Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Towards a framework on tabular synthetic data generation: a minimalist approach: theory, use cases, and limitations

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Sparse PCA plus XGBoost generates synthetic tabular data and controllable perturbations, an interpretable, tuning-free alternative to autoencoder-based generation and raw/quantile perturbation.

desk verdict Sparse PCA + XGBoost for tabular synthetic data is a plausible, honestly reported idea, but the robustness-testing claim is weakened by an incomplete perturbation space and the empirical support is mostly visual. read the letter →

arxiv 2411.10982 v2 pith:Z3NB4S2H submitted 2024-11-17 cs.LG stat.MEstat.ML

Yueyang Shen , Agus Sudjianto , Arun Prakash R , Anwesha Bhattacharyya , Maorong Rao , Yaqun Wang , Joel Vaughan , Nengfeng Zhou This is my paper · ORCID
classification cs.LGstat.MEstat.ML
keywords tabularsyntheticdatasparsePCAXGBoostmodelrobustnesstestinglatent-spaceperturbationinterpretablegenerativevariationalautoencodercomparisonpopulationstabilityindex
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 argues that a deliberately simple pipeline, an unsupervised sparse-PCA encoder followed by an XGBoost decoder, is enough to generate synthetic tabular data and controllable perturbations for model robustness testing. The authors compare this minimalist approach with variational autoencoders on toy and credit-style data, showing it can match or beat them on interpretability, ease of use, and robustness of sampling, while avoiding VAE-style tuning. The central practical claim is that perturbing the PCA latent codes and decoding through XGBoost yields a model-based perturbation protocol that sits between raw and quantile perturbation, and that this protocol is useful for downstream robustness analysis. If true, the result matters because it offers a tractable, transparent path to synthetic data and stress-testing for high-dimensional heterogeneous tables.

What carries the argument

The central object is the sparse-PCA-plus-XGBoost encoder-decoder pipeline, used in two modes: synthetic generation (sample latents, decode) and controllable robustness testing (add $\epsilon$ noise to latents, decode, and add back residuals). The pipeline's load-bearing components are (i) sparse PCA as a linear, interpretable encoder whose loadings show which raw variables drive each latent; (ii) a per-column XGBoost decoder that handles mixed continuous, binary, and categorical data; and (iii) the residual add-back formula $X_{\mathrm{syn}} = X + \hat{X}_\epsilon - \hat{X}$, which separates the perturbation effect from the model's reconstruction error. The clustering or log-transform step is what lets the linear encoder cope with nonlinear structure, and the paper's failure-mode analysis shows that the pipeline stands or falls on whether the retained principal directions carry the information needed for reconstruction.

What would settle it

Take the paper's compactified-circle or cuboid example and run the pipeline with latent dimension equal to the number of retained components; if the synthetic data fail a two-sample distributional test (e.g., a Wasserstein or maximum-mean-discrepancy test) against the original data while an autoencoder passes it, then the central claim is bounded to near-linear datasets. More directly, on any real tabular dataset, compute the reconstruction error with $k$ and $k+1$ PCA components: a large drop when adding one component is evidence that the fixed latent dimension is discarding structure the decoder needs.

Watch

Extended reading notes

Core claim

The discovery that the authors are trying to establish is that a sparse-PCA encoder together with an XGBoost decoder forms a complete, interpretable pipeline for tabular synthetic data generation and controllable robustness testing. The encoder projects each cluster's data onto a sparse set of principal components, after an optional clustering or log-transform step to absorb nonlinearity, and the decoder learns a separate XGBoost tree for each original column to recover the raw space from the latent codes. Synthetic samples are drawn by sampling the latent distribution and decoding; for robustness testing, Gaussian noise of size $\epsilon$ is added to the latents and the decoded output is adjusted by adding back the reconstruction residual, giving $X_{\mathrm{syn}} = X + \hat{X}_\epsilon - \hat{X}$. The paper demonstrates on low-dimensional toys and an 11-dimensional simulated credit dataset that this protocol produces PSI and correlation patterns comparable to the raw and quantile baselines, while being fully interpretable and requiring no additional hyperparameter tuning beyond the chosen latent dimension and tree settings.

Load-bearing premise

The pipeline assumes the tabular data lie close enough to a low-dimensional linear subspace (after optional clustering or log transforms) that sparse PCA retains the information XGBoost needs; if a discarded singular direction carries essential structure, as in the paper's circle and cuboid examples, reconstruction collapses.

Editorial extensions

If this is right

  • If the central claim is right, a sparse-PCA/XGBoost pipeline gives practitioners an interpretable baseline for synthetic tabular data that does not require the hyperparameter search typical of VAEs.
  • Model-based latent perturbation can serve as a middle ground between raw and quantile perturbation: it preserves row-level correspondence and produces perturbation sizes that adapt to the model's learned density.
  • The residual add-back step means reconstruction error is not conflated with perturbation, so robustness tests can isolate the effect of latent changes.
  • For variables that are highly skewed or heavy at zero, the model-based approach naturally produces smaller perturbations in low-density regions, mimicking quantile perturbation without explicit quantile transforms.
  • The documented failure modes (compactified circle, cuboid surfaces) provide a diagnostic: if a dataset is isotropic or symmetric in a way that discards a singular direction, the pipeline should not be used without adding nonlinear encoding or more components.

Reading between the lines

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

  • Beyond the paper's case studies, the same residual-add-back protocol could be tested on any encoder-decoder pair (including autoencoders) to see whether latent-space perturbation plus residual correction generalizes the model-based robustness testing idea.
  • The paper's failure modes suggest a simple pre-check for practitioners: compare reconstruction quality when the latent dimension is $k$ versus $k+1$; if the incremental gain is large for a specific singular direction, the linear subspace assumption is violated and the pipeline's synthetic output should be treated with caution.
  • A natural next experiment would measure downstream utility, not just marginal PSI: train the same classifier on real versus synthetic data and compare performance; this would test whether the pipeline preserves the joint relationships that marginal metrics can miss.
  • Because sparse PCA yields interpretable loadings, the method could be extended to explainability of perturbations, e.g., attributing which raw features are most affected by a given latent shift, which is promising for model governance.
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

4 major / 5 minor

Summary. The paper proposes a minimalist pipeline for synthetic tabular data generation: a SparsePCA encoder (with optional clustering or log transformation) followed by an XGBoost decoder. For ordinary synthetic generation, Algorithm 1 Step 4 fits empirical CDFs to each PCA latent, samples from independent uniforms, and inverts the empirical transforms; for robustness testing, Step 4* instead adds Gaussian noise to the retained PCA latents and adds back reconstruction residuals. The authors present toy experiments (half-circle, cube, cuboid, mammoth) and a simulated 21-variable credit dataset with an 11-dimensional designed latent structure, using 10 PCA components. They report PSI values, visualizations of within- and across-variable perturbation sizes, a correlation-map SSIM comparison, and AUROC/logloss curves from robustness testing of three XGBoost classifiers. The paper claims the method is simple, interpretable, requires no extra tuning, and provides an alternative to raw and quantile perturbation for model robustness testing.

Significance. If the central claims were fully established, the paper would offer a lightweight and interpretable alternative to VAE-based synthetic tabular generation and a model-based perturbation scheme for robustness testing. The paper is transparent about several important limitations: Section II.D documents PCA failure modes on circle and cuboid data, Section III explicitly notes that marginalizing PCA latents ignores dependence beyond second order, Section VI.B calls for further validation of the latent-space raw-perturbation assumption, and Section VI.D attributes anomalous perturbation sizes to insufficient latent variables. This candor is a strength. The toy examples usefully illustrate where the method does and does not work. However, the central use-case claim—suitability as an alternative to raw and quantile perturbation—is not established because the perturbation protocol cannot exercise directions outside the retained PCA subspace, and the 'no extra tuning' claim conflicts with the manual hyperparameter choices in the experiments. The paper is better read as an exploratory case study than as a validated framework in its current form.

major comments (4)
  1. [Section VI] The robustness-testing protocol is structurally incomplete. Step 4* adds Gaussian noise only to the l=10 PCA latents and outputs X + (h(Z_epsilon) - h(Z)), so every perturbation lies in the fixed 10-dimensional span of the retained components. Since the simulated credit data have a designed latent dimension of 11 (Section VI, Figure 9-left), directions outside this span—including the 11th latent and residual variation—are never exercised. A downstream model that is highly sensitive only to such a discarded direction will appear arbitrarily robust under this protocol, whereas raw and quantile perturbations act in the original variable space and do move those directions. The paper itself acknowledges this risk indirectly: Section VI.D attributes the anomalous Balance-variable perturbation sizes to 'insufficient latent variables,' and Section VI.B states that the raw-perturbation assumption 'requires further testing and validation.' This is a scope limitation rather than an internal contradiction, but it is load-bearing because the central use-case claim is precisely that the method provides an alternative to raw and quantile perturbation for robustness testing. A concrete demonstration that the protocol can miss sensitivity to a discarded direction, or a modified protocol that perturbs the full latent-plus-residual space, is needed before the claim can stand.
  2. [Sections I and VI] The claim that the method 'does not require extra tuning' is contradicted by the paper's own experimental setup. Section VI fixes the PCA latent dimension to 10, XGBoost depth to 2, boosting rounds to 200, bin size to 40, and applies manual log transformations to skewed variables; Section V uses up to 500 boosting rounds and depths 1–9 in the toy examples. Algorithm 1 also takes cluster size c and latent dimension l as inputs. No sensitivity analysis, cross-validation, or data-driven selection is provided for any of these choices, so the 'no extra tuning' claim is unsupported. At minimum, the claim should be restricted to 'uses standard hyperparameters' or, better, supplemented with a sensitivity analysis over l, depth, and boosting rounds in the credit-data example.
  3. [Section V.B] The PSI-based evaluation is partly circular. Step 4 transforms each latent variable to a uniform distribution using its empirical CDF and then samples from the same empirical transform; consequently, each univariate marginal of the synthetic latents matches the original empirical marginal by construction, so univariate PSI values near zero do not validate the joint distribution. The paper itself acknowledges in Section III that 'marginalization limits expression of uncorrelated but dependent information.' The PSI results should be supplemented with joint or multivariate measures—for example, covariance or correlation distance, SSIM on correlation maps as in Figure 16, or Wasserstein distance in the original data space—and the claims about distributional fidelity should be correspondingly qualified.
  4. [Section VI.E] The robustness-testing comparison is reported as mean AUROC over 10 replicates without error bars, confidence intervals, or significance tests. The claim that quantile and model-based perturbations 'can better distinguish between the two models' is based on visual inspection of unquantified differences; with only 10 replicates, the apparent ordering could be noise. Since the practical recommendation (prefer xgb1 over xgb2) rests on these curves, the comparison needs at least standard errors or a paired test across replicates.
minor comments (5)
  1. [Abstract and Section I] There is a grammatical typo: 'minimalistic unsupervised a SparsePCA encoder' should read 'a minimalistic unsupervised SparsePCA encoder.'
  2. [Algorithm 1, Step 4*] The notation conflates clusters and rows: the fitted map h_i is applied to latent Z_i, but the residual is defined as X_i - h_i(Z_i), where X_i is also used for the target of a single row. Please clarify whether the index i denotes a cluster or a row and define the dimensions of all quantities.
  3. [Figure 2 caption] The caption contains a garbled expression: 'X + ˆXϵ − ˆXϵ=0' and a trailing '2' appear to be artifacts; the intended formula seems to be X + (X_epsilon - X_hat_0), but it should be restated cleanly with proper subscripts.
  4. [Section VI.B] The sentence 'The model-based perturbation did not significantly alter the perturbation pattern across different regions' uses 'significantly' without a statistical test; rephrase to 'did not appreciably alter' or add a quantitative comparison.
  5. [References] Reference [30] is cited as related concurrent work on multistage VAEs, but the text does not explain how that work relates to or differs from the proposed method; one sentence of discussion would help the reader.

Circularity Check

1 steps flagged · score 4.0 of 10

The reported univariate PSI evaluation is circular by construction because Step 4 samples from the same empirical CDFs used as the comparison target; the robustness-testing claim itself is benchmarked against raw and quantile baselines and is not circular.

  1. self definitional [Algorithm 1, Step 4 (Section I.A); quantitative evaluation in Section V.B; limitation acknowledged in Section III.]
    "Step 4: Synthetic data generation(Ex post density estimation): For each cluster i obtain empirical transformations gi to transform observed latents Zi to a uniform hypercube H : [0, 1]l ... (a) Sample from the uniform distribution uj ∼ (U nif[0, 1])l ... and inverse transform the observed sample back to latent space Zi = g−1 i (uuu)"

    The transformation g_i is constructed as the empirical CDF of the observed latents, so inverse-sampling from uniform produces latent values whose univariate empirical distributions coincide with the training latents by construction. The PSI metric reported in Section V.B then compares synthetic variables with the original variables after a fitted XGBoost decoder; for marginals the decoder reconstructs well, PSI is near zero regardless of whether the joint distribution is preserved. This makes the low PSI values in Figure 9 an artifact of the fitting procedure rather than independent evidence of synthetic-data quality.

full rationale

The only substantive circularity is the marginal PSI evaluation: Algorithm 1 Step 4 fits an empirical CDF to each latent and samples from it, so univariate PSI close to zero is guaranteed up to decoder error, and the paper explicitly acknowledges that this ignores dependence between PCA components. The robustness-testing use case is not circular: it compares model-based perturbations against raw and quantile perturbations on AUC curves, so the alternative-perturbation claim is externally benchmarked. The paper's documented coverage limitation—perturbing only the 10 retained PCA latents while the designed latent dimension is 11—is a scope restriction, not a circular reduction; the paper even attributes the Balance-variable anomaly to 'insufficient latent variables' and says the latent-space raw-perturbation assumption 'requires further testing and validation.' Self-citations (e.g., refs. 10, 27, 31) are background only and not load-bearing. Overall score reflects the by-construction PSI result while recognizing that the central robustness claim has independent empirical content.

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

The central claims depend on several free hyperparameters and domain assumptions; the manifold/linear-subspace assumption and the empirical marginal sampling are explicitly acknowledged by the authors as limitations. No new physical or conceptual entities are introduced.

free parameters (6)
  • latent dimension l = 10 for credit data (underlying designed latent dimension is 11)
    User choice in Algorithm 1; the paper notes increasing PC dimension can improve performance for high-dimensional data.
  • XGBoost boosting rounds = 200 for credit data, 500 for toy examples
    Chosen by authors; Section VI.B says depth 2 with 200 rounds to limit three-way interaction and overfitting.
  • XGBoost tree depth = depth 2 for credit data, up to 9 for toy examples
    Selected to control interaction order; no tuning procedure described.
  • perturbation size epsilon = 0.001 to 1 explored; recommended 0.01 to 0.1
    Chosen based on observed behavior in Figure 15; not derived from first principles.
  • bin size for continuous variables = 40
    Used in XGBoost generation; no justification given.
  • log transformation for skewed variables = applied to selected variables
    Used to make PCA operate on a more linear scale; exact transformation details are not fully specified.
assumptions (5)
  • domain assumption Manifold hypothesis: real tabular data can be embedded in a low-dimensional latent manifold that sparse PCA can capture.
    Invoked in Section II.A; the method's encoder projects onto a PCA subspace, and the paper acknowledges failure on isotropic and symmetric data.
  • domain assumption PCA variance-maximization captures the structure needed for downstream XGBoost reconstruction; discarded components are unnecessary.
    Section II.B notes information loss when distinct points project to the same latent point; the method relies on this not happening for the data.
  • ad hoc to paper Empirical marginal CDF transformation and independent uniform sampling in latent space is an adequate approximation of the joint latent distribution.
    Step 4 of Algorithm 1; Section III admits PCA components are uncorrelated but not independent, so marginalization limits dependence modeling.
  • domain assumption XGBoost can learn a reliable nonlinear reconstruction map from the PCA latent space to the original data for heterogeneous tabular types.
    Used throughout Algorithm 1; underpins the decoder and the residual-preserving perturbation scheme.
  • ad hoc to paper In robustness testing, the reconstruction residual X - h(Z) is treated as fixed instance-level noise that can be added back after latent perturbation.
    Step 4* defines X_syn = X + X_eps_hat - X_hat; assumes residual is independent of the perturbation and preserves instance characteristics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards a framework on tabular synthetic data generation: a minimalist approach: theory, use cases, and limitations." pith.science (2026). https://pith.science/paper/Z3NB4S2H

@misc{pith2026241110982,
  author       = {Pith},
  title        = {Pith review of: Towards a framework on tabular synthetic data generation: a minimalist approach: theory, use cases, and limitations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3NB4S2H}},
  note         = {Machine review of arXiv:2411.10982}
}
read the original abstract

We propose and study a minimalist approach towards synthetic tabular data generation. The model consists of a minimalistic unsupervised SparsePCA encoder (with contingent clustering step or log transformation to handle nonlinearity) and XGboost decoder which is SOTA for structured data regression and classification tasks. We study and contrast the methodologies with (variational) autoencoders in several toy low dimensional scenarios to derive necessary intuitions. The framework is applied to high dimensional simulated credit scoring data which parallels real-life financial applications. We applied the method to robustness testing to demonstrate practical use cases. The case study result suggests that the method provides an alternative to raw and quantile perturbation for model robustness testing. We show that the method is simplistic, guarantees interpretability all the way through, does not require extra tuning and provide unique benefits.

Figures

Figures reproduced from arXiv: 2411.10982 by the authors.

Figure 1
Figure 1. FIG. 1. The workflow and landscape of latent representation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. The robustness analysis pipeline for computational tractability in perturbation and to facilitate robustness analysis. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Schematic for the half circle example. The red dashed [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (left). A failure mode for autoencoders: no uniform sampling from 0 to [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. The flow-based landscape of latent representation [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. The red is the ground truth and the black scatters [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Latent traversals for autoencoder (left) and XGboost (right) fixed range latent traversals. The autoencoder consists [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. (Left) Case studies to illustrate how does PCA projection work in practice. The key limitation for PCA is its [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. (Left) The data generation mechanism for the simulated credit data. (Right) Quantitative Comparison for the 21 [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. A table with tabularized model-based perturbation hyperparameters [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. The baselines selected distribution landscape for different variables across selected “Lower”, “Middle”, “Upper” group. [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12. A comparison table for different strategy corresponding to the perturbation result for the [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: FIG. 13. A comparison table for different strategy corresponding to the perturbation result for the [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]
Figure 14
Figure 14. Figure 14: FIG. 14. A comparison table for different strategy corresponding to the perturbation result for the [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]
Figure 15
Figure 15. Figure 15: FIG. 15. The y axis data are calculated from the perturbation sizes divided by the standard deviation of each feature. We [PITH_FULL_IMAGE:figures/full_fig_p011_15.png]
Figure 16
Figure 16. Figure 16: FIG. 16. From left to right: (A) Correlation maps of ground truth [PITH_FULL_IMAGE:figures/full_fig_p012_16.png]
Figure 17
Figure 17. Figure 17: FIG. 17. The robustness testing results for Raw, Quantile and model based approach for the simulated models. [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Generative Augmentation of Imbalanced Flight Records for Flight Diversion Prediction: A Multi-objective Optimisation Framework

    cs.LG 2026-04 unverdicted novelty 5.0 of 10

    Hyperparameter-optimized generative models augment scarce flight diversion records and substantially improve prediction accuracy over real data alone.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Linear disentangling A partial remedy for the non-independence in PCA latents is adding ICA on top of PCA or replace ICA for preprocessing (See VIII A). In the linear mixing case (i.e., data generated from Y = AX where X contains independent sources on the rows), if the mixture’s fourth order kurtosis spectrum is distinct and strictly ordered, then the IC...

    work page

  2. [2]

    Multiply the fitted linear mixing matrix Wica with ˆx and obtain the synthetic latent space codesWica ˆx

    work page

  3. [3]

    Sample from the disentangled empirical marginal distribution ˆx

    work page

  4. [4]

    Tishby and N

    N. Tishby and N. Zaslavsky, Deep learning and the in- formation bottleneck principle, in 2015 ieee information theory workshop (itw)(IEEE, 2015) pp. 1–5

    work page 2015

  5. [5]

    We mainly re- visit two types of exact distribution modeling method: autoregressive models and flow-based methods[2]

    The nonlinear case: modeling complex distribution In the nonlinear case, we have to resort to density es- timation schemes where functional approximation of the density is achieved by neural networks. We mainly re- visit two types of exact distribution modeling method: autoregressive models and flow-based methods[2]. Flow-based method: The backbone of flo...

    work page

  6. [6]

    ap- proximate

    The half circle in 2D An illustration for the half-circle simulated setup is previously shown schematically (Figure 3). We can per- form latent traversals on the compressed one latent di- mension as 1D real space is explicitly ordered. We present the latent traversal comparison between autoen- coders and our approach (See Figure 7). We observe that rule-b...

    work page

  7. [7]

    As expected, the 3D point clouds are linearly projected on a plane 7 FIG

    The 3D mammoth and the Cube We graphically illustrate our approach using two 3D point cloud objects (Cube and Mammoth). As expected, the 3D point clouds are linearly projected on a plane 7 FIG. 7. Latent traversals for autoencoder (left) and XGboost (right) fixed range latent traversals. The autoencoder consists of various intermediate sizes j ∈ {1, 2, 4,...

    work page

  8. [8]

    The XGboost generator has a depth 2 with 200 boosting rounds to limit three-way variable interaction and over- fitting

    Visualizing the Model-based raw ˆXϵ, X − ˆX + ˆXϵ, and X − ˆX + ˆXϵ(truncated) across variables We add Gaussian noise on each latent i with standard deviation ϵ∗std(latent[:,i]) and obtain the correspond- ing synthetic data X + ˆXϵ − ˆX for robustness testing. The XGboost generator has a depth 2 with 200 boosting rounds to limit three-way variable interac...

    work page

Show all 51 references
  1. [9]

    Lower”, “Middle

    Baselines of the corresponding variables We first visualize the corresponding ground truth value landscape (Figure 11). We focus on the result of the perturbation for Mortgage x1(Figure 12), Credit Inquiry 6 month x13(Figure 13), Amount Past Due x3(Figure 14). We bin the resul...

  2. [10]

    Borisov, T

    V. Borisov, T. Leemann, K. Seßler, J. Haug, M. Pawel- czyk, and G. Kasneci, Deep neural networks and tabular data: A survey, IEEE Transactions on Neural Networks and Learning Systems (2022)

    2022

  3. [11]

    J. M. Tomczak, Why deep generative modeling?, in Deep Generative Modeling (Springer, 2021) pp. 1–12

    2021

  4. [12]

    Achille and S

    A. Achille and S. Soatto, Emergence of invariance and disentanglement in deep representations, The Journal of 15 Machine Learning Research 19, 1947 (2018)

    2018

  5. [13]

    Van der Maaten and G

    L. Van der Maaten and G. Hinton, Visualizing data using t-sne., Journal of machine learning research 9 (2008)

    2008

  6. [14]

    J. B. Kruskal, Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis, Psychometrika 29, 1 (1964)

    1964

  7. [15]

    J. W. Sammon, A nonlinear mapping for data struc- ture analysis, IEEE Transactions on computers 100, 401 (1969)

    1969

  8. [16]

    J. B. Tenenbaum, V. d. Silva, and J. C. Langford, A global geometric framework for nonlinear dimensionality reduction, science 290, 2319 (2000)

    2000

  9. [17]

    Belkin and P

    M. Belkin and P. Niyogi, Laplacian eigenmaps for di- mensionality reduction and data representation, Neural computation 15, 1373 (2003)

    2003

  10. [18]

    Silva and J

    V. Silva and J. Tenenbaum, Global versus local meth- ods in nonlinear dimensionality reduction, Advances in neural information processing systems 15 (2002)

    2002

  11. [19]

    2: Interpretable Minimal Compression: F or eachcluster, ap- ply sparsePCA to learn a linear subspace

    in the paper writing process Algorithm 1Proposed Algorithm Framework Input: Input tabular data matrix XXX ∈ Rn×p, Cluster size c, Latent dimension l Output: Output synthetic data ˜X˜X˜X, XsynXsynXsyn ∈ Rn×p 1: Interpretable Nonlinearity handling: Apply clustering to kill isotr...

  12. [20]

    Huang, Y

    H. Huang, Y. Wang, C. Rudin, and E. P. Browne, To- wards a comprehensive evaluation of dimension reduction methods for transcriptomic data visualization, Commu- nications biology 5, 719 (2022)

    2022

  13. [21]

    Ghojogh, A

    B. Ghojogh, A. Ghodsi, F. Karray, and M. Crowley, Multidimensional scaling, sammon mapping, and isomap: Tutorial and survey, arXiv preprint arXiv:2009.08136 (2020)

    2020 arXiv

  14. [22]

    Fefferman, S

    C. Fefferman, S. Mitter, and H. Narayanan, Testing the manifold hypothesis, Journal of the American Mathemat- ical Society 29, 983 (2016)

    2016

  15. [23]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville, Umap: Uniform manifold approximation and projection for dimension re- duction, arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  16. [24]

    D. P. Kingma and M. Welling, Auto-encoding variational bayes, arXiv preprint arXiv:1312.6114 (2013)

    2013 arXiv

  17. [25]

    S. Wold, K. Esbensen, and P. Geladi, Principal compo- nent analysis, Chemometrics and intelligent laboratory systems 2, 37 (1987)

    1987

  18. [26]

    d’Aspremont, L

    A. d’Aspremont, L. Ghaoui, M. Jordan, and G. Lanck- riet, A direct formulation for sparse pca using semidefi- nite programming, Advances in neural information pro- cessing systems 17 (2004)

    2004

  19. [27]

    Chen and C

    T. Chen and C. Guestrin, Xgboost: A scalable tree boost- ing system, in Proceedings of the 22nd acm sigkdd in- ternational conference on knowledge discovery and data mining (2016) pp. 785–794

    2016

  20. [28]

    Gorishniy, I

    Y. Gorishniy, I. Rubachev, N. Kartashev, D. Shlen- skii, A. Kotelnikov, and A. Babenko, Tabr: Unlocking the power of retrieval-augmented tabular deep learning, arXiv preprint arXiv:2307.14338 (2023)

    2023 arXiv

  21. [29]

    Rezende and S

    D. Rezende and S. Mohamed, Variational inference with normalizing flows, in International conference on ma- chine learning (PMLR, 2015) pp. 1530–1538

    2015

  22. [30]

    Kamthe, S

    S. Kamthe, S. Assefa, and M. Deisenroth, Copula flows for synthetic data generation, arXiv preprint arXiv:2101.00598 (2021)

    2021 arXiv

  23. [31]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Ka- plan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Advances in neural information processing systems 33, 1877 (2020)

    2020

  24. [32]

    Kaplan, S

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, Scaling laws for neural language models, arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  25. [33]

    Papamakarios, T

    G. Papamakarios, T. Pavlakou, and I. Murray, Masked autoregressive flow for density estimation, Advances in neural information processing systems 30 (2017)

    2017

  26. [34]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber, Long short-term mem- ory, Neural computation 9, 1735 (1997)

    1997

  27. [35]

    Kalchbrenner, E

    N. Kalchbrenner, E. Grefenstette, and P. Blunsom, A convolutional neural network for modelling sentences, arXiv preprint arXiv:1404.2188 (2014)

    2014 arXiv

  28. [36]

    Marino, N

    S. Marino, N. Zhou, Y. Zhao, L. Wang, Q. Wu, and I. D. Dinov, Hdda: Datasifter: statistical obfuscation of electronic health records and other sensitive datasets, Journal of statistical computation and simulation89, 249 (2019)

    2019

  29. [37]

    Germain, K

    M. Germain, K. Gregor, I. Murray, and H. Larochelle, Made: Masked autoencoder for distribution estimation, in International conference on machine learning(PMLR,

  30. [38]

    Ghosh, M

    P. Ghosh, M. S. Sajjadi, A. Vergari, M. Black, and B. Sch¨ olkopf, From variational to deterministic autoen- coders, arXiv preprint arXiv:1903.12436 (2019)

    2019 arXiv

  31. [39]

    Srivastava, Y

    A. Srivastava, Y. Bansal, Y. Ding, C. Hurwitz, K. Xu, B. Egger, P. Sattigeri, J. Tenenbaum, D. D. Cox, and D. Gutfreund, Improving the reconstruction of disentan- gled representation learners via multi-stage modelling, arXiv preprint arXiv:2010.13187 (2020)

    2020 arXiv

  32. [40]

    Y. Wang, H. Huang, C. Rudin, and Y. Shaposhnik, Un- derstanding how dimension reduction tools work: an em- pirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization, The Journal of Machine Learning Research 22, 9129 (2021)

    2021

  33. [41]

    Ramdas, N

    A. Ramdas, N. Garc ´ ıa Trillos, and M. Cuturi, On wasser- stein two-sample testing and related families of nonpara- metric tests, Entropy 19, 47 (2017)

    2017

  34. [42]

    Chhachhi and F

    S. Chhachhi and F. Teng, On the 1-wasserstein dis- tance between location-scale distributions and the effect of differential privacy, arXiv preprint arXiv:2304.14869 (2023)

    2023 arXiv

  35. [43]

    V. M. Panaretos and Y. Zemel, Statistical aspects of wasserstein distances, Annual review of statistics and its application 6, 405 (2019)

    2019

  36. [44]

    Dowson and B

    D. Dowson and B. Landau, The fr´ echet distance between multivariate normal distributions, Journal of multivari- ate analysis 12, 450 (1982)

    1982

  37. [45]

    Barratt and R

    S. Barratt and R. Sharma, A note on the inception score, arXiv preprint arXiv:1801.01973 (2018)

    2018 arXiv

  38. [46]

    L. Cai, H. Gao, and S. Ji, Multi-stage variational auto- encoders for coarse-to-fine image generation, in Proceed- ings of the 2019 SIAM International Conference on Data Mining (SIAM, 2019) pp. 630–638

    2019

  39. [47]

    Higgins, L

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, beta-vae: Learning basic visual concepts with a constrained varia- tional framework, in International conference on learning representations (2016)

    2016

  40. [48]

    R. T. Chen, X. Li, R. B. Grosse, and D. K. Duvenaud, Isolating sources of disentanglement in variational au- toencoders, Advances in neural information processing systems 31 (2018)

    2018

  41. [49]

    J. M. Tomczak, Deep generative modeling, in Deep Gen- erative Modeling (Springer, 2022) pp. 63–66. 16

    2022

  42. [50]

    M. E. Tipping and C. M. Bishop, Probabilistic princi- pal component analysis, Journal of the Royal Statistical Society Series B: Statistical Methodology 61, 611 (1999)

    1999

  43. [51]

    G.-I. Yu, S. Amizadeh, S. Kim, A. Pagnoni, C. Zhang, B.-G. Chun, M. Weimer, and M. Interlandi, Windtunnel: towards differentiable ml pipelines beyond a single model, Proceedings of the VLDB Endowment 15, 11 (2021)

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.