acceptodds
Under review as a conference paper at ICLR 2027

Back to the roots: revisiting tree-based autoregressive models for tabular synthesis

Abstract

We challenge the hegemony of deep generative models, especially diffusion models, for mixed-type tabular data generation. We revisit a popular method for mixed-type tabular synthesis from the statistical community – tree-based autoregressive models – and target two existing issues: (1) dependence on a single autoregressive factorization, and (2) using the same data to learn tree partitions and perform sampling at inference-time. Our method (Synthpop+) forms a mixture over random-order tree-based autoregressive models and uses out-of-bag observations for honest leaf-value estimation. Across 16 datasets and 11 modern baselines, Synthpop+ scores best in terms of multivariate fidelity: it decreases the error of a classifier 2-sample test by 49% w.r.t. the best-performing baseline (TabCascade, a recent cascaded diffusion model for tabular data), while simultaneously decreasing overfitting risk (DCR Share) and membership disclosure risk. We achieve this performance gain at 4% of the training time while training on consumer-grade CPUs instead of GPUs. These results indicate that these relatively simpler classical methods still comfortably beat deep generative models when given enough attention; we challenge practitioners to reconsider when deep generative models are actually required for tabular synthesis.

Primary area:
generative models
Keywords:
Generative model, Tabular synthesis, Autoregressive model, Tabular data, Mixed-type data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TUQCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits tree-based autoregressive models for mixed-type tabular data synthesis and proposes Synthpop+, an improved version of the classical Synthpop approach. Synthpop+ ensembles models trained with randomized feature orders and uses out-of-bag observations for honest leaf-value estimation, aiming to improve fidelity while reducing overfitting and membership disclosure risk. The method is evaluated on 16 datasets against 11 recent tabular generative baselines, including tree-based, GAN, VAE, autoregressive, and diffusion-based approaches. The results show that Synthpop+ achieves strong synthetic-data fidelity and competitive utility while substantially reducing training costs compared with recent deep generative models.

Critical strengths
  • The paper addresses an important question for tabular data synthesis. It provides a timely examination of whether increasingly complex deep generative models are necessary when simpler classical approaches may offer competitive or better performance at substantially lower training cost.

  • The paper is generally well written and clearly presents both the motivation and the proposed improvements. The connections among feature-order sensitivity, tree regularization, fidelity, and memorization are explained in an accessible and logically structured manner.

  • The experiments are thorough within the considered benchmark setting. The paper compares Synthpop+ against a broad range of recent baselines and evaluates multiple dimensions, including fidelity, utility, overfitting, membership disclosure risk, and computational cost. The component-wise ablations and sensitivity analyses also provide clear evidence for the effects of the proposed design choices.

Critical weaknesses
  • Synthpop+ primarily combines established techniques within the existing Synthpop framework. The proposed method ensembles CART-based autoregressive models over randomized feature orders, trains each model on a data subsample, and uses out-of-bag observations for honest leaf-value estimation. Although this combination produces substantial empirical gains, the paper could more clearly explain which methodological insights are specific to tabular synthesis rather than direct adaptations of established bagging and honest-estimation techniques.

  • The experimental scope may not fully support the paper’s broad conclusions about the necessity of deep generative models. The evaluation covers 16 static i.i.d. tabular datasets with at most 86 features. It remains unclear whether the conclusions extend to settings that may more strongly motivate expressive deep models, such as substantially higher-dimensional tables, complex continuous distributions, or data with structural and domain-specific constraints.

  • The computational advantage is concentrated primarily in training rather than generation. Synthpop+ requires autoregressive sampling, and the reported average sampling time is higher than that of several deep generative baselines. The paper also shows that sampling time and memory usage increase with the number of features and ensemble members. Therefore, the claimed computational advantage may not hold in applications requiring repeated or large-scale synthetic-data generation.

  • The handling of baselines that fail to complete within the resource limit may affect the aggregate comparison. UnmaskingTrees and ForestFlow fail on several datasets, and the paper assigns them the worst observed metric values and the maximum two-hour training time. Although their failures are operationally relevant, this imputation strategy may amplify their disadvantage in the averaged results. Reporting aggregate comparisons separately over datasets on which all methods complete would make the comparison more transparent.

Questions for authors
  • What aspects of Synthpop+ provide methodological insights specific to tabular synthesis beyond combining random-order ensembling, subsampling, and honest estimation?

  • How well does Synthpop+ generalize to substantially higher-dimensional or structurally constrained tabular datasets?

  • How does Synthpop+ compare with deep generative models when large volumes of synthetic data must be generated repeatedly?

  • Do the main conclusions remain unchanged when the comparison is restricted to datasets on which all baseline methods complete successfully?

6 comments

Thank you for your thorough review. We will address your mentioned weaknesses and questions point-by-point below.

Critical Weaknesses

1

The applications of these existing ideas are naturally specific to tabular synthesis. Namely, bagging over random feature orders only makes sense if we're autoregresively factorizing the joint distribution. For discriminative models there is no ordering to the features.

Also, it is very typical for ML research to reuse existing ideas. The important point is whether how it is implemented is a good application considering the problem, theoretically justified, and whether it substantially improves the results. We have shown all in the paper.

2

We provide in-depth discussion in the Limitations section regarding this. We do not claim that our method is the best choice for any and all dataset; this would never be possible.

However, we can say that our benchmarking is extremely intensive compared to other works published in this venue: 16 mixed-type tabular datasets, with a big range of number of rows, features, and cardinality. Asking us to extend this already very intensive benchmark even further seems harsh. We have shown great performance of our method on an already great variety of tabular datasets.

Regarding data with structural- or domain-specific constraints: see our answer to Reviewer v8cD (Question 1). In short, such constraints can be easily enforced through pre- and postprocessing schemes in any generative model, and their ability to directly model such constraints is not as practically relevant as it may seem.

3

Yes, sampling time is a widely discussed limitation in the paper. However, note that our method still outperforms 4/11 baselines, on average. Also, consider that sampling time is measured in seconds, whereas training times are measured in minutes. The practical implications in terms of absolute runtimes are therefore smaller than they might seem.

Also, note that our method can easily be scaled up in terms of sampling time and memory by simply using more CPU cores and RAM. Since we parallelize over the random feature orders at each feature index, we can scale up roughly linearly as long as we have less cores than random orders. This is in contrast to GPU-based methods which would require much more expensive GPUs (either extra cards or more capable cards) to scale up.

Finally, consider again that we've already considered much more varied, large, and complex datasets than most similar works, and that we've shown that the runtime and memory consumption of our method are low considering the quite restrictive hardware we used.

4

If we restrict the results to datasets on which UTrees and ForestFlow complete, we can only include very small datasets. Then, the whole evaluation loses its value. The interesting part is whether our method can indeed scale to large datasets; it can!

If you do wish to evaluate performance on individual datasets, Appendix B.6 provides this. Here we see that our method greatly outperforms both baselines. Therefore your suggestion does not change the conclusions, but it would harm the information on the scaling abilities of our method.

Finally, note that this imputation scheme is typical in this field. See for example the CDTD and TabCascade paper.

Questions

1

Random-order ensembling is specific to tabular synthesis, as it is only relevant when factorizing the joint distribution, and not for, e.g., discriminative models. Also, we've shown that these methodological improvements substantially improve generalization of the generative distribution. The work should be about evaluating whether the improvements are sound, sensible for the problem at hand, not used in this manner before, and provide good results. They do all of that.

2

As mentioned above, structurally constrained datasets are not necessarily interesting from a practical perspective, as structural constraints can be easily enforced in any generative model through pre- and postprocessing schemes.

Scaling performance to higher-dimensional datasets can be directly investigated from Table 4 (dataset characteristics) and Appendix B.6 (dataset-wise results). We find that our method generalizes well to higher-dimensional datasets, also considering high-cardinality datasets.

3

We expect you mean in terms of sampling times. You can inspect the dataset characteristics (Table 4) and the dataset-wise sampling times (Table 24). This way you can find how our method compares to other baselines for the larger datasets. Notably, this already considers repeated sampling.

4

Yes, the conclusions remain the same. This can be seen directly from the dataset-wise results in Appendix B.6.

Thank you again for your thorough review. If you have further questions, please let us know.

Thank you for the detailed clarifications. While the arguments regarding the fidelity-generalization tradeoff (via test-set metrics and DCR/MIA) and scalable CPU-based parallelization are helpful, the core concern regarding the marginal improvement in training-set C2ST relative to the original Synthpop remains only partially addressed.

To help finalize this evaluation, could you briefly clarify whether a direct ablation study (or concrete metrics) was conducted to isolate the exact performance gain contributed specifically by the out-of-bag honest estimation versus random-order ensembling alone?

Thank you for your reply. Yes, the paper provides a direct comparison for:

  • solely using random order ensembling
  • gains from also including out-of-bag honest estimation

Specifically, Table 3 provides this. Note that this table provides additive model improvements, such that the model in each row also includes the improvements from the rows above.

Here you can compare the following rows:

  • +subsample data: this contains both the random order ensemble (+ensemble feature order) and the subsampling of the training data. This should be considered as the full random order ensembling, since the subsampling is used to further decorrelate the ensemble members, as is typical.
  • +honest trees: this contains the full random order ensembling, and out-of-bag honest estimation. This therefore indicates the exact performance gain contributed specifically by the out-of-bag honest estimation versus random-order ensembling alone, as you requested.

As you can see from the results, out-of-bag honest estimation substantially improves on DCR Share, and therefore reduces overfitting. Please consider for each of the metrics the scale of the original error when evaluating the scale of the improvement.

Finally, note line 408 in the paper: the various improvements we supply are ``pick and choose". If you're less worried about overfitting and membership disclosure risk you may omit out-of-bag honest estimation and/or subsampling of the training data. If you do this, our improvements actually do provide a substantial improvement on training-set C2ST (see Table 3)!

Thank you again for your reply. If you feel like we've substantially addressed some of your concerns, we humbly request to consider reflecting this in your score, as per the conference guidelines.

If you have any further questions regarding our work we're happy to help.

Reviewer_v8cDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Current dominant methods for mixed-type tabular data generation are deep learning based approaches. This paper revisits the classical tree-based methods for the tabular data generation by asking whether they can offer a competitive alternative approach. The paper extends Synthpop, a tree-based auro-regressive model for sequentially column generation, to Synthpop+ through three changes:

  1. an ensemble of models with different random column orders,
  2. data subsampling, and
  3. honest trees.

The ensemble of column order addresses the order dependencies for column generation (different column order can provide different performances). Subsampling provides each ensemble model a different random half of training data reducing the correlation between the ensemble members and also decreasing overfitting and membership disclosure risk. To generate a row, Synthpop+ randomly and uniformly selects an ensemble member and sequentially generates all column values using that member’s feature order and fitted trees, sampling each value from the corresponding leaf's out-of-bag observations.

Based on the paper's results, on average, Synthpop+ generates more faithful synthetic mixed-type tabular data than the 11 methods it was compared with across 16 datasets. Synthpop+ has a training time much less compared to deep learning methods but it is slower in sampling stage as it generates each row column by column.

Critical strengths
  1. The paper asks an important questions in todays deep learning focused world. Are deep learning based complex generators necessarily the best choice for tabular synthesis? Can we bring back traditional statistical methods with more modern ML work and beat/compete with those deep learning based approaches?

  2. To answer those questions, the paper proposed a tree-based auto-regressive approach which extends Synthpop. To address the ideal order complexity which we have in Synthpop, the random order ensembling method is employed as a trade-off to increase fidelity without a statistically significant increase in the reported overfitting metrics.

  3. The paper has a clear communication on main ideas and implementations.

  4. The benchmark models and datasets cover a broad model classes and regimes with different size, dimension and categorical cardinalities in datasets.

  5. The Synthpop+ has the lowest training cost on average over all the models.

Critical weaknesses
  1. The paper seems more or less incremental over the known ideas such as Synthpop, random-order mixtures, bagging/subsampling, and honest trees while the combination are novel in how they are used. But I am not sure that is a strong enough novelty factor with the evidence they show.

  2. Synthpop+ has a longer average sampling time than 7/11 models and is 100x longer to sample than the fastest. Thus although the proposed approaches are efficient in training, sampling costs narrow the practical advantage.

  3. There are some other classic approaches for tabular data generation such as such as copula or KDE based models which are sometimes fast in training or sampling. The papers did not include those baselines.

Questions for authors
  1. The paper's evaluations considers global distributional fidelity but does not explicitly measure preservation of exact/ approximate functional dependencies (FD). It is not clear to me if the proposed approach can provably/experimentally capture FDs between columns. For example, if two rows agree on column A, they should agree on column B, or more complex dependencies; given the values of columns A and B, one can deterministically determine the value of column C. It is important to have a comparison between most promising DNN baselines and Synthpop+ in terms of how well they can capture and maintain FDs.

  2. The proposed tree-based method stores many trees and leaf-value lookup tables, but the paper does not provide a systematic cross-baseline memory comparison.

  3. It seems that CPU and GPU based methods use different hardware, this is not a normalized compute-cost comparison.

  4. In the paper, Table 1 summarizes the results in which for baselines could not be trained for a specific dataset received the worst observed scores for each metric. I think we should have the results restricted to datasets on which both methods UTrees and ForestFlow complete.

4 comments

Thank you for your thorough review. We address your mentioned weaknesses and questions below, point-by-point.

Critical Weaknesses

1

It is very typical for ML research to reuse existing ideas. The important point is whether how it is implemented is a good application considering the problem, and whether it substantially improves the results. We have shown both in the paper.

We do not believe it is a good idea to only publish papers since the math is novel. It should be about bringing the field forward. Importantly, we show that the focus on complex deep generative models in the tabular data domain is somewhat misplaced, as we can achieve much(!) better results by simply reusing existing ideas in a smart way. We’d say this is a very important insight for the field.

2

Put differently, Synthpop+ is faster in sampling than 4/11 baselines, is fastest to train, and achieves the best-generalizing generative distributions by far! The framing seems too negative in your comment, considering the results from the paper. We do acknowledge the limitation of the sampling time, also in the Limitations section. However, please also consider that sampling times are measured in seconds, whereas training times are measured in minutes. A larger sampling time, but still in the order of seconds, will not have huge practical implications. Also, our solution is very scalable by simply throwing more CPUs against the problem; this is a cheap solution compared to deep generative models which would need more expensive GPUs.

3

Popular copula implementations for tabular data generation are known to perform very poorly, at least for somewhat complex datasets (which we consider as well). This is know both theoretically and empirically. Theoretically, these copulas are not expressive at all, as they cannot represent higher-order interactions well. Empirically, this can be seen from, e.g., the ForestDiffusion paper. As we therefore know that even our baseline models strongly outperform copulas, and our method strongly outperforms all baselines, it seems very wasteful of resources to include such a weak method. Finally, please note that we already consider a great variety of (strong and recent) baselines; much more than most similar work published in similar venues.

Questions

1

We do not include analysis of such Functional Dependencies (FDs) for the following reasons:

  • these analyses are not common in the field. That is, none of the baseline model papers -- which are all published in similar venues as this one -- include this. We need to follow the standard evaluation practice.
  • these analyses are less interesting than it seems, from a practical perspective. Exact FDs, but also inequalities, can always be 100% preserved through simple pre- and postprocessing schemes. Take colA = colB + colC. Simply remove colA, generate data, then generate and append syn_colA as syn_colA = syn_colB + syn_colC. The same argument holds for inequalities, but through modelling the difference between the arguments directly. Because of this, whether a generative model can directly preserve FDs is not really interesting. Any model can do this by simple pre- and postprocessing with 100% guarantee.

2

Everything fits comfortably within our modest hardware (see the paper), so we know that it is (far) below this cap. Also, such a cross-baseline memory comparison is not common in this field. We do already provide the empirical scaling of the memory consumption on simulated data, which seems even more interesting, and is already beyond what most works provide.

3

CPU- and GPU-based methods indeed use different hardware; one uses CPUs, and one uses GPUs. There is no widely-accepted way to normalize this. Nowhere do we claim that our results are normalized compute-cost comparisons, nor is this typical in this field. We simply selected a somewhat restricted hardware setting as to evaluate whether our method is sensible for users with access to typical hardware. This was indeed the case. We are not sure what more can be expected.

4

If we restrict the results to datasets on which UTrees and ForestFlow complete, we can only include very small datasets. Then, the whole evaluation loses its value. The interesting part is whether our method can indeed scale to large datasets; it can!

If you do wish to evaluate performance on individual datasets, Appendix B.6 provides this. Here we see that our method greatly outperforms both baselines. Therefore your suggestion does not change the conclusions, but it would harm the information on the scaling abilities of our method.

Finally, note that this imputation scheme is typical in this field. See for example the CDTD and TabCascade paper.

We thank you again for your thorough review. If you have any further questions, please let us know.

Dear reviewer v8cD,

we'd be interested to hear any remaining questions you may have regarding our work. Considering our exchange, it seems that many of your earlier concerns should be addressed. If so, we'd kindly like to remind you to consider reflecting this in your score of our work.

Thanks again for your insightful review of our work.

Thank you for the response. Some concerns still remain about the evaluation.

Regarding FDs, I suggest using them to check the validity of generated synthetic data. Ii is clear that a known simple rule such as f(A,B)=C can be satsified by calculating the missing column. However, the question here is whether the model itself can preserve the dependencies hidden in the real data (called dependency preservation, a aspect of synthetic data quality). Simple FDs provide a basic test of this ability.

To make it more clear, note that the real data may also contain more complex or overlapping dependencies and constraints, such as f(A,B)=g(D,E). These cannot always be handled by simply removing and rebuilding a column. Reconstruction may require solving a difficult equation, and some generated combinations may have no valid solution without changing other columns.

I am keeping my current score, but I may reconsider it after discussing the paper and your clarifications with the other reviewers.

Dear reviewer v8cD,

Thank you for your response.

Regarding FDs, I suggest using them to check the validity of generated synthetic data. Ii is clear that a known simple rule such as f(A,B)=C can be satsified by calculating the missing column. However, the question here is whether the model itself can preserve the dependencies hidden in the real data (called dependency preservation, a aspect of synthetic data quality). Simple FDs provide a basic test of this ability.

Our point is that, if these FDs can be easily satisfied by calculating the missing column, the question whether the model itsel can preserve this is not as interesting. Then, it is mostly interesting whether the model can accurately model the joint distribution (which we measure). Any FDs can simply be enforced through pre- and postprocessing. Again, we have to mention that such FD checks also aren't a standard evaluation measure in similar works published in this venue. See any of the papers of the baseline models we have included. Therefore we feel this is not really a fair ask, especially if it is preventing acceptance of this work.

To make it more clear, note that the real data may also contain more complex or overlapping dependencies and constraints, such as f(A,B)=g(D,E). These cannot always be handled by simply removing and rebuilding a column. Reconstruction may require solving a difficult equation, and some generated combinations may have no valid solution without changing other columns.

To be able to evaluate FDs, f() and g() must be known. If f() and g() are known, they can typically be rewritten in the form A=z(B,D,E); then, the same pre- and postprocessing trick applies. If the function is indeed more complex, such that A cannot be uniquely determined by B, D, and E, this may be more difficult. However, note that the discovery of such FDs purely from the data is not very realistic. These most likely have to stem from domain knowledge; we consider relatively simple datasets from general-purpose tabular benchmark, so such domain-specificity is far beyond the scope of this work. Again, we also are not aware of any similar works in tabular synthesis which do evaluate FDs with this degree of complexity.

In conclusion, we have to remain insistent on the fact that FDs are not typical evaluation in tabular synthesis, considering similar works published in this venue. This can directly be seen from the papers of the included baseline models. Secondly, we remain insistent that this FD evaluation is less relevant than evaluating aspects of the modelling of the joint distribution, for which we provide a very detailed evaluation.

Thank you for you comments. We are glad to answer any remaining questions!

Reviewer_uWeBOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits tree-based autoregressive models for tabular data synthesis and proposes Synthpop+, which combines random feature-order ensembling with out-of-bag estimation to improve generation fidelity and reduce overfitting. Experiments on 16 datasets demonstrate strong performance against recent deep generative models, with substantially lower training costs.

Critical strengths
  1. The paper provides a valuable empirical reassessment of classical tree-based methods for tabular synthesis, challenging the common reliance on deep generative models.

  2. The experimental evaluation is comprehensive.

Critical weaknesses
  1. The additional benefit of the proposed improvements over the original Synthpop is not fully convincing. As shown in Table 3, the original method already achieves strong fidelity, while the improvement in the primary training-set C2ST metric is relatively small and not statistically significant. It remains unclear how much practical value the proposed modifications provide beyond the existing method.

  2. The practical advantages of Synthpop+ require further qualification. The method directly resamples observed feature values, which may expose sensitive information in certain data-sharing scenarios. Moreover, its sampling time is considerably higher than that of some deep generative baselines. These limitations raise questions about its suitability for applications requiring privacy-sensitive or large-scale synthetic data generation.

Questions for authors
  1. Could the authors further clarify the practical benefits of Synthpop+ over the original Synthpop, particularly given the relatively limited improvement in C2ST?

  2. Could the authors elaborate on the intended application scenarios and how the sampling overhead and potential disclosure of original feature values affect the method's practical usefulness?

  3. Given that autoregressive sampling remains relatively time-consuming, could the authors discuss the potential for GPU-based batch generation? In particular, it would be interesting to understand whether sample-level parallelism and batched tree traversal could further improve sampling efficiency. Just for discussion. No need for more experiments.

1 comment

Thank you for your thorough review. We will address your mentioned weaknesses and questions point-by-point.

Critical Weaknesses

1

Training-set C2ST is not the only important metric. Specifically, we need to learn a generative distribution which generalizes as well as possible. This means we fit the training data without overfitting. This is called the fidelity-generalization tradeoff – or in tabular generative modelling sometimes also fidelity-privacy tradeoff.

DCR Share, MIA Lift, and test-set C2ST improve substantially in Table 3. If we can maintain very high fidelity while majorly decreasing overfitting risk, we are learning a generative distribution which generalizes much better.

It is important not to underestimate the importance of this tradeoff, and therefore the strength of our results. Namely, fitting the training data accurately is not difficult at all if we would simultaneously be allowed to overfit – i.e., we could just train a hugely overparameterized model.

2

Regarding resampling observed values: we address this directly in the Limitations section. Also, it is important to consider that our method does not yield increased overfitting or membership disclosure risk in our experiments. Notably, resampling observed values is not necessarily a privacy risk. This is only a privacy risk if that particular value is identifiable, or a particular value in combination with particular values in other features. Actually, this is what the MIA measures, and we do not find increased risk there. If you do wish to decrease this risk, there are simple solutions, e.g., smoothing sensitive features, differential-privacy, etc. However, these are method-agnostic; they could be applied to any of the included baselines as well. We are mainly concerned with learning an accurate and well-generalizing generative distribution, for which we’ve provided ample evidence in the paper.

Regarding sampling time: Synthpop outperforms 4/11 baselines, so it definitely does not perform very poorly. Also, consider that sampling time is measured in seconds; the absolute sampling times are still very low, so the practical implications of increased sampling time are limited. Of course, increased sampling time is somewhat of a limitation, and we already mention it explicitly in the Limitations section. But, there is no free lunch: our method learns much(!) better generalizing generative distributions, and trains much faster than any other baseline. The only real limitation is this sampling time.

Questions

1

As mentioned above, the main benefit is that we substantially decrease overfitting (DCR Share, MIA Lift). We also substantially increase fidelity w.r.t. the test set (C2ST test). Therefore our improvements ensure much better generalization of the generative distribution. Again, simply having high fidelity w.r.t. training data is not sufficient, we actually want to generalize. Therefore our improvements are indeed substantial and important.

2

Our method can be used in any scenario where tabular synthetic data generation is relevant. Namely, our method does this with much greater fidelity, while incurring less overfitting and membership disclosure risk than high-fidelity competitors. Regarding sampling overhead, there is a very simple solution; just consider more capable hardware, i.e., more CPU cores and RAM. However, note that we already consider relatively large datasets (compared to other work published in similar venues), and our method performs well here in absolute runtimes.

3

We would generally not recommend this. We already parallelize across random feature orders at each feature index during generation. It is unclear how this could be easily combined with sample-level parallelism and/or tree traversal within a single GPU. A much easier solution is to use more CPU cores such that each feature is generated faster. Namely, we find that the tree traversal is not the bottleneck, but rather, the parallel generation over feature orders at each feature index.

Thanks again for your review. If you have any further questions, we're glad to address them.

Comments on the paper

1 comment

Dear reviewers,

Coming up on the end of the discssusion period, we'd like to thank you again for taking the time to review our work.

Considering the exchange below, we can conclude that there are two remaining concerns, which we'd like to address one final time:

  • evaluation of Functional Dependencies (FDs). These FDs are not a typical evaluation in research on tabular generative models, for good reason: they can typically be enforced through pre- and postprocessing schemes. Very complex FDs for which this is not possible, are typically not discoverable from the data, but only from domain knowledge. Evaluating those is far beyond the scope of our work, since we consider general-purpose tabular datasets from TabArena and TabZilla. We also were not provided any references of similar work which evaluate FDs, where the goal of the work is advancing the frontier of tabular generative models, like in our work.
  • improvements w.r.t. the original Synthpop method. We'd like to re-emphasize that the original Synthpop method already yields very high training-set fidelity, but relatively poor generalization (high overfitting and privacy risk). Our improvements substantially enhance generalization while yielding similar training-set fidelity (Table 3). This yields a method with very desirable properties: high training-set fidelity ánd generalization; in fact, our method yields a considerable Pareto-improvement on recent state-of-the-art tabular generative models (Figure 1). Our improvements therefore should not be understated!

Thanks again for your time!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.