Towards High Supervised Learning Utility Training Data Generation: Data Pruning and Column Reordering

Chi-Hua Wang; Guang Cheng; Tung Sum Thomas Kwok; Zeyong Zhang

REVIEW 4 major objections 5 minor 81 references

Towards High Supervised Learning Utility Training Data Generation: Data Pruning and Column Reordering

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PRRO pipeline lifts synthetic-data training utility by 26.74% on average

desk verdict The column-reordering idea is worth a look, but the evaluation as reported does not support the headline claims. read the letter →

arxiv 2507.10088 v1 pith:3IWZVSWK submitted 2025-07-14 cs.LG stat.ML

Tung Sum Thomas Kwok , Zeyong Zhang , Chi-Hua Wang , Guang Cheng This is my paper

classification cs.LGstat.ML

keywords synthetictabulardatasupervisedlearningutilitypruningcolumnreorderingclassimbalanceLLMgenerationPRROpipelineimbalancedclassification

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 establish that the low supervised-learning utility of synthetic tabular data can be largely repaired before generation, by pruning rows that carry little signal for the class of interest and by reordering columns so the predictor sits last. The proposed PRRO pipeline combines these two preprocessing modules to make a tabular generator model the same predictor-given-features relationship that supervised learning models assume. If the claim is correct, synthetic data can serve as a much stronger replacement for, or supplement to, original training data, especially on imbalanced classification tasks. Across 22 datasets, the authors report average utility gains of 26.74% in the replacement setting and 6.13% in the appendant setting, with a 43% improvement in class-distribution similarity on six highly imbalanced datasets.

What carries the argument

The load-bearing mechanism is the PRRO pipeline itself, a two-stage pre-synthesis transformation. The first stage, Signal-based Data Pruning, treats the Spearman rank correlation between a majority-class row and any minority-class row as a signal-to-noise measurement and keeps only majority rows exceeding a fixed threshold of 0.3. The second stage, Column Conditional ReOrdering, standardizes the predictor column to the last position so that an LLM-based generator, which models columns sequentially, learns the conditional distribution of the predictor given the features. Together these modules reshape the training table before generation and are inverted after synthesis, and they are what the paper credits for both balanced-dataset utility gains and imbalanced-dataset class-distribution improvements.

What would settle it

Run the same experiments with the pruning module replaced by random selection of the same number of majority-class rows; if random selection matches PRRO's utility gains on the imbalanced datasets, then the signal-selection mechanism is not what drives the reported improvement.

Watch

Extended reading notes

Core claim

The central claim is that PRRO, a pre-synthesis pipeline, improves the supervised learning utility of synthetic tabular data by addressing two generator failures: exaggeration of class imbalance and neglect of the predictor-feature conditional relationship. The Signal-based Data Pruning module keeps minority-class observations and only those majority-class observations whose Spearman correlation with some minority observation exceeds 0.3, increasing signal density for the class of interest. The Column Conditional ReOrdering module moves the label column to the last position, aligning the generator's sequential column-by-column modeling with the supervised learning assumption that the predictor is conditional on the features. In experiments using an LLM-based tabular generator, replacing original training data with PRRO-processed synthetic data improves precision by 20.96%, recall by 24.23%, F1 by 26.45%, and AUC by 6.92% on average, with the replacement scenario improving by an average of 26.74% and up to 871.46%; the appendant scenario improves by an average of 6.13% and up to 200.32%. On six highly imbalanced CTR subdatasets, pruning reduces the generator's discount of minority-class positive rate from 51% to 8%, a 43% improvement in class-distribution similarity.

Load-bearing premise

The pruning step assumes that a Spearman correlation above 0.3 between an individual majority-class row and any minority-class row, computed across feature values, means that row carries useful signal for the minority class.

Editorial extensions

If this is right

If PRRO is correct, synthetic replacement data can approach the predictive performance of original data much more closely, making privacy-preserving or data-sharing scenarios more viable.
On imbalanced datasets, appending PRRO-processed synthetic minority data increases recall by 3.82% with only a 1.04% precision drop, yielding a net F1 gain over training on original data alone.
The column reordering module alone should generalize to balanced datasets because it aligns any sequential tabular generator with the SL conditional assumption, explaining the consistent improvements across 22 datasets.
The pruning module reduces the generator's class-imbalance exaggeration substantially, cutting the synthetic positive-rate discount from 51% to 8% on highly imbalanced CTR data.
The fixed Spearman threshold of 0.3 suggests a tunable hyperparameter: datasets with different sparsity or feature types may need different thresholds, and per-dataset selection could increase gains further.

Reading between the lines

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

A natural extension the paper leaves implicit is to test whether the pruning gains come from signal selection or simply from rebalancing class proportions; comparing PRRO pruning against random majority-class undersampling of the same size would disentangle these effects.
The reordering logic implies an ordering of features by predictive importance could further guide sequential generators, and the paper sketches this as future work using permutation importance and multicollinearity-aware placement.
Because the Spearman correlation is computed across rows with mixed categorical and continuous features, its validity as a signal measure is the most fragile link; replacing it with a learned similarity or class-conditional distance could make the pipeline more robust.
The reported gains are tied to one generator family, so a testable extension is to run PRRO with non-LLM generators such as diffusion or GAN-based synthesizers to see whether the utility improvements persist across generator architectures.

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.

Desk Editor's Note

read the letter

You should know this paper proposes a preprocessing pipeline (PRRO) for LLM-based tabular data synthesis: prune majority-class rows that correlate with minority rows, then put the label column last. The column-reordering idea is sensible and testable; the pruning idea is essentially a variant of undersampling. The formal utility definitions are clean.

What the paper does well: experiments span 22 datasets and multiple classifiers, and the authors provide an anonymous code link. They also honestly disclose that several baseline synthetic sets had 0% positive rate and that they arbitrarily flipped one label to make evaluation possible. That transparency is good, but it doesn't save the evaluation.

Now the soft spots, and they are load-bearing. The class-distribution similarity claim (43% improvement) is computed against two different reference distributions. In Table 1, "Original" is 1.95% positive in the no-pruning condition but 11.29% in the pruning condition. So the 8% discount after pruning is relative to a pruned dataset that is itself the treatment, not the original data. Against the true original 1.95%, the PRRO synthetic positive rate of 10.38% is off by a factor of five, not improved. The 26.74% and 871% utility improvements also rest on baseline datasets that included the arbitrary label flips, and no absolute metrics or per-dataset breakdowns are given. The Spearman correlation between observations across feature values is a nonstandard similarity measure, and the tau=0.3 threshold is not justified; with mixed categorical and continuous features, this needs much more support.

These are not minor issues because the central claim is empirical. The paper's own Table 1 and caption show the problem explicitly. That said, the column-reordering hypothesis — LLM generators have position bias, so put the label last — is plausible and cheap to test properly. The paper deserves a serious referee who can ask for re-analysis with honest baselines, absolute numbers, and error bars.

If I were editor, I'd send it to review rather than desk reject: the idea is falsifiable and the authors have shared code. But I'd expect heavy revision or rejection after review.

Referee Report

4 major / 5 minor

Summary. The paper proposes PRRO, a pre-synthesis pipeline for tabular data that combines signal-based data pruning and column conditional reordering, with the aim of improving the supervised learning utility of synthetic data. The claimed contributions are (i) a formal definition of synthetic data utility, (ii) a pruning module that retains non-target-class observations correlated with the target class to mitigate class-imbalance exaggeration by generators, and (iii) a reordering module that places the predictor column last to align the generator's sequential modeling with the predictor-given-features assumption of supervised learning models. Empirical results over 22 datasets are reported for both replacement and appendant scenarios, with headline claims of a 26.74% average improvement (up to 871.46%) in replacement utility and a 43% improvement in class-distribution similarity for imbalanced datasets. The paper also contains a future-work section proposing an importance-based feature ordering extension.

Significance. The problem addressed is practically important: synthetic tabular data often underperforms original data in downstream supervised learning, and the two failure modes identified (class-imbalance exaggeration and mismatched conditional modeling assumptions) are plausible. The paper provides a reproducible-code link and experiments across a range of public datasets, which are useful assets. If the reported effects were validated, a simple pre-processing pipeline that consistently improves synthetic-data utility would be valuable to practitioners. However, the current experimental evidence contains load-bearing flaws: the headline similarity improvement is computed against two different reference distributions, and several baseline synthetic datasets with zero positive labels were arbitrarily relabeled before evaluation. These issues invalidate the paper's central empirical claims as reported, so the significance of the contribution is not yet established.

major comments (4)

[§4.2.3, Table 1] The claimed 43% similarity improvement is computed as the difference between a 51% discount (from 1.95% to 0.95%) and an 8% discount (from 11.29% to 10.38%) relative to two different 'original' distributions. Against the actual original data, PRRO synthesis achieves a positive rate of 10.38% versus the true original rate of 1.95%, an overshoot of roughly 432%. The paper's own Table 1 therefore contradicts the claim that pruning makes the synthetic class distribution resemble the original data more closely; the comparison must be made against a fixed reference distribution.
[Table 1 caption, §4.2] The Table 1 caption states that several baseline synthetic datasets with a true 0% positive rate were made usable by arbitrarily choosing one random point to be positive. Any precision, recall, F1, or AUC computed on such relabeled baselines is not a valid measure of the baseline generator's utility, and a single flipped row can swing these metrics substantially in small subdatasets (e.g., CDR5 has only a few hundred synthetic rows). Because these baselines feed into the averages behind the headline 26.74% and 'up to 871.46%' improvements, the reported utility gains cannot be interpreted as measuring PRRO's effect.
[§3.1, Algorithm 1, §4.1] The signal-based pruning module rests on the assumption that the Spearman correlation between two observation vectors, computed across their feature values, indicates whether a majority-class point carries signal for the minority class. No justification is given for this nonstandard use, no handling of mixed continuous/categorical features is described, and the only threshold mentioned is the hand-picked tau=0.3. Since the pruning decision and consequently the imbalanced-dataset gains depend entirely on this choice, the method needs either a defense of the correlation-as-signal assumption or a sensitivity analysis over tau; without this, the imbalanced-dataset results are not supported.
[§4.2.1, §4.2.2, Figure 2] The reported comparisons compare a pipeline that applies both pruning and reordering against a baseline without either, so the individual contribution of the Column Conditional ReOrdering module is not isolated in the main results; attributing the generalized improvements to reordering (Sec. 4.2.1) requires an ablation that the paper does not provide. Similarly, the paper reports no paired statistical tests or confidence intervals, so it is unclear whether the averaged improvements are significant or driven by a few datasets.

minor comments (5)

[§3.1, Algorithm 1] The notation around D′, ̸D′, and the expression 'D∗ ← D∗ ∪ d' in Algorithm 1 is not properly defined; in particular, the variable d on the right-hand side is not introduced, and the correlation function corr is used without a formal definition.
[§4.1] The text states that Spearman correlation is used because it handles ordinal and continuous data, but it does not describe how categorical columns are rank-encoded before computing the correlation; this matters for the validity of the pruning step on the mixed-type datasets used in the experiments.
[§4.2.2] There are several typos in this section, including 'Colunn Conditional ReOrdering', 'Contradictorily', and the phrase 'the addition of synthetic data does not improve overall data quality in fact'; these should be corrected.
[Figure 2] The Figure 2 caption refers to 'bold numbers' and a 'more than 10% improvement for illustrative purpose', but the figure itself is not included in the manuscript text, and the axes, color coding, and what exactly is being compared are not described.
[References] Reference [1] is listed as 'Anonymous ... under review'; citing an anonymous under-review submission is not verifiable and should be replaced with a published or preprint version, or removed.

Circularity Check

1 steps flagged · score 6.0 of 10

The '43% similarity improvement' headline is computed against the pruning module's own output as the 'original' reference, so that claim reduces by construction; the main utility numbers are empirical and not circular.

self definitional [Sec. 4.2.3 and Table 1]
"After pruning, with the positive rate improving to 11.29%, a significantly smaller 8% discount rate from 11.29% to 10.38% can be observed. This accounts for a 43% improvement in similarity of class distribution between original and synthetic data. ... Pruning does not only improve the positive rate for subsequent data analytics ease, but also allows the generator to generate minority class at a rate more similar to the original data."

The 'original' reference used in the with-pruning condition is the pruned dataset D*, which is the output of the pruning module itself, not the original data D. Because the pruning module retains all positive observations and only a subset of negatives, the pruned positive rate (11.29%) is definitionally higher than the true original rate (1.95%). The '8% discount' compares the synthetic rate (10.38%) to this pruned reference, while the '51% discount' compares the un-pruned synthetic rate (0.95%) to the true original reference. The claimed 43% improvement is thus the difference between two discount rates computed against two different baselines; the improvement is manufactured by changing the reference distribution to the pruning module's own output.

full rationale

The paper's central utility claims, such as the 26.74% average and 871.46% maximum improvement from PRRO, are empirical comparisons against external datasets and are not defined in terms of the paper's own outputs or fitted parameters. The Spearman-correlation pruning threshold tau=0.3 is hand-picked with reference to external statistical conventions and is not fitted to the headline numbers. The column reordering comparison is an empirical A/B test of predictor position. The self-citations to the authors' prior work on utility theory (Ref. [69]) and on LLM tabular generation (Ref. [31]) are used for motivation and background, not as load-bearing derivations of PRRO's reported gains. The one load-bearing circular step is the 43% 'similarity improvement' claim in Sec. 4.2.3: the improvement is computed relative to the pruning module's own output as the 'original' distribution, so the favorable discount rate is built into the changed reference rather than measured against the actual original data. Additionally, Table 1's caption discloses that baseline synthetic datasets with a 0% positive rate had 'one random point arbitrarily chosen to be positive' to enable evaluation; this is a serious validity concern for the headline utility comparisons, but it is a baseline-manipulation issue rather than a circular-reasoning issue. Apart from the constructed class-similarity metric, the derivation chain is not circular, so the score reflects partial circularity confined to that headline claim.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on one hand-picked threshold and two domain assumptions. The pruning threshold tau=0.3 is a free parameter. The no-multicollinearity assumption and the observation-level Spearman correlation assumption are both load-bearing. No new theoretical entities are introduced.

free parameters (1)

Spearman correlation threshold tau = 0.3
Hand-chosen threshold in Algo 1 for retaining majority-class points; no sensitivity analysis or principled selection.

assumptions (3)

domain assumption Features are assumed to have negligibly low multicollinearity, so the conditional distribution of a feature given previous features equals its marginal distribution.
Invoked in Sec 3.2 to justify that only the predictor-feature conditional relationship matters for column reordering.
ad hoc to paper Spearman correlation between two observations (vectors of feature values) is a valid measure of signal relevance for pruning.
Core to Algo 1; no statistical justification provided, and it is not a standard similarity measure for tabular data.
domain assumption A single 40/40/20 train/holdout/validation split is sufficient to evaluate synthetic data utility without repeated runs.
Used in Sec 4.1; no cross-validation or multiple seeds, so variance of the utility estimates is unknown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards High Supervised Learning Utility Training Data Generation: Data Pruning and Column Reordering." pith.science (2026). https://pith.science/paper/3IWZVSWK

@misc{pith2026250710088,
  author       = {Pith},
  title        = {Pith review of: Towards High Supervised Learning Utility Training Data Generation: Data Pruning and Column Reordering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IWZVSWK}},
  note         = {Machine review of arXiv:2507.10088}
}

read the original abstract

Tabular data synthesis for supervised learning ('SL') model training is gaining popularity in industries such as healthcare, finance, and retail. Despite the progress made in tabular data generators, models trained with synthetic data often underperform compared to those trained with original data. This low SL utility of synthetic data stems from class imbalance exaggeration and SL data relationship overlooked by tabular generator. To address these challenges, we draw inspirations from techniques in emerging data-centric artificial intelligence and elucidate Pruning and ReOrdering ('PRRO'), a novel pipeline that integrates data-centric techniques into tabular data synthesis. PRRO incorporates data pruning to guide the table generator towards observations with high signal-to-noise ratio, ensuring that the class distribution of synthetic data closely matches that of the original data. Besides, PRRO employs a column reordering algorithm to align the data modeling structure of generators with that of SL models. These two modules enable PRRO to optimize SL utility of synthetic data. Empirical experiments on 22 public datasets show that synthetic data generated using PRRO enhances predictive performance compared to data generated without PRRO. Specifically, synthetic replacement of original data yields an average improvement of 26.74% and up to 871.46% improvement using PRRO, while synthetic appendant to original data results with PRRO-generated data results in an average improvement of 6.13% and up to 200.32%. Furthermore, experiments on six highly imbalanced datasets show that PRRO enables the generator to produce synthetic data with a class distribution that resembles the original data more closely, achieving a similarity improvement of 43%. Through PRRO, we foster a seamless integration of data synthesis to subsequent SL prediction, promoting quality and accessible data analysis.

Figures

Figures reproduced from arXiv: 2507.10088 by the authors.

**Figure 2.** The graph shows SL utility comparison between original and synthetic data. Bold numbers indicate that the change [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Conducting the same experiments with two under [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 43 canonical work pages

[1]

Anonymous. 2024. Eliminating Position Bias of Language Models: A Mechanistic Approach. In Submitted to The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=fvkElsJOsN under review

2024
[2]

Tom Archer. 2023. Tech Translated: Synthetic data. https://www.pwc.com/gx/ en/issues/technology/synthetic-data.html

work page 2023
[3]

Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. 2023. Language Models are Realistic Tabular Data Generators. In The Eleventh International Conference on Learning Representations. https://openreview. net/forum?id=cEygmQNOeI

work page 2023
[4]

Leo Breiman. 2001. Random Forest. doi:10.1023/A:1010933404324

doi:10.1023/a:1010933404324 2001
[5]

Broadstreet. 2022. What is a Successful Click-Through Rate? https:// broadstreetads.com/successful-click-through-rate/

work page 2022
[6]

Hsin-Yu Chang, Pei-Yu Chen, Tun-Hsiang Chou, Chang-Sheng Kao, Hsuan-Yun Yu, Yen-Ting Lin, and Yun-Nung Chen. 2024. A Survey of Data Synthesis Ap- proaches. arXiv:2407.03672 [cs.LG] https://arxiv.org/abs/2407.03672

arXiv 2024
[7]

N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. 2002. SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research 16 (June 2002), 321–357. doi:10.1613/jair.953

doi:10.1613/jair.953 2002
[8]

Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16) . ACM, 785–794. doi:10.1145/ 2939672.2939785

arXiv 2016

Show all 81 references

[9]

Chi, Xuezhi Wang, and Denny Zhou

Xinyun Chen, Ryan A. Chi, Xuezhi Wang, and Denny Zhou. 2024. Premise Order Matters in Reasoning with Large Language Models. arXiv:2402.08939 [cs.AI] https://arxiv.org/abs/2402.08939

2024 arXiv
[10]

Waseem Ahmad Chishti and Shahid Mahmood Awan. 2019. Deep Neural Net- work a Step by Step Approach to Classify Credit Card Default Customer. In 2019 International Conference on Innovative Computing (ICIC) . 1–8. doi:10.1109/ ICIC48496.2019.8966723

2019
[11]

Jacob Cohen. 1988. Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates, Hillsdale, NJ. ISBN: 0-8058-0283-5

1988
[12]

DeepInvolution. 2022. CTR Prediction - 2022 DIGIX Global AI Challenge. Kaggle. DOI: https://www.kaggle.com/datasets/xiaojiu1414/digix-global-ai-challenge

2022
[13]

Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J

DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei L...

2024 arXiv
[14]

Kailaash Devi. 2023. Understanding Hold-Out Methods for Training Machine Learning Models. https://www.comet.com/site/blog/understanding-hold-out- methods-for-training-machine-learning-models/

2023
[15]

Duda, Peter E

Richard O. Duda, Peter E. Hart, and David G. Stork. 2000. Pattern Classification. (2000). Chapter on Naive Bayes Classifier, including Gaussian Naive Bayes

2000
[16]

Yotam Elor and Hadar Averbuch-Elor. 2022. To SMOTE, or not to SMOTE? arXiv:2201.08528 [cs.LG] https://arxiv.org/abs/2201.08528

2022 arXiv
[17]

Xi Fang, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Qi, Scott Nickleach, Diego Socolinsky, Srinivasan Sengamedu, and Christos Faloutsos
[18]

Candace Flatt and Ronald Jacobs. 2019. Principle Assumptions of Regression Analysis: Testing, Techniques, and Statistical Reporting of Imperfect Data Sets. Advances in Developing Human Resources 21 (11 2019), 484–502. doi:10.1177/ 1523422319869915

2019
[19]

Friedman

Jerome H. Friedman. 2001. Greedy function approximation: A gradient boosting machine. The Annals of Statistics 29, 5 (2001), 1189–1232

2001
[20]

Gallegos, Ryan A

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. Bias and Fairness in Large Language Models: A Survey. arXiv:2309.00770 [cs.CL] https://arxiv.org/abs/2309.00770

2024 arXiv
[21]

Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv:2312.10997 [cs.CL] https://arxiv.org/abs/2312.10997

2024 arXiv
[22]

Seda Guney, Irem Kilinc, Alaa Hameed, and Akhtar Jamil. 2022. Abalone Age Prediction Using Machine Learning . 329–338. doi:10.1007/978-3-031-04112-9_25

2022 doi
[23]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. arXiv:1703.04247 [cs.IR] https://arxiv.org/abs/1703.04247

2017 arXiv
[24]

Román Salmerón Gómez, Catalina García García, and Ainara Rodríguez Sánchez. 2024. Enlarging of the sample to address multicollinearity. arXiv:2407.01172 [stat.AP] https://arxiv.org/abs/2407.01172

2024 arXiv
[25]

Tawfiq Hasanin and Taghi Khoshgoftaar. 2018. The Effects of Random Under- sampling with Simulated Class Imbalance for Big Data. In2018 IEEE International Conference on Information Reuse and Integration (IRI) . 70–79. doi:10.1109/IRI.2018. 00018

2018 doi
[26]

Hassanat, Ahmad S

Ahmad B. Hassanat, Ahmad S. Tarawneh, Ghada A. Altarawneh, and Abdullah Almuhaimeed. 2022. Stop Oversampling for Class Imbalance Learning: A Critical Review. arXiv:2202.03579 [cs.LG] https://arxiv.org/abs/2202.03579

2022 arXiv
[27]

Tech Lab

iab. Tech Lab. 2024. ADMAP: Attribution Data Matching Protocol: A Data Clean Room Interoperability Protocol for Attribution Measurement. iab. Tech Lab. DOI: https://iabtechlab.com/wp-content/uploads/2024/10/ADMAP-Version-1.0- for-PUBLIC-COMMENT.pdf

2024
[28]

Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. 2023. Generating Images with Multimodal Language Models. arXiv:2305.17216 [cs.CL] https://arxiv.org/ abs/2305.17216

2023 arXiv
[29]

Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. 2022. TabDDPM: Modelling Tabular Data with Diffusion Models. https://arxiv.org/ abs/2209.15421

2022 arXiv
[30]

Rohit Kundu. 2022. Precision vs. Recall: Differences, Use Cases & Evaluation. https://www.v7labs.com/blog/precision-vs-recall-guide

2022
[31]

Tung Sum Thomas Kwok, Chi hua Wang, and Guang Cheng. 2024. DEREC- SIMPRO: unlock Language Model benefits to advance Synthesis in Data Clean Room. arXiv:2411.00879 [cs.DB] https://arxiv.org/abs/2411.00879

2024 arXiv
[32]

Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. 2015. Deep learning. Nature 521, 7553 (2015), 436–444

2015
[33]

Cheng-Te Li, Yu-Che Tsai, Chih-Yao Chen, and Jay Chiehen Liao. 2024. Graph Neural Networks for Tabular Data Learning: A Survey with Taxonomy and Directions. arXiv:2401.02143 [cs.LG] https://arxiv.org/abs/2401.02143

2024 arXiv
[34]

Zongjie Li, Chaozheng Wang, Pingchuan Ma, Daoyuan Wu, Shuai Wang, Cuiyun Gao, and Yang Liu. 2024. Split and Merge: Aligning Position Biases in LLM-based Evaluators. arXiv:2310.01432 [cs.CL] https://arxiv.org/abs/2310.01432 Conference acronym ’XX, June 03–05, 2018, Woodstock, N...

2024 arXiv
[35]

McCallum and K

A. McCallum and K. Nigam. 1998. Improving text classification by discovering word dependencies. Proceedings of the 15th International Conference on Machine Learning (ICML 1998) (1998), 359–367

1998
[36]

Meta. 2024. Introducing Llama 3.2. https://www.llama.com/

2024
[37]

Robert L. Miller. 2005. Applying Statistical Methods in the Social Sciences (2nd ed.). Prentice Hall, Upper Saddle River, NJ. ISBN: 0-13-108703-4

2005
[38]

Alhassan Mumuni and Fuseini Mumuni. 2022. Data augmentation: A compre- hensive survey of modern approaches. Array 16 (11 2022), 100258. doi:10.1016/j. array.2022.100258

2022
[39]

Dang Nguyen, Sunil Gupta, Kien Do, Thin Nguyen, and Svetha Venkatesh
[40]

Geoff Norman. 2010. Likert scales, levels of measurement and the ‘laws’ of statistics. Advances in Health Sciences Education 15, 5 (2010), 625–632. doi:10. 1007/s10459-010-9222-y

2010
[41]

arXiv:2410.21717 [cs.LG] https://arxiv.org/abs/2410.21717

Generating Realistic Tabular Data with Large Language Models. arXiv:2410.21717 [cs.LG] https://arxiv.org/abs/2410.21717

arXiv
[42]

Shyam Nuggehalli, Jifan Zhang, Lalit Jain, and Robert Nowak. 2024. DIRECT: Deep Active Learning under Imbalance and Label Noise. arXiv:2312.09196 [cs.LG] https://arxiv.org/abs/2312.09196

2024 arXiv
[43]

Northcutt, Lu Jiang, and Isaac L

Curtis G. Northcutt, Lu Jiang, and Isaac L. Chuang. 2022. Confident Learning: Estimating Uncertainty in Dataset Labels. arXiv:1911.00068 [stat.ML] https: //arxiv.org/abs/1911.00068

2022 arXiv
[44]

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fan...

2019 arXiv
[45]

Jane Ostler and Ashok Kalidas. 2024. Synthetic Data: The Real Deal? The opportunities and challenges of synthetic data for market research. https: //www.kantar.com/inspiration/ai/synthetic-data-the-real-deal

2024
[46]

Pedregosa, G

F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour- napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine ...

2011
[47]

Neha Patki, Roy Wedge, and Kalyan Veeramachaneni. 2016. The Synthetic Data Vault. In 2016 IEEE International Conference on Data Science and Advanced Analytics (DSAA). 399–410. doi:10.1109/DSAA.2016.49

2016 doi
[48]

Joanne Peng, Kuk Lee, and Gary Ingersoll. 2002. An Introduction to Logistic Regression Analysis and Reporting. Journal of Educational Research - J EDUC RES 96 (09 2002), 3–14. doi:10.1080/00220670209598786

2002 doi
[49]

Dehua Peng, Zhipeng Gui, and Huayi Wu. 2024. Interpreting the Curse of Dimensionality from Distance Concentration and Manifold Effect. arXiv:2401.00422 [cs.LG] https://arxiv.org/abs/2401.00422

2024 arXiv
[50]

Michael Platzer and Thomas Reutterer. 2021. Holdout-Based Empirical Assess- ment of Mixed-Type Synthetic Data. Frontiers in Big Data 4 (06 2021), 679939. doi:10.3389/fdata.2021.679939

2021
[51]

Pezoulas, Dimitrios I

Vasileios C. Pezoulas, Dimitrios I. Zaridis, Eugenia Mylona, Christos Androutsos, Kosmas Apostolidis, Nikolaos S. Tachos, and Dimitrios I. Fotiadis. 2024. Synthetic data generation methods in healthcare: A review on open-source tools and methods. doi:10.1016/j.csbj.2024.07.005

2024 doi
[52]

Janes, Neal Navani, and Mihaela van der Schaar

Zhaozhi Qian, Thomas Callender, Bogdan Cebere, Sam M. Janes, Neal Navani, and Mihaela van der Schaar. 2024. Synthetic data for privacy-preserving clinical risk prediction. doi:10.1038/s41598-024-72894-y

2024 doi
[53]

Zahra Pourbahman, Saeedeh Momtazi, and Alireza Bagheri. 2023. Deep neural ranking model using distributed smoothing. doi:10.1016/j.eswa.2023.119913

2023
[54]

Data Reply. 2023. How synthetic data is revolutionising Industries. https://www. reply.com/data-reply/en/how-synthetic-data-is-revolutionising-industries

2023
[55]

J. R. Quinlan. 1986. Induction of decision trees. In Machine Learning , Vol. 1. Springer, 81–106

1986
[56]

Sandeep Singh Sengar, Affan Bin Hasan, Sanjay Kumar, and Fiona Carroll
[57]

Patrick Schober, Christa Boer, and Lothar A Schwarte. 2018. Correlation Coeffi- cients: Appropriate Use and Interpretation. doi:10.1213/ANE.0000000000002864

2018 doi
[58]

Noora Shrestha. 2020. Detecting Multicollinearity in Regression Analysis. American Journal of Applied Mathematics and Statistics 8 (06 2020), 39–42. doi:10.12691/ajams-8-2-1

2020 doi
[59]

arXiv:2405.11029 [cs.LG] https://arxiv.org/abs/2405.11029

Generative Artificial Intelligence: A Systematic Review and Applications. arXiv:2405.11029 [cs.LG] https://arxiv.org/abs/2405.11029

arXiv
[60]

Jonathon Shlens. 2014. A Tutorial on Principal Component Analysis. arXiv:1404.1100 [cs.LG] https://arxiv.org/abs/1404.1100

2014 arXiv
[61]

Samuel Stocksieker, Denys Pommeret, and Arthur Charpentier. 2023. Generalized Oversampling for Learning from Imbalanced datasets and Associated Theory. arXiv:2308.02966 [stat.ML] https://arxiv.org/abs/2308.02966

2023 arXiv
[62]

Aakriti Singh, Shipra Saraswat, and Neetu Faujdar. 2017. Analyzing Titanic disaster using machine learning algorithms. In 2017 International Conference on Computing, Communication and Automation (ICCCA) . 406–411. doi:10.1109/ CCAA.2017.8229835

2017
[63]

Solatorio and Olivier Dupriez

Aivin V. Solatorio and Olivier Dupriez. 2023. REaLTabFormer: Generating Realis- tic Relational and Tabular Data using Transformers. arXiv:2302.02041 [cs.LG] https://arxiv.org/abs/2302.02041

2023 arXiv
[64]

Weixi Tong and Tianyi Zhang. 2024. CodeJudge: Evaluating Code Generation with Large Language Models. arXiv:2410.02184 [cs.LG] https://arxiv.org/abs/ 2410.02184

2024 arXiv
[65]

Aofeng Su, Aowen Wang, Chao Ye, Chen Zhou, Ga Zhang, Gang Chen, Guangcheng Zhu, Haobo Wang, Haokai Xu, Hao Chen, Haoze Li, Haoxuan Lan, Jiaming Tian, Jing Yuan, Junbo Zhao, Junlin Zhou, Kaizhe Shou, Liangyu Zha, Lin Long, Liyao Li, Pengzuo Wu, Qi Zhang, Qingyi Huang, Saisai Ya...

2024 arXiv
[66]

Yasiru Tishan. 2023. Understanding the Difference Between Supervised and Unsupervised Learning Techniques. doi:10.13140/RG.2.2.36176.48641

2023
[67]

Wei and Y

X. Wei and Y. Zhou. 2005. Complement naive bayes: A new approach to classi- fication of imbalanced data. In Proceedings of the 9th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2005) . Springer, 370–380

2005
[68]

Gil Trotino. 2024. Synthetic Data Use Cases for Every Company. https://www. k2view.com/blog/synthetic-data-use-cases/#Understanding-Synthetic-Data

2024
[69]

Artem Vysogorets, Kartik Ahuja, and Julia Kempe. 2024. DRoP: Distributionally Robust Pruning. arXiv:2404.05579 [cs.LG] https://arxiv.org/abs/2404.05579

2024 arXiv
[70]

Yilin Ye, Jianing Hao, Yihan Hou, Zhan Wang, Shishi Xiao, Yuyu Luo, and Wei Zeng. 2024. Generative AI for Visualization: State of the Art and Future Directions. arXiv:2404.18144 [cs.LG] https://arxiv.org/abs/2404.18144

2024 arXiv
[71]

Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni
[72]

Yijiong Yu, Huiqiang Jiang, Xufang Luo, Qianhui Wu, Chin-Yew Lin, Dongsheng Li, Yuqing Yang, Yongfeng Huang, and Lili Qiu. 2024. Mitigate Position Bias in Large Language Models via Scaling a Single Dimension. arXiv:2406.02536 [cs.CL] https://arxiv.org/abs/2406.02536

2024 arXiv
[73]

Shirong Xu, Will Wei Sun, and Guang Cheng. 2024. Utility Theory of Synthetic Data Generation. arXiv:2305.10015 [stat.ML] https://arxiv.org/abs/2305.10015

2024 arXiv
[74]

Tongtong Yuan, Weihong Deng, Jian Tang, Yinan Tang, and Binghui Chen
[75]

Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi, Nayeong Kim, Suha Kwak, and Tae-Hyun Oh. 2024. SYNAuG: Exploiting Synthetic Data for Data Imbalance Problems. arXiv:2308.00994 [cs.CV] https://arxiv.org/abs/2308.00994

2024 arXiv
[76]

Yan-Ping Zhang, Li-Na Zhang, and Yong-Cheng Wang. 2010. Cluster-based Majority Under-Sampling Approaches for Class Imbalance Learning. Proceedings - 2010 2nd IEEE International Conference on Information and Financial Engineering, ICIFE 2010 (09 2010). doi:10.1109/ICIFE.2010.56...

2010
[77]

Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. 2024. Real-Fake: Effective Training Data Synthesis Through Distribution Matching. arXiv:2310.10402 [cs.LG] https://arxiv.org/abs/2310.10402

2024 arXiv
[79]

arXiv:1904.02616 [cs.CV] https://arxiv.org/abs/1904.02616

Signal-to-Noise Ratio: A Robust Distance Metric for Deep Metric Learning. arXiv:1904.02616 [cs.CV] https://arxiv.org/abs/1904.02616

1904 arXiv
[80]

Bo Zhang, Ruotong Yu, Haipei Sun, Yanying Li, Jun Xu, and Hui Wang. 2020. Privacy for All: Demystify Vulnerability Disparity of Differential Privacy against Membership Inference Attack. arXiv:2001.08855 [cs.CR] https://arxiv.org/abs/ 2001.08855

2020 arXiv
[2019]

In Advances in Neural Information Processing Systems

Modeling Tabular data using Conditional GAN. In Advances in Neural Information Processing Systems
[2024]

arXiv:2402.17944 [cs.CL] https://arxiv.org/abs/ 2402.17944

Large Language Models(LLMs) on Tabular Data: Prediction, Generation, and Understanding – A Survey. arXiv:2402.17944 [cs.CL] https://arxiv.org/abs/ 2402.17944

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Anonymous. 2024. Eliminating Position Bias of Language Models: A Mechanistic Approach. In Submitted to The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=fvkElsJOsN under review

2024

[2] [2]

Tom Archer. 2023. Tech Translated: Synthetic data. https://www.pwc.com/gx/ en/issues/technology/synthetic-data.html

work page 2023

[3] [3]

Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. 2023. Language Models are Realistic Tabular Data Generators. In The Eleventh International Conference on Learning Representations. https://openreview. net/forum?id=cEygmQNOeI

work page 2023

[4] [4]

Leo Breiman. 2001. Random Forest. doi:10.1023/A:1010933404324

doi:10.1023/a:1010933404324 2001

[5] [5]

Broadstreet. 2022. What is a Successful Click-Through Rate? https:// broadstreetads.com/successful-click-through-rate/

work page 2022

[6] [6]

Hsin-Yu Chang, Pei-Yu Chen, Tun-Hsiang Chou, Chang-Sheng Kao, Hsuan-Yun Yu, Yen-Ting Lin, and Yun-Nung Chen. 2024. A Survey of Data Synthesis Ap- proaches. arXiv:2407.03672 [cs.LG] https://arxiv.org/abs/2407.03672

arXiv 2024

[7] [7]

N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. 2002. SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research 16 (June 2002), 321–357. doi:10.1613/jair.953

doi:10.1613/jair.953 2002

[8] [8]

Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16) . ACM, 785–794. doi:10.1145/ 2939672.2939785

arXiv 2016