acceptodds
Withdrawn from ICLR 2027

GDIFFTABLE: GRAPH DIFFUSION FOR TABULAR DATA SYNTHESIS

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Tabular synthesis requires preserving dependencies among heterogeneous columns, yet diffusion progressively corrupts the row-level evidence for those dependencies. We introduce GDiffTable, a mixed-type tabular diffusion model that represents columns as graph nodes and jointly learns a dataset-level weighted adjacency with a time-conditioned message-passing denoiser. The adjacency is shared across diffusion timesteps to provide a stable scaffold for cross-column coupling, while the exchanged messages adapt to the current noisy state and time embedding. We further introduce correctable graph initialization: an offline schema-derived relation prior determines only the initial adjacency, and all off-diagonal weights remain trainable under the denoising objective. Consequently, initially absent relations can emerge, selected relations can be weakened, and the formulation remains compatible with non-LLM initializers. Across seven benchmarks, GDiffTable reports the highest aggregate point estimates for Shape (98.74%) , together with the smallest average downstream gap to real data (4.50%) among the evaluated synthesizers. Adult and News comparisons indicate that cross-column communication improves over node-wise denoising, learned non-uniform routing improves over fixed uniform mixing, and refinement improves over a frozen prior. The full and data-only graph variants attain similar final quality; in the reported runs, LLM initialization reduces their early-stopping times by 38.5% and 59.3%. These results support graph-structured diffusion denoising while positioning the LLM prior as an optional optimization warm start.

Primary area:
generative models
Keywords:
Tabular Data Synthesis, Diffusion Models, Graph Structure Learning, Feature Dependency Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MtBmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a tabular diffusion model that represents rows as graphs, where nodes are column attributes and (soft) edges are learned, and uses a graph neural network-based denoiser to learn the reverse diffusion process. The denoiser network and the graph edge weights are jointly trained via a regularized TabDiff-style [1] denoising loss. The paper also proposes a (optional) way to initialize the graph edge weights using an external LLM and the available schema information.

Critical strengths

The proposed methodology is a clean and useful application of graph models to tabular data generation, explicitly capturing inter-column dependencies.

Critical weaknesses
  • The advantage of the proposed method is not clear from the reported experiments, which do not match the initial motivation. Specifically, when looking at the main metrics for inter-column dependency evaluation (Trend and alpha-Precision), GDiffTable performs very similarly to the most similar diffusion baseline, TabDiff (Table 2): 97.92 vs. 97.86 on Trend and 97.99 vs. 97.76 on alpha-Precision. Missing statistical significance tests make the evaluation harder to assess.
  • Additional experiments are needed to assess the benefits of the proposed graph component. One suggestion is to test whether the proposed method can produce high-quality synthetic data using fewer denoising steps than diffusion baselines. An alternative could be to evaluate the method on more complex datasets, where the baselines could fail.
Questions for authors
  1. Can the authors elaborate on [Their reverse process, however, must recover coherent feature relationships from states in which the evidence for those relationships has itself been corrupted.] (L39-41)? How is this different than marginal distributions or diffusion models in other domains?
  2. How is Equation (1) derived?
  3. What is the "common early-stopping criterion" (L450-451)?
  4. Some hyperparameter details are missing. E.g., how many GNN layers are used? What is the total number of parameters?

References

[1] Shi, Juntong, et al. "Tabdiff: a mixed-type diffusion model for tabular data generation." International Conference on Learning Representations. Vol. 2025. 2025.

Reviewer_W4L9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces GDiffTable which is a graph-based diffusion model for mixed-type tabular data generation. It treats columns as graph nodes and try to learns relationships through a trainable adjacency matrix and a time-conditioned message-passing network. Additionally, an optional LLM is used as graph prior provides an initialization which can be refined during training.

Critical strengths

Synthetic tabular data generation is a challenging task especially when the diverse columns are present. Considering different applications like in data augmentation, privacy-sensitive data sharing, and downstream prediction having such a method is beneficial to generate training data.

Critical weaknesses
  • Even though the problem is not new, the paper does not sufficiently establish its novelty relative to existing graph-based tabular generation and diffusion methods. It combines learnable adjacency, time-conditioned message passing, and schema-derived initialization without clearly identifying the primary methodological contribution.

  • The relative importance of graph-based denoising compared to correctable graph initialization is vague.

  • The ablations focus on Adult and News, providing limited evidence that the reported benefits generalize across all seven datasets. Although the paper emphasizes heterogeneous column types, it would make paper strong if the impact of how column diversity and dependency structures affect performance considering limitation and capability of GDiffTable.

  • The reported results of the GDiffTable are close to its baseline. Having the statistical significance of the improvements would help to know if the improvement is useful or not. Also, SPADA's higher β-recall also suggests that GDiffTable does not consistently outperform existing approaches.

  • Since the adjacency matrix is figured out via denoising objective, it is unclear if it captures meaningful, stable column relationships or not. Having study around sensitivity to initialization, and comparisons with random or statistically derived graphs would help refine the contributions.

  • The paper requires improvement in writing, and proofreading. The research gap, limitations of prior methods, and rationale for the proposed approach are not clearly established.

Questions for authors

See above

Reviewer_o7TJOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce a method for learning a weighted adjacency matrix in GNN-based tabular synthetic data generators that use columns as nodes and time-conditioned message passing. They ablate the mechanism and isolate this learned adjacency matrix as a core component for improvements in both MLE and fidelity metrics compared to fixed uniform weights. They compare means across five runs (though what a “run” is is not specified), 12 total methods, and 7 datasets (4 classification, 3 regression). The authors also try an LLM warm start that is shown to reduce the amount of time it takes to reach the early stopping point.

Critical strengths
  1. The ablation, while only a point estimate and on two datasets, is interesting and provides meaningful insight into what makes GDiffTable useful. The learned adjacency is an interesting idea and appears to be a core component of the mechanism evidenced by the ablations: Trend goes from ~85% to 97% for Adult/News and AUROC from ~0.81 to ~0.91 for Adult when learning the weighted adjacency matrix A instead of using fixed uniform weights. I believe this is the strongest part of the paper and it would serve the work to highlight this more prominently.
  2. The paper is honest about how the warm start primarily serves to reduce the amount of time it takes to achieve the method’s early stopping criteria, while not necessarily improving the final generated output. It is important to frame findings in a way that does not overstate what is able to be claimed, and the authors avoided overclaiming here.
Critical weaknesses
  1. Benchmarks: The authors fail to cite and compare to prior tabular synthetic GNN diffusion denoiser work (GATD, see questions below) - half of their 1st bullet point in contributions to the literature already exists (columns as nodes with time-conditioned message passing in mixed-type diffusion). The TabDiff results reported in this paper do not reproduce the published TabDiff numbers and are weaker by about a percentage point. Missing prior/relevant work and failing to reproduce one of the core benchmark method’s results make it difficult to assess GDiffTable’s performance and novelty. This contributes heavily to how I have scored this paper.
  2. Reproducibility: The paper does not include important information about its protocol and tables appear to be incomplete - e.g. Table 5 does not include all methods and datasets (only 5 methods), Figure 3 does not specify what datasets GDiffTable was tuned on, and it is unclear what “5 runs” means in Table 1. These details are important for the reproducibility of the paper and I would rank the concerns as equally high as my concern with benchmarks.
Questions for authors
  1. The authors’ TabDiff results call into question how the authors’ benchmarks were run. The authors were unable to reproduce TabDiff’s results (published AUROC for Adult TabDiff is 0.912 while the authors reproduced 0.901. For Default it is 0.763 vs. 0.758, Magic is 0.936 vs. 0.932, News RMSE is 0.866 vs. 0.884) but reproduced TabSyn’s results (as published by TabDiff) to within 0.003 on AUROC and RMSE. Each utility score is worse than the results in TabDiff’s published paper. The authors state in section 4.1 “All experiments were conducted on a single NVIDIA RTX 6000 GPU using CUDA 12.1, Python 3.10, and PyTorch 2.2.2.” TabDiff does not use PyTorch 2.2.2. Did the authors train TabDiff’s models in the published TabDiff environment, or their own?
  2. Furthermore, TabDiff’s default hyperparameters in one entrypoint in the codebase are not the default hyperparameters from their published paper (the config file, particularly referring to learnable noise schedules and per-column hyperparameter choices, does not use the paper’s core methods, while the paper reproduction run entrypoint does). Did the authors mistakenly use the fixed schedule ablation in Shi et al.’s TabDiff paper? Why are the TabDiff results worse than reported in TabDiff’s paper?
  3. Prior work “Geometry-Aware Tabular Diffusion” (published ICML 2026, publicly available on OpenReview since April 30th, 2026, code available on GitHub) is not cited. Its GNN geometry denoiser overlaps significantly with your own (columns as nodes, time-conditioned message passing with the same EDM/masked diffusion, etc). Their GNN geometry architecture achieves similar utility scores under a different protocol (GDiffTable AUROC Magic 0.935 vs. GATD 0.934, Adult 0.913 vs. 0.911, Default 0.764 vs. 0.763, California RMSE 0.264 (16.3% error) vs 0.216 (14.3% error), News 0.857 (1.7% error) vs. 0.855 (2.5% error)). GDiffTable provides a meaningful improvement over GATD GNN Trend for Adult and News comparing an average of 97.5% to 93.5% preservation for the two datasets. GATD’s MLP reports gains over TabDiff and would be an appropriate benchmark for GDiffTable. Given the similarity and timing (GATD was public more than 4 months prior to ICLR submissions), why was GATD neither cited nor used as an additional benchmark? Please cite GATD, clarify how GDiffTable might provide something better, more interesting, or more novel than the geometric GNN published in GATD, and include GATD as a baseline under GDiffTable’s experimental protocol.
  4. Table 1 specifies “5 runs.” Is this 5 independent training runs? 5 independent generation seeds from a single trained model? 5 evaluator fits? How many independent samples were drawn and what random seeds (exactly) were used for each run? What is a run?
  5. Table 3 says “point estimate.” Is this single seed? Why is there no variance? How many runs contribute to these results?
  6. Table 5 wall clock only includes a handful of methods and datasets (5 out of 12 methods, 3 out of 7 datasets). Are you able to complete this table for all methods and datasets?
  7. Tables, presentation: Could the authors please provide tables similar to how TabSyn, TabDiff, and others in the tabular synthetic data literature have presented their results? A handful of tables with datasets as columns, methods as rows, and a far right column showing each method’s aggregate metric or distance from real data performance. This would make it much easier to interpret the results. Right now, there is only 1 table in the paper that does this (AUROC/RMSE).
  8. Figure 3 showing the relationship between hyperparameter sensitivity analysis and different metrics is interesting and provides useful insight into the mechanism. I didn’t see it specified in the manuscript - how many datasets and seeds was GDiffTable tuned on? Is Figure 3 the result of fitting GDiffTable on all of the datasets in your benchmark, or was GDiffTable tuned on a held-out set of benchmark data not included in the primary paper scores? The Figure 3 DCR scores are very high compared to Table 2’s 51.95% - again, what was GDiffTable tuned on?
  9. Could the authors please, in Table 3, put runtime in a separate column or find a different way to present it? The parentheticals that contain the runtimes are easy to miss when reading by column name.
  10. I would generally lump papers like this into two categories; a methods paper, or a results paper. A core component of what is written here and claimed to be novel already exists in the literature (GNN diffusion denoiser for tabular synthetic data has been done before), while I believe the LLM warm start and adjacency notions are not something I’ve seen applied to tabular data yet (at least not in a tabular GNN diffusion denoiser). In order for this to be a successful methods paper, greater focus should be placed on the truly novel parts of your methodology (removing claims to what has already been done), and provide a more thorough systems analysis. If this is to be a successful results paper, I would need to see convincing evidence that GDiffTable consistently outperforms other methods on at least one of the following metrics: Fidelity (Shape, Trend), MLE (classification, regression), runtime, memory consumption, or something else of importance. Given that TabDiff numbers are in question and GATD is not present, I do not see evidence to support strong results claims compared to existing methods. After clarifying TabDiff’s results and benchmarking against GATD, what compelling evidence exists that shows GDiffTable consistently outperforms other existing tabular synthetic data generation methods in some important area? Does the Shape result hold up? How about Utility?

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

Related papers

Discussion (0)

Sign in to comment.