Pith. sign in

REVIEW 5 major objections 5 minor 33 references

Large Language Models for Imbalanced Classification: Diversity makes the difference

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that an LLM-based oversampling method that conditions on label plus features, fixes the label at the start, and interpolates minority samples produces more diverse synthetic data and outperforms eight baselines on imbalance

desk verdict Promising LLM oversampling recipe, but the fine-tuning data definition has a load-bearing hole (interpolated rows lack categoricals) and the theory is heuristic; worth a serious review with major revision. read the letter →

arxiv 2510.09783 v2 pith:SXYCEOSM submitted 2025-10-10 cs.LG cs.AIstat.ML

Dang Nguyen , Sunil Gupta , Kien Do , Thin Nguyen , Taylor Braund , Alexis Whitton , Svetha Venkatesh This is my paper · ORCID
classification cs.LGcs.AIstat.ML
keywords imbalancedclassificationoversamplinglargelanguagemodelstabulardatasyntheticgenerationdiversityentropyinterpolation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that the right way to use a language model for oversampling is to make the generated minority samples diverse, not merely plausible. It proposes ImbLLM, built from three changes to the standard LLM pipeline: prompt the model with a minority label plus a randomly chosen feature-value pair, keep the label at the beginning of the sentence while permuting only the features during fine-tuning, and fine-tune on minority samples plus linearly interpolated continuous values instead of on the full imbalanced dataset. If the paper is right, these cheap changes let a classifier trained on rebalanced data nearly match a classifier trained on the full original data, using only 20 percent of the original minority samples, and they beat eight competing oversamplers on F1 and AUC. The paper also gives entropy-based arguments that each modification increases the entropy of the model's output distribution, which is its formal way of saying the method generates more diverse data.

What carries the argument

The central object is the conditional distribution of a fine-tuned decoder-only LLM over tabular rows expressed as text. Three mechanisms modify it: (1) condition_yx, prompting with 'Y is y_minor, X_i is v_i' to create a mixture of context-dependent next-token distributions; (2) fix_y, permuting only features and placing the label at the sentence start so the attention mask lets the label attend to all features; (3) minor+interpolate, fine-tuning only on minority samples plus convexly interpolated continuous values, which the paper likens to smoothing the empirical distribution. The supporting theoretical machinery is three entropy proofs: Jensen's inequality on a mixture of distributions, a

What would settle it

Fine-tune the same pre-trained LM on the same data with fix_y versus permute_xy, prompt both models with only the minority-label sentence, generate a large batch of continuations, and compare the empirical entropy of the generated feature sequences; if the permute_xy model's entropy is not below the fix_y model's, Proposition 2 is false. A second check: run the full ImbLLM pipeline but replace condition_yx with condition_y; if F1 does not drop, diversity of prompts is not the driver claimed.

Watch

Extended reading notes

Core claim

ImbLLM outperforms eight baselines (traditional, GAN-based, and LLM-based) on 10 tabular datasets for imbalanced classification, achieving the best average F1 of 0.8002 versus 0.7724 for the runner-up CTGAN and 0.7238 for training on the imbalanced data without oversampling. The synthetic minority samples score higher on close-probability and coverage, and their distance-to-closest-record distribution is concentrated near zero relative to unseen test minority samples, indicating the generated data cover the real minority manifold. The paper claims each of the three design choices—condition_yx, fix_y, and minor+interpolate—contributes to this outcome, with three entropy propositions intended

Load-bearing premise

The proof of Proposition 2 assumes that after permute_xy fine-tuning, the label in the middle of the context acts as an 'attentional anchor' that makes the next-token distribution nearly independent of the preceding prefix, so the marginal distribution over unseen prefixes stays low-entropy; that assumption is asserted rather than derived, and the claimed entropy ordering between the two fine-tuning strategies collapses without it.

Editorial extensions

If this is right

  • A classifier trained on ImbLLM-rebalanced data nearly reaches the F1/AUC of one trained on the full original data, despite seeing only a fifth of the original minority samples.
  • The verification step used by TapTap and Pred-LLM is unnecessary and can backfire: when the imbalanced-trained classifier is weak, rejecting synthetic samples removes useful diversity, and ImbLLM's no-verification design avoids this failure mode.
  • Each of the three modifications independently raises average F1 relative to the Great baseline (from 0.6833 to roughly 0.74 alone), and combining all three yields the best result, so the gains are complementary.
  • Performance improves smoothly as the interpolation ratio grows from 0 to 1 times |D_major|, suggesting the interpolation step controls a trade-off between fidelity and diversity.
  • Larger imbalance ratios (more minority samples) shrink the gap between ImbLLM and the baselines, implying the method's advantage is largest in the most extreme data-scarce regimes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the entropy argument for condition_yx is the operative mechanism, the same 'inject random context tokens into the prompt' trick could be applied to any autoregressive class-conditional generator, not just tabular oversampling—e.g., class-conditional text or code augmentation.
  • A cleaner test would isolate prompt diversity from the other two changes by ablating only the random feature-value prefix while keeping fix_y and minor+interpolate fixed; the paper's main table bundles all three, though the ablation table partially separates them.
  • The downstream evaluation uses XGBoost only; whether the diversity gain transfers to other classifiers (logistic regression, MLPs, or other LLM-based predictors) is an open question the paper does not address.
  • The paper's own comparison shows ImbLLM does not beat the Original method—it approaches it (0.8002 vs 0.8238)—so the claim is better stated as closing most of the gap, not matching full-data training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces ImbLLM, an LLM-based oversampling method for imbalanced tabular classification. The method makes three changes relative to prior LLM oversamplers: (i) at generation time, prompts condition on the minority label plus one randomly sampled feature value ('condition_yx'); (ii) during fine-tuning, only features are permuted while the label is fixed at the start ('fix_y'); and (iii) the LLM is fine-tuned on minority samples plus synthetic interpolated samples rather than on the full imbalanced table. The authors evaluate downstream XGBoost F1/AUC on ten tabular datasets against eight baselines, report quality/diversity metrics, and present three entropy-based propositions intended to prove that each component increases diversity. They claim ImbLLM significantly outperforms all baselines.

Significance. If the empirical claims survive scrutiny, ImbLLM is a simple and potentially useful contribution: it is an LLM-based oversampler that avoids categorical encoding losses and produces diverse minority samples. The evaluation is broad (10 datasets, 8 baselines, 3 seeds) and includes an ablation of the three components. However, the paper's central empirical claim is weakened by model selection on the test sets and the absence of significance testing, and the theoretical analysis is not rigorous. Most importantly, the method description is internally inconsistent concerning the treatment of categorical features in interpolation. These issues are repairable in a revision, but they are load-bearing for the current claims.

major comments (5)
  1. [§III-A3, Eq. (5), Fig. 3] The definition of interpolated samples is inconsistent with the generation target. Eq. (5) constructs x'_i by interpolating only continuous variables, and the text states that x'_i ∈ D_inter has only M_con features. Yet |D_inter| = |D_major| while D_minor is only 20% of the original minority set, so D_train = D_minor ∪ D_inter consists mostly of rows with no categorical features. At inference, condition_yx prompts contain the label plus one feature, and the model is expected to autoregressively complete all M features. The paper does not explain how the model learns to generate categorical features from a fine-tuning set that is dominated by incomplete rows, nor how XGBoost is trained on synthetic samples if categorical columns are absent. Please clarify: either the implementation retains categorical features in interpolation (contradicting the text), or the pipeline is incomplete. This
  2. [§IV-B, Eqs. (9)–(12)] Proposition 2's proof rests on the claim that the fixed recent context C acts as an 'attentional anchor', making Q_theta_mid(G|S,C) nearly independent of the prefix S and hence low-entropy. This is asserted without derivation or empirical support. The fine-tuning objective for D_mid is over full contexts S⊕C; at inference only C is provided. The marginal Q_theta_mid(G|C) = ∫ Q_theta_mid(G|S,C)p(S)dS can be high-entropy even if each component is low-entropy, because the mixture may spread mass. The step from Eqs. (10)–(11) to Eq. (12) is therefore unjustified. The claimed inequality H(Q_begin)>H(Q_mid) does not follow. This is the central theoretical argument for fix_y.
  3. [§V-C1, Table III] The ablation study selects the component combination that maximizes average F1 on the same test sets used for the final comparison. The final row (0.8002) is the maximum over eight configurations, so the reported result is a selected maximum, not the performance of a pre-specified method. No validation set or nested selection is used. Moreover, no significance tests are reported; in Table II several per-dataset differences between ImbLLM and a baseline are within one standard deviation (e.g., adult: 0.6543±0.044 vs. Great 0.6592±0.040; housing: 0.8750±0.011 vs. CTGAN 0.8690±0.011). The abstract's claim of 'significantly outperforms' is not supported by the statistical evidence.
  4. [§IV-A, Eq. (6)] Proposition 1's proof assumes in Eq. (6) that appending a random suffix R_i to the fixed condition C flattens the next-token distribution: H(p(G|C⊕R_i)) ≥ H(p(G|C)). This is not generally true; a specific suffix can reduce entropy (e.g., a suffix that makes the next token nearly deterministic). The Jensen step in Eq. (7) yields H(¯p) ≥ (1/n)ΣH(p_i), but the desired comparison to H(p(G|C)) requires the individual entropy increase in Eq. (6), which is not established. The proof needs tightening or the proposition should be reframed as an intuition.
  5. [§IV-C] The proof of Proposition 3 applies the entropy power inequality to empirical distributions defined as sums of Dirac deltas. Differential entropy is -∞ for such distributions, so the inequality H(p_aug)>H(p_minor) is not meaningful in the standard sense. In addition, the interpolated samples omit categorical features, so the 'augmented manifold' is only a subspace of the feature space. The theoretical claim that fine-tuning on interpolated samples increases model entropy is not rigorously established.
minor comments (5)
  1. [§III-A3 heading] 'Fine-tunning' should be 'Fine-tuning'.
  2. [§II and §III-A3] Notation: D_train is used for the imbalanced classification set in §II and again for the LLM fine-tuning set in §III-A3; use different symbols (e.g., D_ft) to avoid ambiguity.
  3. [Figure 2] The two panels appear to show the same matrix layout in the extracted text; please verify the illustration actually distinguishes permute_xy and fix_y.
  4. [Table II] Add significance markers or a statistical comparison (e.g., paired test across datasets) to support the claimed improvement.
  5. [§V-A3 and reproducibility] The paper does not state whether ImbLLM's code will be released; for reproducibility, please include code or a detailed pseudo-code of the generation procedure, especially how categorical features are produced.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity found: ImbLLM's empirical claim is benchmarked on held-out test sets against external baselines; self-citations are not load-bearing, and the entropy proofs' issues are unsupported assumptions rather than reductions.

full rationale

The central empirical claim—higher F1/AUC on ten tabular datasets—is evaluated by an external pipeline: train on an imbalanced train split, generate synthetic minority samples, retrain XGBoost, and score on a held-out test set. No fitted parameter encodes the test score, and the method is compared with published implementations of eight baselines. The self-citations that exist (Pred-LLM as a baseline, and the GPT-2 hyperparameter convention following [21]) are not load-bearing: they do not justify ImbLLM's superiority; the comparison is reproduced empirically. The ablation study (Table III) selects components on the same evaluation benchmark, which is a selection-bias concern rather than circular reasoning. The theoretical section is weak: Proposition 1's Eq. (6) assumes random feature conditions flatten the predictive distribution; Proposition 2's 'attentional anchor' claim (Eqs. 9-12) is asserted rather than derived, and Jensen's inequality is used in a way that does not imply low mixture entropy; Proposition 3 treats interpolation as smoothing. These are unsupported or mathematically questionable steps, but they are not circular reductions: the conclusions are not identical to the premises by definition. The internal inconsistency in Section III-A3—interpolated samples contain only continuous features while generation must output all M features—is a correctness/pipeline issue, not a circularity. Overall, no circular step is present; the score reflects only minor non-load-bearing self-citation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. The main assumptions are concentrated in the entropy-based theoretical analysis, where several unproved statements are needed for the claimed diversity guarantees. The method also relies on hand-set hyperparameters that are not tuned per dataset but affect the generated sample diversity.

free parameters (6)
  • temperature T = 0.7
    Sampling temperature in Eq. (2); chosen following prior works, controls token diversity; not fitted but hand-set.
  • interpolation ratio r = 1.0 (|D_inter| = |D_major|)
    Number of interpolated samples added to the fine-tuning set; Figure 5 shows performance changes with r, and r=1 is the selected setting.
  • fine-tuning epochs = 50
    Number of epochs for distilled GPT-2; chosen, not tuned per dataset.
  • batch_size = 32
    LLM fine-tuning batch size; chosen, not tuned.
  • imbalance ratio q = 0.2
    Fraction of original minority samples retained in D_train; evaluation setup, affects all methods.
  • quality-metric thresholds = alpha=0.2, k=2
    Threshold for Close probability and neighbor count for Coverage; chosen by hand.
assumptions (5)
  • ad hoc to paper A random suffix R_i flattens the LLM's next-token distribution: H(p(G|C⊕R_i)) ≥ H(p(G|C)) (Eq. 6).
    Stated without proof in Proposition 1 Step 2; central to showing condition_yx increases diversity.
  • ad hoc to paper The fixed context C acts as an 'attentional anchor', making Q_theta_mid(G|S,C) nearly independent of prefix S, so the marginal over S is low-entropy.
    Proposition 2 Step 3; not derived from attention mechanics; load-bearing for fix_y's claimed entropy advantage.
  • ad hoc to paper Adding interpolated points increases entropy of the empirical distribution and forces the optimized LLM distribution to become smoother with higher entropy.
    Proposition 3; applied to Dirac empirical distributions and to the LLM family without justification.
  • domain assumption Linear interpolation of continuous features between two minority samples produces valid minority-distribution samples (Eq. 5).
    Needed for D_inter; categorical features are excluded, so interpolated rows contain only continuous features plus label.
  • domain assumption Fine-tuning via MLE minimizes KL divergence to the empirical distribution, and the optimized model faithfully reflects target entropy ordering.
    Used in Propositions 2 and 3; assumes optimization reaches the optimum and the model family is flexible enough; unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models for Imbalanced Classification: Diversity makes the difference." pith.science (2026). https://pith.science/paper/SXYCEOSM

@misc{pith2026251009783,
  author       = {Pith},
  title        = {Pith review of: Large Language Models for Imbalanced Classification: Diversity makes the difference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SXYCEOSM}},
  note         = {Machine review of arXiv:2510.09783}
}
read the original abstract

Oversampling is one of the most widely used approaches for addressing imbalanced classification. The core idea is to generate additional minority samples to rebalance the dataset. Most existing methods, such as SMOTE, require converting categorical variables into numerical vectors, which often leads to information loss. Recently, large language model (LLM)-based methods have been introduced to overcome this limitation. However, current LLM-based approaches typically generate minority samples with limited diversity, reducing robustness and generalizability in downstream classification tasks. To address this gap, we propose a novel LLM-based oversampling method designed to enhance diversity. First, we introduce a sampling strategy that conditions synthetic sample generation on both minority labels and features. Second, we develop a new permutation strategy for fine-tuning pre-trained LLMs. Third, we fine-tune the LLM not only on minority samples but also on interpolated samples to further enrich variability. Extensive experiments on 10 tabular datasets demonstrate that our method significantly outperforms eight SOTA baselines. The generated synthetic samples are both realistic and diverse. Moreover, we provide theoretical analysis through an entropy-based perspective, proving that our method encourages diversity in the generated samples.

Figures

Figures reproduced from arXiv: 2510.09783 by the authors.

Figure 1
Figure 1. Training and evaluation of an oversampling method. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Overview of our ImbLLM. First step: it converts each row to a sentence (a), permutes X and fixes Y at the beginning (b), and fine-tunes an LLM with minority samples Dminor and interpolated samples Dinter (c). Second step: it constructs prompts based on both minority label and features to query the fine-tuned LLM to generate minority samples xˆ. Step 3: Increasing total process entropy We consider the entire set of g… view at source ↗
Figure 4
Figure 4. The average quality and diversity scores. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: F1-score vs. interpolation-ratio over 10 datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: F1-score vs. imbalance-ratio over 10 datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: DCR distributions for the dataset gem_price. While other methods show that their synthetic minority samples are far away from the unseen minority samples (their mode is around 1), our ImbLLM shows that our synthetic minority samples are close to the unseen minority sam…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 2 linked inside Pith

  1. [1]

    Generating synthetic data in finance: opportunities, challenges and pitfalls

    Samuel Assefa, Danial Dervovic, Mahmoud Mahfouz, Robert Tillman, Prashant Reddy, and Manuela Veloso. Generating synthetic data in finance: opportunities, challenges and pitfalls. InACM International Conference on AI in Finance, 2020

    2020

  2. [2]

    Table-to-text: Describing table region with natural language

    Junwei Bao, Duyu Tang, Nan Duan, Zhao Yan, Yuanhua Lv, Ming Zhou, and Tiejun Zhao. Table-to-text: Describing table region with natural language. InAAAI, 2018

    2018

  3. [3]

    SciBERT: A Pretrained Language Model for Scientific Text

    Iz Beltagy, Kyle Lo, and Arman Cohan. SciBERT: A Pretrained Language Model for Scientific Text. InEMNLP-IJCNLP, 2019

    2019

  4. [4]

    Deep neural networks and tabular data: A survey.IEEE Transactions on Neural Networks and Learning Systems, 2022

    Vadim Borisov, Tobias Leemann, Kathrin Seßler, Johannes Haug, Martin Pawelczyk, and Gjergji Kasneci. Deep neural networks and tabular data: A survey.IEEE Transactions on Neural Networks and Learning Systems, 2022

    2022

  5. [5]

    Language models are realistic tabular data generators

    Vadim Borisov, Kathrin Seßler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. InICLR, 2023

    2023

  6. [6]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. InNeurIPS, 2019

    2019

  7. [7]

    SMOTE: synthetic minority over-sampling technique.Journal of Artifi- cial Intelligence Research, 2002

    Nitesh Chawla, Kevin Bowyer, Lawrence Hall, and Philip Kegelmeyer. SMOTE: synthetic minority over-sampling technique.Journal of Artifi- cial Intelligence Research, 2002

    2002

  8. [8]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InKDD, 2016

    2016

Show all 33 references
  1. [9]

    TabFact : A Large- scale Dataset for Table-based Fact Verification

    Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and Yang Wang. TabFact : A Large- scale Dataset for Table-based Fact Verification. InICLR, 2020

    2020

  2. [10]

    Class-balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. InCVPR, 2019

    2019

  3. [11]

    Generative Adversarial Nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. InNeurIPS, 2014

    2014

  4. [12]

    Why do tree- based models still outperform deep learning on tabular data?arXiv preprint arXiv:2207.08815, 2022

    Léo Grinsztajn, Edouard Oyallon, and Gaël Varoquaux. Why do tree- based models still outperform deep learning on tabular data?arXiv preprint arXiv:2207.08815, 2022

    2022 arXiv

  5. [13]

    Borderline-SMOTE: a new over-sampling method in imbalanced data sets learning

    Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-SMOTE: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, 2005

    2005

  6. [14]

    ADASYN: Adaptive synthetic sampling approach for imbalanced learning

    Haibo He, Yang Bai, Edwardo Garcia, and Shutao Li. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In IJCNN, 2008

    2008

  7. [15]

    Learning from imbalanced data.IEEE Transactions on knowledge and data engineering, 2009

    Haibo He and Edwardo Garcia. Learning from imbalanced data.IEEE Transactions on knowledge and data engineering, 2009

    2009

  8. [16]

    Tabllm: Few-shot classification of tabular data with large language models

    Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. Tabllm: Few-shot classification of tabular data with large language models. InAISTAT, 2023

    2023

  9. [17]

    TaPas: Weakly Supervised Table Parsing via Pre-training

    Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Müller, Francesco Piccinno, and Julian Eisenschlos. TaPas: Weakly Supervised Table Parsing via Pre-training. InACL, 2020

    2020

  10. [18]

    Oct-GAN: Neural ODE-based conditional tabular GANs

    Jayoung Kim, Jinsung Jeon, Jaehoon Lee, Jihyeon Hyeong, and Noseong Park. Oct-GAN: Neural ODE-based conditional tabular GANs. In Proceedings of the Web Conference, 2021

    2021

  11. [19]

    An introduction to variational autoencoders.Foundations and Trends in Machine Learning, 2019

    Diederik Kingma, Max Welling, et al. An introduction to variational autoencoders.Foundations and Trends in Machine Learning, 2019

    2019

  12. [20]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. InICML, 2020

    2020

  13. [21]

    Generating realistic tabular data with large language models

    Dang Nguyen, Sunil Gupta, Kien Do, Thin Nguyen, and Svetha Venkatesh. Generating realistic tabular data with large language models. InICDM, 2024

    2024

  14. [22]

    Fairness Improvement for Black-box Classifiers with Gaus- sian Process.Information Sciences, 2021

    Dang Nguyen, Sunil Gupta, Santu Rana, Alistair Shilton, and Svetha Venkatesh. Fairness Improvement for Black-box Classifiers with Gaus- sian Process.Information Sciences, 2021

    2021

  15. [23]

    Data Synthesis Based on Generative Adversarial Networks.Proceedings of the VLDB Endowment, 2018

    Noseong Park, Mahmoud Mohammadi, Kshitij Gorde, Sushil Jajodia, Hongkyu Park, and Youngmin Kim. Data Synthesis Based on Generative Adversarial Networks.Proceedings of the VLDB Endowment, 2018

    2018

  16. [24]

    Synthcity: facilitating innovative use cases of synthetic data in different data modalities.arXiv preprint arXiv:2301.07573, 2023

    Zhaozhi Qian, Bogdan-Constantin Cebere, and Mihaela van der Schaar. Synthcity: facilitating innovative use cases of synthetic data in different data modalities.arXiv preprint arXiv:2301.07573, 2023

    2023 arXiv

  17. [25]

    A novel SMOTE-based re- sampling technique trough noise detection and the boosting procedure

    Fatih Sa ˘glam and Mehmet Ali Cengiz. A novel SMOTE-based re- sampling technique trough noise detection and the boosting procedure. Expert Systems with Applications, 2022

    2022

  18. [26]

    Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in ultra low-data regimes

    Nabeel Seedat, Nicolas Huynh, Boris van Breugel, and Mihaela van der Schaar. Curated LLM: Synergy of LLMs and Data Curation for tabular augmentation in ultra low-data regimes. InICML, 2024

    2024

  19. [27]

    Tabular data: Deep learning is not all you need.Information Fusion, 2022

    Ravid Shwartz-Ziv and Amitai Armon. Tabular data: Deep learning is not all you need.Information Fusion, 2022

    2022

  20. [28]

    Some inequalities satisfied by the quantities of information of Fisher and Shannon.Information and Control, 1959

    Aart Stam. Some inequalities satisfied by the quantities of information of Fisher and Shannon.Information and Control, 1959

    1959

  21. [29]

    Table meets LLM: Can large language models understand structured ta- ble data? a benchmark and empirical study

    Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. Table meets LLM: Can large language models understand structured ta- ble data? a benchmark and empirical study. InInternational Conference on Web Search and Data Mining, 2024

    2024

  22. [30]

    RPT: relational pre-trained trans- former is almost all you need towards democratizing data preparation

    Nan Tang, Ju Fan, Fangyi Li, Jianhong Tu, Xiaoyong Du, Guoliang Li, Sam Madden, and Mourad Ouzzani. RPT: relational pre-trained trans- former is almost all you need towards democratizing data preparation. VLDB Endowment, 2021

    2021

  23. [31]

    Modeling tabular data using Conditional GAN

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veera- machaneni. Modeling tabular data using Conditional GAN. InNeurIPS, 2019

    2019

  24. [32]

    Language-interfaced tabular oversampling via progressive imputation and self-authentication

    June Yong Yang, Geondo Park, Joowon Kim, Hyeongwon Jang, and Eunho Yang. Language-interfaced tabular oversampling via progressive imputation and self-authentication. InICLR, 2024

    2024

  25. [33]

    Generative table pre-training empowers models for tabular prediction

    Tianping Zhang, Shaowen Wang, Shuicheng Yan, Jian Li, and Qian Liu. Generative table pre-training empowers models for tabular prediction. In EMNLP, 2023

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.