acceptodds
Desk rejected at ICLR 2027

Tabular-to-Graph Diffusion: Structured Generative Oversampling for Imbalanced Tabular Data

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Class imbalance presents severe challenges for supervised classification on tabular data. Traditional oversampling techniques synthesize minority records via linear interpolation, which often crosses decision boundaries in non-linear distributions. Tabular diffusion models alleviate this by learning generative densities; however, flattening records into one-dimensional vectors and initiating reverse trajectories from standard Gaussian noise across the full feature space frequently leads to uncalibrated score estimates in low-density regions, resulting in mode drift toward majority classes and damaged column correlations. We propose Tabular-to-Graph Diffusion (TGD), an oversampling framework designed to preserve feature dependencies under acute sample scarcity. TGD structures each minority record alongside its nearest minority neighbors into a local feature graph, where nodes represent tabular columns and node attribute vectors contain feature values across ordered neighboring records. A two-layer residual graph convolutional network updates feature representations using empirical correlation coefficients and a categorical mask that isolates mutually exclusive dummy variables, while linear layers project across neighbor coordinates to capture local variations. Reverse diffusion begins from a localized perturbation of the empirical neighborhood and integrates along deterministic trajectory paths, preventing synthetic records from drifting outside populated minority regions. Evaluations across eight benchmark datasets show that TGD achieves competitive downstream classification performance across heuristic interpolators, deep generative models, and vector-based tabular diffusion methods, establishing a practical trade-off between decision boundary adherence and multi-feature correlation preservation.

Primary area:
generative models
Keywords:
Tabular Data Synthesis, Diffusion Models, Class-Imbalanced Learning, Tabular-to-Graph Transformation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_FRG6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Tabular-to-Graph Diffusion (TGD) for minority-class oversampling. It represents local minority neighborhoods as feature graphs and combines graph convolutions with localized diffusion sampling. Experiments on eight datasets show improved balanced accuracy and competitive performance against traditional oversampling and tabular generative methods.

Critical strengths
  1. I liked the idea of combining local neighborhood information with feature-level graph representations. This is a reasonable way to preserve feature dependencies when minority samples are limited.

  2. The experimental evaluation is fairly comprehensive. The authors compare against several strong baselines and include ablation studies, distribution visualizations, and runtime comparisons.

Critical weaknesses
  1. I am not fully convinced that the improvement comes mainly from the proposed graph representation. The method also uses localized warm-start sampling and post-processing, which can independently improve generation quality. Although the paper includes ablations, a matched vector-diffusion baseline using the same sampling strategy would make the contribution clearer.

  2. The improvement over SMOTE seems relatively limited given the additional model complexity. Although TGD achieves better average ranks, the Nemenyi analysis does not establish a significant difference from SMOTE. The statistical analysis also treats multiple correlated metrics from the same datasets as separate tasks, which makes the reported significance harder to interpret.

Questions for authors
  1. How much of the improvement comes from the graph representation rather than localized sampling? Could a standard vector diffusion model with the same warm-start strategy achieve similar results?

  2. Since TGD does not significantly outperform SMOTE under the reported Nemenyi test, what is the main practical advantage that justifies the additional computational cost?

  3. Would the statistical conclusions change if significance were evaluated across the eight independent datasets rather than treating all 24 dataset-metric combinations as separate tasks?

4 comments
AuthorsOpenReview ↗

### General Response and Summary of Revisions

Thank you for your constructive review and insightful questions on attribution, practical utility, and statistical rigor. Following your suggestions, we have revised the manuscript to include:

(1) a strictly aligned Matched 1D Vector Diffusion ablation across all eight benchmark datasets (sharing the exact same localized warm-start perturbation twarm=30t_{\text{warm}}=30, deterministic PF-ODE integration, and post-processing) to isolate the contribution of the transposed feature-graph representation (Appendix H.3, Figure 11, Tables 7–8);

(2) block-independent non-parametric statistical tests evaluated strictly across the N=8N=8 independent datasets to eliminate inter-metric correlation, yielding p=0.0156p = 0.0156 against SMOTE and p=0.0078p = 0.0078 against the remaining baselines under two-sided Wilcoxon signed-rank tests (Section 5.1; Appendix C.5);

(3) relative net gain analyses over unaugmented training (Δ\Delta in Appendix F.3, Table 4) alongside high-dimensional scalability benchmarks up to D=299D=299 (Appendix H.4, Table 9), further clarifying TGD's practical advantages over linear interpolation in boundary adherence, multi-feature covariance preservation, and categorical validity. Below are our point-by-point responses.

AuthorsOpenReview ↗

### Response to W1 & Q1: Comparing Full TGD against 1D Vector Diffusion under Matched Sampling

Thank you for raising this point regarding performance attribution. To investigate whether a standard vector diffusion model with localized sampling could achieve comparable performance (Q1), we implemented the exact matched 1D vector baseline in Appendix H.3 (Tables 7–8).

Our empirical results show that it falls short: even with identical localized warm-start perturbation and post-processing, standard vector diffusion degrades, with its overall mean rank dropping from 1.645 to 6.479 across all eight benchmark datasets.

The strictly matched comparison from the extended ablation suite (Appendix H.3, Tables 7–8) is summarized below:

Method / VariantFeature ArchitectureOverall Mean Rank (Tables 7–8)Phoneme (D=5D=5) Bal-Acc (Rank)Churn (D=33D=33) Bal-Acc (Rank)Ol8 (D=72D=72) Bal-Acc / AUC / F1 (Rank)
Full TGD (Ours)Transposed Patch + GCN1.6450.843 (1.000)0.830 (1.000)0.741 / 0.817 / 0.388 (2.333)
Matched 1D Vector Diff.Flat Vector + MLP6.4790.833 (8.667)0.814 (7.000)0.716 / 0.792 / 0.371 (7.667)
  1. Controlled Experimental Alignment (Appendix H.3): To isolate architectural contributions from localized sampling and post-processing, the Matched 1D Vector baseline strictly shares:
  • Equivalent Capacity: A standard residual MLP backbone with matched hidden dimensions (dh=64d_h = 64) operating on flattened 1D records x∈RD\mathbf{x} \in \mathbb{R}^D;
  • Identical Sampling: The exact same localized warm-start perturbation (twarm=30t_{\text{warm}} = 30, T=100T = 100) and deterministic reverse PF-ODE integration (η=0.0\eta = 0.0);
  • Identical Post-Processing: Empirical moment matching, coordinate interval clamping, and categorical Argmax discretization.
  1. Disentangling the Gain Hierarchy (Connecting Table 6 and Tables 7–8): Analyzing these ablation suites reveals a clear, complementary two-level contribution hierarchy across representation and message passing:
  • Level 1 (Local Coordinate Smoothing via Transposed Patch): In the representation ablation suite (9 variants, Tables 7–8), transitioning from flat 1D vectors to structured transposed patches accounts for the primary performance gain, improving the overall mean rank from 6.479 (Matched 1D) to 1.645 (Full TGD). Organizing ordered local neighbors along the coordinate axis allows linear projection layers (Win,Wout\mathbf{W}_{\text{in}}, \mathbf{W}_{\text{out}}) to smooth variations across neighboring minority records, which isolated 1D MLPs cannot capture.
  • Level 2 (Column Covariance Regularization via Graph Convolutions): In the architectural ablation suite (7 variants, Table 6), incorporating correlation-guided GCN layers over the transposed patch provides the remaining critical gains, advancing overall rank from 4.667 ('w/o Graph') to 2.291 (Full TGD). As analyzed in Section 5.3, while neighbor-axis projection alone maintains local support, graph message passing explicitly penalizes discordant column updates (minimizing Dirichlet energy), preventing the severe high-dimensional covariance degradation observed in flat diffusion (e.g., raising Balanced Accuracy on Ol8 from 0.716 to 0.738 in Table 6, versus 0.716 for Matched 1D and 0.741 for Full TGD in Table 8).

These comparisons verify that localized sampling provides a reliable neighborhood anchor, but the transposed graph architecture is essential for preserving multi-feature dependencies and achieving top-tier classification fidelity. Complete quantitative metrics are reported in Appendix H.1 (Table 6) and Appendix H.3 (Tables 7–8).

AuthorsOpenReview ↗

### Response to W2 & Q2: Practical Advantages over SMOTE: Decision Boundaries and Feature Correlations

Thank you for this perceptive assessment regarding the comparison with SMOTE and the statistical evaluation protocol.

We fully share your respect for SMOTE. In tabular learning, classical heuristics like SMOTE remain remarkably competitive, mirroring the ongoing dynamic between tree-based ensembles (GBDTs) and deep tabular architectures. While neural diffusion inevitably introduces higher computational overhead than linear interpolation, this modest cost delivers essential geometric and structural capabilities that heuristic methods cannot provide.

The practical advantages of TGD over SMOTE, alongside our updated statistical evaluation, are summarized below:

Evaluation DimensionSMOTE (Heuristic Baseline)TGD (Transposed Graph Diffusion)Practical Advantage of TGD
Statistical Significance (N=8N=8, Bal-Acc)Mean Rank: 2.500 (24-Task: 3.167)Mean Rank: 1.125 (24-Task: 1.875)Reject H0H_0 (p=0.0156p = 0.0156, Wilcoxon)
Boundary FidelityLinear straight-line segmentsBounded reverse ODE trajectoriesEliminates crossing into majority clusters
Feature CovarianceUnmodeled (independent axes)Graph convolutions (τ=0.8\tau = 0.8)Preserves multi-column dependencies
Categorical ValidityFractional interpolation valuesMcat\mathbf{M}_{\text{cat}} + Categorical ArgmaxEnforces strict one-hot mutual exclusivity
Computational Footprint< 1 second (CPU)14.6s–92.5s (D≈300D \approx 300, Laptop GPU)Real-time feasible without pre-training
  1. Statistical Significance on N=8N=8 Independent Datasets (p=0.0156p = 0.0156): The global Nemenyi test is conservative when controlling family-wise error across all 45 algorithm pairs (k=10k=10). When evaluating Balanced Accuracy strictly across the N=8N=8 independent dataset blocks, TGD achieves an average rank of 1.125 (ranking 1st on 7 datasets and 2nd on SD) compared to 2.500 for SMOTE, yielding p=0.0156p = 0.0156 (W=34.0W = 34.0) against SMOTE under a two-sided Wilcoxon signed-rank test (Appendix C.5). Furthermore, evaluating net improvements across all 24 evaluation tasks (Appendix F.3, Table 4) confirms that TGD maintains the leading overall rank of 1.875 with an average gain of Δ‾=+0.0299\overline{\Delta} = \mathbf{+0.0299} over raw training, compared to 3.167 and +0.0251+0.0251 for SMOTE.

  2. Elimination of Decision Boundary Violations (Boundary Fidelity): SMOTE connects neighboring minority instances via unconstrained straight lines. On non-convex or multi-modal distributions, these segments cut across intermediate majority-class valleys, introducing severe label noise (Figure 4 and Appendix Figure 8). On datasets like SD and BCCP, this causes synthetic points to scatter into majority territory, dropping SMOTE's F1 rank to 8th on SD. In contrast, TGD perturbs local patches at twarm=30t_{\text{warm}} = 30 and integrates inward, confining synthetic samples strictly within authentic minority clusters.

  3. Preserving Feature Correlations and Discrete States: Linear interpolation treats attributes independently, failing to preserve column covariance and producing fractional floats on one-hot indicators. TGD preserves pairwise feature dependencies via correlation-guided GCN layers and strictly enforces discrete validity via categorical masking Mcat\mathbf{M}_{\text{cat}} and post-hoc Argmax discretization, scaling reliably up to 299 features (Musk and Scene, Appendix H.4, Table 9).

  4. Practical Runtime: Unlike global tabular diffusion models that incur quadratic attention overhead across columns, TGD operates on sparse local graphs (τ=0.8\tau = 0.8) without pre-training. Computational cost is governed primarily by dataset sample size and the target synthesis volume rather than feature dimensionality DD. Consequently, TGD scales efficiently to high-dimensional tables, requiring only 14.62 seconds on Scene (D=299,N=2,407D=299, N=2,407) and 92.52 seconds on Musk (D=268,N=6,598D=268, N=6,598) on a standard consumer RTX 3060 Laptop GPU (Appendix H.4, Table 9).

AuthorsOpenReview ↗

### Response to Q3: Statistical Significance Evaluated Across N=8 Independent Datasets

Thank you for pointing out the issue of metric correlation. In our original formulation, evaluating across 24 dataset-metric tasks was intended to provide a comprehensive, multi-faceted benchmark overview across complementary performance criteria (macro balance, global ranking, and minority precision/recall trade-offs). Nevertheless, we fully agree with your perspective that from a rigorous mathematical statistics standpoint, eliminating potential inter-metric co-dependencies is essential.

Accordingly, in Appendix C.5, we have re-executed the statistical testing protocol by treating the eight distinct datasets strictly as N=8N=8 independent evaluation blocks (k=10k=10), and find that our statistical conclusions remain consistent.

The re-evaluated non-parametric statistical tests across the N=8N=8 independent datasets are summarized below:

Statistical Test ProtocolEvaluation ScopeTest Statisticp-valueDecision / Verdict
Friedman TestBalanced Accuracy aloneχF2=60.33\chi_F^2 = 60.33p=1.16×10−9\mathbf{p = 1.16 \times 10^{-9}}Reject H0H_0 (Difference among 10 methods)
Friedman TestAUC-ROC aloneχF2=34.77\chi_F^2 = 34.77p=6.53×10−5\mathbf{p = 6.53 \times 10^{-5}}Reject H0H_0 (Difference among 10 methods)
Friedman TestDataset-Level Mean RankχF2=47.26\chi_F^2 = 47.26p=3.51×10−7\mathbf{p = 3.51 \times 10^{-7}}Reject H0H_0 (Difference among 10 methods)
Wilcoxon Signed-RankTGD vs. SMOTE (Bal-Acc)W=34.0W = 34.0p=0.0156\mathbf{p = 0.0156}Reject H0H_0 at α=0.05\alpha = 0.05 (p<0.02p < 0.02)
Wilcoxon Signed-RankTGD vs. Other 8 BaselinesW=36.0W = 36.0p=0.0078\mathbf{p = 0.0078}Reject H0H_0 at α=0.05\alpha = 0.05 (p<0.01p < 0.01)
  1. Multi-Classifier ANOVA (Friedman): Whether testing Balanced Accuracy alone, AUC alone, or collapsing all metrics into a single unweighted dataset-level rank, the Friedman test rejects the null hypothesis of equivalent performance across the 10 algorithms (p≤6.53×10−5p \le 6.53 \times 10^{-5}).

  2. Pairwise Comparisons against Baselines (Wilcoxon): Treating datasets as independent blocks, the two-sided Wilcoxon signed-rank test on Balanced Accuracy yields W=34.0W = 34.0 with p=0.0156p = 0.0156 (p=0.0078p = 0.0078 under one-sided testing) against SMOTE, rejecting the null hypothesis at α=0.05\alpha = 0.05. Against each of the remaining eight baselines (Original, ADASYN, CTGAN, TVAE, TabDDPM, GOGGLE, TabSyn, and TabDiff), the test yields W=36.0W = 36.0 with p=0.0078p = 0.0078 (two-sided).

Therefore, eliminating inter-metric correlation and restricting statistical tests strictly to independent dataset blocks leads to the same conclusions (p<0.02p < 0.02 across all pairwise comparisons). Full details are documented in Appendix C.5.

Reviewer_7pMVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Tabular-to-Graph Diffusion (TGD), a generative oversampling framework for imbalanced tabular classification. TGD constructs local minority-neighborhood matrices, transposes them into feature graphs, and employs correlation-guided graph convolutions with categorical masking to preserve feature dependencies. A deterministic diffusion sampling strategy generates synthetic minority observations from locally perturbed neighborhoods. Experiments across eight datasets demonstrate competitive downstream classification performance against interpolation-based, deep generative, and diffusion-based oversampling methods. This is a decent paper from a tabular generative perspective. Eventually, TGD structures each minority record alongside its nearest minority neighbors into a local feature graph, where nodes represent tabular columns and node attribute vectors contain feature values across ordered neighboring records.

Critical strengths
  • Transforming local tabular neighborhoods into feature graphs offers an interesting approach to preserving feature dependencies during generative oversampling.
  • Localized deterministic diffusion and categorical masking provide useful mechanisms for limiting unrealistic samples under minority-class scarcity.
  • TGD achieves the best overall average rank across eight datasets, particularly demonstrating strong balanced accuracy.
  • The paper includes architectural ablations, covariance comparisons, distribution visualizations, and nearest-record distance analyses, providing useful evidence for its design choices.
Critical weaknesses
  • Limited investigation of ordering strategies. Sec. 3.1, L149-L152, only Euclidean distance is used for minority neighbors for forming a local matrix which is sorted in ascending order only to get an ordered local matrix. The authors should provide an ablation on this choice. It sorts neighboring records by ascending Euclidean distance naively without investigating alternative neighborhood orderings. Since the subsequent projection layers operate on these ordered coordinates, the ordering choice may affect representation quality. Structure-aware ordering approaches, such as GO-LR [1], could motivate additional investigations, although feature ordering and neighborhood ordering are distinct problems.

[1]. Habib et al. "GOTabPFN: From Feature Ordering to Compact Tokenization for Tabular Foundation Models on High-Dimensional Data." ICML 2026.

  • Unexplored distance metrics. The local neighborhood construction relies exclusively on Euclidean distance. There are other alternatives like Manhattan distance, KL Divergence, cosine distance etc. Please see [1] for more insight on this. Alternative metrics, including Manhattan, cosine, or appropriately defined distribution-aware distances, could produce different neighborhood structures. An ablation would strengthen the justification for this choice.

  • Limited intrinsic dimensionality analysis. Sec. 3.2, the paper mentions intrinsic dimensionality and provides a rule. However, can’t we also find this using PCA? It adopts TwoNN for estimating intrinsic dimensionality and selecting neighborhood sizes. However, the paper does not compare this strategy with alternatives such as PCA-based dimensionality estimation [1] or other local intrinsic dimension estimators. Furthermore, the estimated dimension determines a dataset-level neighborhood size rather than a genuinely adaptive size for each local patch.

  • Insufficient high-dimensional evaluation. The paper uses only 8 datasets and all of them are low-dimensional (Table 1). It would be interesting to see how TGD does in high-dimensional settings. These 8 datasets with 5-72 features show a possible loop hole in high-dimensionality for TGD. Consequently, the claims concerning high-dimensional feature dependencies and computational scalability are not adequately validated. Please see CtrTab [2] and BSTabDiff [3] for reference to some high-dimensional datasets from tabular generative perspective for further evaluation.

[2]. Li, Zuqing, Junhao Gan, and Jianzhong Qi. "Towards Synthesizing High-Dimensional Tabular Data with Limited Samples." AAAI 2026.

[3]. Habib et al. "BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation." NeurIPS 2026 / ICLR 2026 DeLTa Workshop

  • Incomplete diffusion baselines. The paper omits several relevant tabular generative approaches. Some other diffusion-based baselines are omitted or at least need to be discussed. Please refer to BSTabDiff [3], BinaryDiffusion [4], and ForestDiffusion [5] for more related baselines including CtrTab [2] Although these methods are not all designed specifically for minority oversampling, discussing them and evaluating suitable adaptations would strengthen the comparisons.

[4]. Kinakh, Vitaliy, and Slava Voloshynovskiy. "Tabular Data Generation Using Binary Diffusion." NeurIPS 2024 Third Table Representation Learning Workshop

[5]. Jolicoeur-Martineau, Alexia, Kilian Fatras, and Tal Kachman. "Generating and Imputing Tabular Data via Diffusion and Flow-based Gradient-Boosted Trees." AISTATS 2024.

  • Limited theoretical guarantees. The paper provides a computational complexity analysis but limited formal characterization of how the proposed graph representation preserves feature dependencies or prevents distributional drift. Furthermore, the claimed computational advantages depend on graph sparsity, while correlation matrix construction still incurs quadratic complexity in the number of features.

  • Theoretical characterization of neighborhood ordering. Section 3.1 adopts a simple Euclidean distance-based ordering of minority neighbors. Could the authors formulate an optimal neighborhood-sequencing objective analogous to the Traveling Salesman Problem (TSP) and investigate whether the generalized ordering problem is NP-Hard? Establishing such a connection through a formal reduction, together with suitable approximation or heuristic strategies, could strengthen the theoretical foundation of TGD. Right now the paper misses a theoretical connection from Computer Science perspective but that can be recovered from NP-Hardness proof via TSP equivalency (if the authors can do it in rebuttal).

  • Statistical evaluation concerns. The Friedman and Nemenyi analyses treat 24 dataset metric combinations as evaluation tasks, although the three metrics calculated on each dataset are correlated. This may overstate the effective number of independent evaluation blocks. Additionally, the manuscript refers to macro F1, whereas Appendix C presents the conventional binary positive class F1 formula, requiring clarification.

  • Minor presentation issues. The caption for Table 1 is placed at the bottom of this table, which is not exactly the ICLR formatting rule. Some explanations regarding neighborhood construction and computational complexity could also be clearer.

Questions for authors

Please see the Critical Weaknesses and refer to the following questions. For initial decision, I am giving it a benefit of doubt and hope the authors can turn it around in rebuttal.

  1. Could alternative neighborhood ordering strategies improve TGD? Have the authors considered structure-aware ordering approaches inspired by GO-LR [1] or other optimization methods?
  2. Why was Euclidean distance selected exclusively? Can the authors provide ablations using Manhattan, cosine, KL divergence or other suitable distance measures?
  3. How does TwoNN compare with PCA-based or other intrinsic dimensionality estimators for selecting neighborhood sizes (e.g. [1] did PCA based)?
  4. Can the authors extend evaluations to genuinely high-dimensional datasets, such as those considered in CtrTab [2] and BSTabDiff [3]?
  5. Could additional diffusion-based baselines, including BinaryDiffusion [4] and ForestDiffusion [5], be considered or discussed?
  6. Can the authors provide stronger theoretical justification for dependency preservation and clarify the worst case computational complexity?
  7. Can the authors clarify the statistical testing protocol and whether the reported F1 metric is macro-averaged or minority-class F1?
  8. Could the authors formulate neighborhood ordering as a TSP-like optimization problem and investigate its NP-hardness through a formal reduction?
14 comments
AuthorsOpenReview ↗

### General Response and Summary of Revisions

Thank you for your thorough review and encouraging feedback to further strengthen our work. Following your suggestions, we have revised the manuscript to include: (1) full comparisons against all four recent tabular diffusion baselines (ForestDiffusion, BinaryDiffusion, CtrTab, and BSTabDiff) across all 10 datasets, including the high-dimensional Musk (D=268D=268) and Scene (D=299D=299) benchmarks (Appendix H.4, Table 9); (2) systematic ablations on neighborhood ordering strategies (inspired by GO-LR), alternative distance metrics, and PCA-based dimensionality estimation across all eight core datasets (Appendix H.3, Tables 7–8); and (3) a formal NP-hardness proof for optimal neighborhood sequencing via Metric Path-TSP (Appendix H.6, Theorem 1), alongside clarified complexity bounds, minority-class F1 definitions, and N=8N=8 block-independent statistical tests (Appendices C.5 and E.6). Below are our point-by-point responses.

AuthorsOpenReview ↗

### Response to W1 & Q1: Ablation of Neighbor Ordering Strategies and Choice of Ascending Sorting

We agree that neighbor ordering interacts with linear projection layers along the coordinate axis. Following your suggestion and GO-LR (Habib et al., 2026), we ablated four ordering strategies across all 8 datasets (Referenced in Section 4; content in Appendix H.3, Tables 7–8):

Ordering StrategyOverall Mean RankNamG Bal-AccNamG F1NamG Rank
Ascending Euclidean (Default)1.6450.9200.4763.667
Descending Euclidean5.5000.9100.4758.000
Random Permutation5.1250.9070.4787.333
TSP-Greedy (Structure-Aware)5.1670.9230.4961.333
  • When Alternative Ordering Improves TGD (Answering Q1): Yes. On the narrow, curved minority distribution of NamG (42:1 imbalance), TSP-Greedy achieves the top rank (1.333), confirming your intuition that shortest-path sequencing preserves step-by-step local continuity.
  • Rationale for Default Ascending: Unlike sequence-based tabular architectures where columns must be serialized into 1D ordered inputs, TGD represents features as graph nodes where message passing is naturally permutation-equivariant across columns. Ascending distance sorting provides an invariant coordinate frame for Win\mathbf{W}_{\text{in}} (coordinate 1 = anchor, coordinate kk = kk-th neighbor) and runs in O(N∗log⁡N∗)\mathcal{O}(N^* \log N^*), outperforming TSP heuristics across the broader benchmark suite.
AuthorsOpenReview ↗

### Response to W2 & Q2: Comparison of Distance Metrics (Manhattan, Cosine) and L1 Distance

We evaluated replacing Euclidean distance with Manhattan (L1L_1) and Cosine distances across all 8 datasets (Appendix H.3, Tables 7–8). The empirical comparison is summarized below:

Distance MetricOverall Mean Rank (8 Datasets)Ol8 Bal-AccOl8 AUCOl8 F1Ol8 Rank
Euclidean (L2L_2, Default)1.6450.7410.8170.3882.333
Manhattan (L1L_1)4.3750.7480.8320.4001.000
Cosine Distance4.8540.7360.8170.3912.667
  1. Empirical Findings Across 8 Benchmarks (Appendix H.3):
  • Global Robustness: Euclidean distance achieves the best overall mean rank of 1.645 across the 8 datasets, outperforming Manhattan (4.375) and Cosine (4.854).
  • High-Dimensional Utility of Manhattan Distance: On the 72-dimensional Ol8 benchmark, Manhattan distance achieves the top rank of 1.000 (improving Bal-Acc from 0.741 to 0.748). This confirms metric space theory, as the L1L_1 norm mitigates distance concentration in higher dimensions.
  1. Design Rationale for Euclidean Default & Applicability of Alternatives:
  • Alignment with Isotropic Diffusion: Features are standardized to zero mean and unit variance. In this standardized space, Euclidean distance corresponds to the negative log-likelihood under identity covariance, matching isotropic Gaussian noise injection (ϵ∼N(0,I)\boldsymbol{\epsilon} \sim \mathcal{N}(\mathbf{0}, \mathbf{I})) and reverse ODE trajectories.
  • Preserving Absolute Feature Scale: Tabular features carry critical diagnostic meaning in absolute coordinate deviations (e.g., severe lab anomalies). Cosine distance normalizes vector norms, discarding magnitude variations between records.
  • Why KL Divergence is Inapplicable: KL divergence operates strictly on probability distributions, whereas tabular rows x∈RD\mathbf{x} \in \mathbb{R}^D are standardized coordinate vectors containing negative values. Using KL divergence would require either non-parametric density estimation per row (which suffers from acute estimation variance under sample scarcity) or artificial softmax normalization across heterogeneous variables, distorting physical column relationships.
Reviewer_QeMKOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new method for oversampling of tabular data. Oversampling is method of addressing class imbalance in classification tasks, where new samples are generated for the minority class(es). The problem is that for tabular data, standard interpolative methods (e.g., convex combinations of minority samples) can generate out-of-distribution synthetic points, where as conventional deep generative models struggle to learn realistic examples conditional on the rare classes. The authors propose a method of very strong constraints on the generative model based on the neighborhood of minority samples (within the minority class). The constraints are generated by only maintaining covariate relations that are very strong before diffusion and by postprocessing with domain constraints. The authors compare their methods with other generators, and show that on a set of tabular datasets it leads to the best classification results after oversampling the minority. (Surprisingly, the best in line is the relatively simple "smote" extrapolator).

Critical strengths

The paper addresses an important problem for classification for (or alternatively, generation of minority-class samples in ) tabular data, and seems to make progress in solving it. We have here a novel approach for light-weight and locally coherent sampling that is more productive the linear interpolation, but stays very close to the sample manifold. The three main components are extremely agressive pruning of the codependence graphs, an flexible neighborhood determination scheme, and the post processign. These can be identified using the informative abelation study.

The paper also exposes the problem by benchmarking the behavior of current generative models, and showing that they often provide only limited assitance in minority class sampling. This is an important exposition to the problem, and is therefore of valoue.

Critical weaknesses

One thing I am missing is more insight about what goes right and wrong about this method and its competitors. The method, its competitors and the datasets are given as is (with the caveat of the abelation study). For example, I'm thinking of the extremely strict corr>0.8 threhsold used in graph estimation within the small neighborhoods. I'm thinking more generally about how the methods fair as we fix a dataset, with or without categorical features, and vary the sample size or the imbalance ration.

Questions for authors
  1. I am not completely sure I understand what parts of the algorithm are "pretrained" specifically: Are there any parameters learned jointly over a larger subset of the dataset, or is everything recomputed locally on N* datasamples at a time.

  2. I'd be interseted to know how does changing the size of the imbalance (for a specific dataset) would change the accuracy of this method compared to other methods. Currently we have a sample of points, ratio's etc and it is not easy to tell apart the different methods.

  3. Do any of the method papers mentioned in parag 2 and 3 of the introduction, that discuss class imbalance, have methods for minority class sampling ? I see that the comparison here is only with general purpose diffusion models? or there was some other criterion for inclusion

Suggestions:

  • current umap is very hard to read, both color-wise and in terms of the story. It is too colorful, hard to tell apart true points from samplers.

  • Table 2 is also dense and hard to read. Consider removing or minimizing the report of the sds, sorting the datasets by some interesting attribute that may correlate to results, subtracting out original, or retaining only a single classification method rather than 3 in the main results.

11 comments
AuthorsOpenReview ↗

### General Response and Summary of Revisions

Thank you for your constructive review and practical suggestions.

We agree with your observation that "Surprisingly, the best in line is the relatively simple 'smote' extrapolator." In imbalanced tabular learning, classical heuristics such as SMOTE remain highly competitive. This observation mirrors the broader tabular domain, where tree-based ensembles continue to perform strongly alongside deep neural networks. Classical methods maintain genuine practical value, and we hold sincere respect for these foundational baselines. At the same time, exploring new modeling paradigms is essential for progress. As shown by recent diffusion models discussed in our Introduction and Related Work, generative modeling provides an alternative approach to geometric linear interpolation. Building on these prior efforts, TGD investigates how localized structural priors can make diffusion models practical and reliable under minority sample scarcity.

To address your questions and suggestions, we have conducted additional experiments and updated the manuscript accordingly. Below are our point-by-point responses.

AuthorsOpenReview ↗

### Response to W1: Effects of Correlation Threshold , Imbalance Ratios, and Categorical Features

To directly address your concerns regarding the correlation threshold and examine how different methods behave as data conditions vary, we have conducted two sets of sensitivity experiments across three representative datasets (Phoneme, Ol8, and NamG), with full quantitative results reported in Appendix H.5 (Tables 10 and 11).

We summarize the core empirical findings and representative subsets below:

  1. Sensitivity to Correlation Threshold tau (Appendix H.5, Table 11) In small local patches (N∗N^*), sample correlations naturally exhibit higher estimation variance. Setting τ=0.8\tau = 0.8 serves as a structural noise filter:
  • Relaxing tau to 0.5 increases graph edge density to 20%–44%, admitting noisy spurious correlations that slightly degrade performance.
  • Tightening tau to 0.9 reduces edge density to 2%–8%, pruning valid feature dependencies.
  • As shown below on Phoneme (D=5) and Ol8 (D=72), setting τ=0.8\tau = 0.8 maintains edge density within a sparse 6.7%–17.0% window, filtering noise while ensuring message-passing complexity scales linearly with active edges:
DatasetSettingEdge Density (%)Bal-AccF1
Phoneme (D=5D=5)τ=0.5\tau = 0.544.33%0.8380.757
τ=0.8\tau = 0.8 (Default)17.03%0.8430.767
τ=0.9\tau = 0.98.84%0.8390.758
Ol8 (D=72D=72)τ=0.5\tau = 0.519.97%0.7250.380
τ=0.8\tau = 0.8 (Default)6.76%0.7360.384
τ=0.9\tau = 0.93.34%0.7290.387
  1. Varying Imbalance Ratios from 10:1 to 100:1 (Appendix H.5, Table 10) We systematically downsampled minority training records while fixing the majority training partition and untouched test sets:
  • What goes wrong for vector diffusion: Unconstrained models (TabDDPM, TabDiff) degrade sharply under acute scarcity. On the 72-dimensional Ol8 benchmark at IR = 100:1 (where only M = 16 minority training instances remain), score estimates across 72 unpopulated dimensions suffer severe trajectory drift into majority regions, collapsing F1 scores to 0.065 and 0.046.
  • What goes right for TGD: Initializing reverse trajectories from localized neighborhood perturbations (twarm=30t_{\text{warm}} = 30) keeps generation anchored within the empirical minority support, sustaining an F1 score of 0.303 and Balanced Accuracy of 0.619:
MethodOl8 (IR = 20:1, M = 83) Bal-Acc / F1Ol8 (IR = 50:1, M = 33) Bal-Acc / F1Ol8 (IR = 100:1, M = 16) Bal-Acc / F1
SMOTE0.699 / 0.3370.662 / 0.3040.634 / 0.278
TabDDPM0.625 / 0.2740.523 / 0.0880.514 / 0.065
TabSyn0.666 / 0.2830.601 / 0.2090.600 / 0.221
TabDiff0.582 / 0.2330.504 / 0.0480.506 / 0.046
TGD (Ours)0.706 / 0.3640.643 / 0.3260.619 / 0.303
  1. Behavior on Datasets With vs. Without Categorical Features
  • With categorical features (e.g., Churn, OCI, BCCP): Standard vector diffusion models treat one-hot dummy columns as independent continuous variables, failing to preserve discrete mutual exclusivity under small sample sizes. Standard GCNs also blur discrete boundaries by propagating messages across indicators belonging to the same attribute. TGD addresses this via two explicit mechanisms: (i) the categorical mask Mcat\mathbf{M}_{\text{cat}} blocks message passing between mutually exclusive indicators (Section 3.3), and (ii) post-hoc Argmax discretization restores continuous outputs into valid single-active one-hot states (Section 3.5). Omitting Mcat\mathbf{M}_{\text{cat}} causes an immediate rank drop (e.g., mean rank falling from 1.333 to 5.333 on BCCP in Table 6).
  • Without categorical features (purely continuous, e.g., Phoneme, Ol8, Musk): Heuristic SMOTE interpolates straight lines that can cut across non-convex decision boundaries into majority clusters (Figure 4), while flat vector diffusion struggles to maintain pairwise column covariances (CMD exceeding 28 on Ol8 in Figure 5). On purely continuous tables, TGD naturally simplifies to Pearson correlation graphs, where neighbor-axis projections smooth local coordinate variations and graph convolutions preserve pairwise column correlations without boundary crossing.

Full experimental results across all evaluated ratios (10:1, 20:1, 50:1, 100:1) and all three benchmark datasets are detailed in Appendix H.5 (Tables 10 and 11).

AuthorsOpenReview ↗

### Response to Q1: Training Setup: Shared Denoising Network across All Local Patches

There is no separate pre-training stage in TGD, and the denoising network is not re-trained locally for each patch. Instead, a single unified denoising network is trained jointly across all local patches extracted from the minority partition Xmin\mathbf{X}_{\text{min}}.

The parameter learning and structural construction are organized as follows:

  1. Globally Estimated and Learned Components:
  • Dataset-Level Neighborhood Scale (N∗N^*): The window size N∗N^* is estimated once across the entire standardized minority partition Xmin\mathbf{X}_{\text{min}} using the TwoNN algorithm (Section 3.2), providing a shared attribute dimension for the whole dataset.
  • Shared Denoising Backbone: A single set of neural network weights is shared across the entire dataset and trained end-to-end via mini-batch gradient descent (Section 3.4). These weights include the input and output linear projections (Win,Wout\mathbf{W}_{\text{in}}, \mathbf{W}_{\text{out}}), the two residual GCN layers (Wg(l)\mathbf{W}_{\text{g}}^{(l)}), and the sinusoidal time embedding MLP.
  1. Locally Constructed Structures (Per Anchor Record): What is computed locally for each minority anchor record is strictly the graph data representation (Section 3.1): retrieving its N∗N^* nearest minority neighbors, transposing the patch H(0)=Xloc⊤∈RD×N∗\mathbf{H}^{(0)} = \mathbf{X}_{\text{loc}}^\top \in \mathbb{R}^{D \times N^*}, and computing the correlation-filtered adjacency matrix A~mask\mathbf{\tilde{A}}_{\text{mask}}.

During training, these local patches act as individual graph instances within mini-batches to optimize the shared network. During sampling, reverse diffusion is executed on each perturbed local patch using the trained, frozen network to synthesize balanced minority records (Section 3.5).

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

Related papers

Discussion (0)

Sign in to comment.