype. Guiding generated samples toward a single prototype may compromise diversity and fail to preserve different minority sub-distributions. The paper does not sufficiently investigate this issue. 2 The novelty and effectiveness of the proposed components are not sufficiently established. The ablation studies do not isolate the contribution of the class-aware Mamba mechanism, and the results do not demonstrate statistically significant improvements over strong baselines such as RelDDPM. Furthermore, some aspects of the sampling procedure and theoretical analysis remain unclear.

Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.