acceptodds
Under review as a conference paper at ICLR 2027

SynFD: Functional Dependency Repair for Tabular Synthesis

Abstract

Tabular data synthesis has made substantial progress in improving distributional fidelity. However, for relational tables, which are standard in relational databases, statistically realistic samples are not necessarily structurally valid. For such tables, functional dependencies (FDs) specify deterministic relationships among attributes, which cannot be easily formulated or guaranteed by tabular generative objectives. Motivated by this observation, we propose SYNFD, a framework that repairs synthesized tabular data for FD consistency. SYNFD exploits mapping lookups for seen determinants (left-hand-side value of an FD), runs conditional generation for unseen determinants, and follows topological dependency propagation to process chained FDs. Experiments across multiple tabular datasets show that SYNFD substantially improves FD consistency, achieving an FD-consistent rate of 98.4% on average and a training-mapping agreement rate of 99.8%, while maintaining strong distributional fidelity and downstream utility.

Primary area:
generative models
Keywords:
Tabular Data Generation, Generative Models, Functional Dependencies, Data Constraints, Structure-Aware Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8QBaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a simple method to fix functional dependency (FD) violations in synthetic tabular data (an FD X --> Y between two sets of attributes X and Y means that X uniquely determines Y). The approach consists of traversing the FDs of the table in topological order and replacing the dependent value (Y) with the correct value. For cases where the determinant's value (X) has not been observed in the training data, a separately trained conditional generative model is used to sample Y conditioned on X. Unsurprisingly, this approach yields higher consistency rates than with unprocessed data.

Critical strengths

The problem the paper addresses is practically interesting, and the solution is simple and seems to work, but I found no critical strengths that would support acceptance at venues like ICLR.

Critical weaknesses

The proposed method is extremely simple and straightforward. The core idea is a post-processing algorithm that can be applied to synthetic data. While the method trains an additional conditional generative model, it does not introduce any new insights of interest to machine learning research.

Questions for authors

The following questions are clarification requests:

  1. Can the authors elaborate on [... FDs ... cannot be easily formulated or guaranteed by tabular generative objectives.] (L15-16)?
  2. In L71-72, the authors say that after repairing a tuple for one FD, the determinant of the second FD also changes, which might create an unseen determinant for the second FD. However, couldn't it also be the case that the determinant A of the first FD is unseen since it is generated by the generative model?
  3. For an FD A --> B, can one simply remove column B in a pre-processing step, train the generative model on the resulting table, then add it back based on generated values in A? How does this relate to the proposed approach?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8QBa (1/2)

We thank the reviewer for the constructive feedback. We address the concerns below and will incorporate the additional results and clarifications into the revised manuscript.

W1. The proposed method appears to be a straightforward post-processing procedure, and the additional conditional generative model provides limited methodological insight.

We respectfully disagree. As discussed in Section 5.2, the challenge is to enforce FD constraints without substantially degrading distributional fidelity and downstream utility. SynFD addresses this through theoretical analysis, DAG-guided repair, and empirical validation.

Theoretical characterization: Section 3 formalizes exact FD preservation under probabilistic generation. Proposition 1 shows that learning a deterministic relationship does not necessarily recover the reference FD mapping, highlighting the distinction between statistical dependency learning and exact FD preservation.

Algorithmic design: Interacting FDs require more than independent mapping corrections. Consider A→BA\rightarrow B and (B,D)→C(B,D)\rightarrow C. Repairing BB may introduce an unseen downstream determinant (B,D)(B,D), for which the original CC may no longer match the repaired context. SynFD addresses this through DAG-guided repair and dynamic mapping updates (Sections 4.1--4.2). It handles three cases: (i) known configurations through reference mapping lookup, (ii) native-unseen configurations by retaining jointly generated values, and (iii) repair-induced unseen configurations through conditional completion. This avoids reconstructing dependent attributes unnecessarily.

Empirical validation: SynFD improves average FD-valid retention from 54.21% to 98.40% and reference-mapping agreement from 87.39% to 99.77%.

We also compare against HFGF-style reconstruction using the same base generator. This baseline reconstructs dependent attributes through mapping lookup and uses empirical random assignment for unseen configurations. SynFD instead selectively repairs jointly generated tuples, achieving higher average Trend (0.799 vs. 0.759) and α\alpha-Precision (0.875 vs. 0.825). On Hospital, which contains 154 FDs, Trend improves from 0.421 to 0.623.

Q1. Why cannot exact FDs be easily formulated or guaranteed by conventional tabular generative objectives (L15--16)?

Section 3 shows that exact FD preservation requires the dependent attribute to agree with the reference mapping with probability one for every determinant configuration with positive generated probability. While FD constraints can be explicitly formulated, conventional generative objectives focus on approximating the joint distribution and do not guarantee this requirement.

Proposition 1 further shows that even a deterministic relationship may differ from the reference FD mapping. Thus, learning statistical dependencies alone does not guarantee exact FD preservation, as also reflected in the FD violations reported in Table 1.

This motivates SynFD's DAG-guided repair framework, which explicitly handles interacting FD constraints while retaining jointly generated values whenever possible.

AuthorsOpenReview ↗

Response to Reviewer 8QBa (2/2)

Q2. Can the determinant of the first FD already be unseen because it is generated by the base model, rather than introduced by an earlier repair?

Yes. This is the native-unseen case explicitly handled in Section 4.1 and Algorithm 2 (Lines 16--19).

SynFD distinguishes two types of unseen determinant configurations. For native-unseen configurations, the determinant is absent from the reference mapping but has not been modified by earlier repairs. SynFD retains the jointly generated dependent value and stores it in the mapping table for consistent reuse.

For repair-induced unseen configurations, an upstream repair changes the determinant, so the original dependent value may no longer match the updated context. SynFD instead uses MCGEN for conditional completion and stores the generated value for subsequent reuse.

Q3. Could FD-dependent columns simply be removed before generation and reconstructed afterward using the determinant values?

For a single FD A→BA\rightarrow B with a complete reference mapping, this is a reasonable approach. However, it faces two challenges.

First, predefined mappings cannot directly determine dependent values for unseen determinant combinations. For example, given (A,B)→C(A,B)\rightarrow C, a generator may produce an unseen (A,B)(A,B) combination even when both attributes have been individually observed.

Second, interacting FDs require coordinated reconstruction. For A→BA\rightarrow B and (B,D)→C(B,D)\rightarrow C, reconstructing BB may create an unseen (B,D)(B,D) configuration, for which mapping lookup alone cannot determine CC.

SynFD handles these cases through DAG-guided repair, dynamic mapping updates, and selective conditional completion (Sections 4.1--4.2), while retaining jointly generated values whenever possible.

We also evaluate this alternative using an HFGF-style reconstruction baseline with the same base generator (Table 4 and Appendix D.1). It reconstructs dependent attributes through mapping lookup and uses empirical sampling for unseen configurations. SynFD achieves higher average Trend (0.799 vs. 0.759) and α\alpha-Precision (0.875 vs. 0.825).

Thank you for answering my questions; that clarifies things.

However, my main concern regarding the paper's contribution still holds, and the rebuttal only summarized it. Therefore, I maintain my score.

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

This paper focuses on addressing issues with functional dependencies when tabular data is synthesized, and proposes SynFD. SynFD repairs generated tables through dependency-ordered lookups, retaining assignments for new determinants and using conditional diffusion when earlier repairs create unseen determinants. Across five datasets, the paper reports 98.4% average FD-consistent retention and 99.77% training-mapping agreement. The practical idea is sensible, but the algorithm and experiments do not yet establish the reliability or added value of conditional completion.

Critical strengths
  • The paper addresses a practical distinction that distributional evaluations can miss. Plausible individual rows do not necessarily form a consistent table. Treating functional dependencies as a table-level requirement makes the problem concrete. Reporting internal consistency separately from agreement with training mappings is also useful, since these measure different properties.
  • The repair design is economical. it leaves the base generator unchanged, preserves jointly generated values where possible, and reserves conditional completion for cases introduced by dependency propagation. This makes it a potentially useful addition to existing synthesis pipelines without requiring a new generator.
Critical weaknesses
  • The paper perhaps has to distinguish FD consistency from recovery of the reference mapping. Definition 1 permits any mapping that is consistent within the generated table, whereas Section 3 appears to require agreement with a particular reference mapping. These are different objectives. Caching one predicted value for an unseen determinant can make subsequent assignments consistent with that prediction; it does not establish that the prediction matches the unknown reference mapping. Likewise, exact preservation of a joint distribution satisfying an FD would preserve that FD almost surely; the practical difficulty lies in approximation and repair. This distinction matters because it changes what the method guarantees and how the reported consistency and mapping-agreement results should be interpreted.
  • The experiments do not yet isolate the value of the method’s distinctive component. Dependency-aware reconstruction and constraint-based repair have substantial precedents: HFGF reconstructs dependent attributes hierarchically [1], TabDiff supports conditional imputation [2], HoloClean combines constraints with statistical evidence for repair [3], and Kamino examines post-generation repair and its utility cost [4]. None is identical to SynFD. The contribution therefore appears to lie in selective completion of repair-induced unseen determinants, rather than dependency-aware generation itself. The HFGF-style ablation is useful, but Table 3’s comparison of filtered baselines with repaired SynFD tables makes it hard to separate repair quality from sample count and residual violations. The decrease in TPC-H beta-recall from 0.7967 to 0.2688, together with the higher reported mean F1 of HFGF-style reconstruction (0.489 versus 0.420), also limits a broad claim of preserved coverage and utility. A comparison including unfiltered TabDiff, common validity criteria, and matched-size results for Hospital and TPC-H would make the trade-offs much clearer.

[1] Umesh et al. Dependency-aware synthetic tabular data generation. arXiv:2507.19211, July 2025. [2] Shi et al. TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation. ICLR 2025, §2.5. [3] Rekatsinas et al. HoloClean: Holistic Data Repairs with Probabilistic Inference. PVLDB, 2017. [4] Ge et al. Kamino: Constraint-Aware Differentially Private Data Synthesis. PVLDB, 2021.

Questions for authors

Is the intended guarantee internal FD consistency, preservation of known reference mappings, or recovery of reference mappings for unseen determinants? Please distinguish these claims and state the conditions under which the procedure guarantees consistency, particularly when dependencies share a dependent attribute. Does the reported 98.4% retention include filtering after repair?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NGjV(1/3)

We thank the reviewer for the constructive feedback. We address the concerns below and will incorporate the additional results and clarifications into the revised manuscript.

Internal FD consistency and reference-mapping agreement (W1).

Definition 1 defines internal FD consistency, while Section 3 examines agreement with the reference mapping frf_r. Proposition 1 shows that conditional determinism does not guarantee reference-mapping agreement, since a learned mapping hrh_r may differ from frf_r. SynFD uses reference mapping lookup for observed determinants and consistent assignments for unseen ones, without assuming their unknown reference values. We evaluate these objectives separately: RFDR_{\mathrm{FD}} measures the proportion of synthetic rows retained by FD filtering, while RmapR_{\mathrm{map}} measures reference-mapping agreement on observed determinants (Appendix C.4). We will clarify this distinction near Definition 1.

Exact versus approximate distribution matching (W1).

Exact recovery of a joint distribution satisfying an FD would indeed preserve that dependency almost surely. Our motivation concerns approximate generative models, whose distribution-learning objectives do not guarantee exact FD preservation. As established in Proposition 1, even learning a deterministic conditional relationship does not necessarily recover the reference FD mapping. SynFD addresses this gap through explicit DAG-guided repair, combining reference mapping lookup with conditional completion. We will clarify this distinction in Section 3.

Methodological distinctions from HFGF, TabDiff, HoloClean, and Kamino (W2).

As discussed in Section 5.2, our goal is to improve FD consistency without substantially compromising distributional fidelity and downstream utility. These methods address different problems. TabDiff learns the joint distribution without explicit FD enforcement. HoloClean focuses on cleaning noisy data, while Kamino studies constraint-aware differentially private synthesis. HFGF is the closest approach, but reconstructs dependent attributes through predefined mappings. Its evaluation focuses on categorical dependencies, and it does not provide context-aware completion for unseen determinant configurations. In contrast, SynFD selectively repairs jointly generated tuples rather than reconstructing all dependent attributes. Its DAG-guided procedure handles interacting FDs through mapping lookup and conditional completion when upstream repairs change downstream determinants. Compared with HFGF-style reconstruction, SynFD achieves higher average Trend (0.799 vs. 0.759) and α-Precision (0.875 vs. 0.825). It also improves FD-valid retention from 54.21% to 98.40%, while increasing Trend from 0.760 to 0.799 relative to raw TabDiff samples.

Appendix D.2 appears to compare filtered baselines with unfiltered SynFD (W2).

This is a wording error in Appendix D.2. Both the baselines and SynFD in Table 3 and Tables 8--12 were evaluated using the same FD-filtering procedure. The reported SynFD results are therefore based on repair followed by filtering. We will clarify this in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer NGjV(2/3)

The incremental effect of FD repair is not isolated from final filtering (W2).

To isolate the contribution of DAG-guided repair, we compare four settings under the same initial generation budget: (i) TabDiff (raw), (ii) TabDiff + Filter, (iii) SynFD (repair), and (iv) SynFD + Filter, as shown in the table below. Before filtering, SynFD increases the average proportion of FD-valid tuples from 54.21% to 98.40% (1.82×) without discarding any synthetic tuples. Trend improves from 0.7595 to 0.7991 and α-Precision from 0.8658 to 0.8776. These improvements are achieved through repair alone, without relying on FD filtering. After applying the same FD-filtering procedure, SynFD outperforms TabDiff + Filter in average Shape (0.9555 vs. 0.9222), Trend (0.7987 vs. 0.6923), and α-Precision (0.8748 vs. 0.7005).

DatasetMethodFinal nFD-valid (%)Shape ↑Trend ↑α-Prec. ↑β-Rec. ↑F1 ↑R² ↑
AirbnbTabDiff (raw)2,43851.41±0.920.9663±0.00220.8983±0.00220.9614±0.01320.2320±0.0087---0.0607±0.0853
TabDiff + Filter1,253100.000.9601±0.00360.8975±0.00640.9336±0.01980.3611±0.0160---0.0155±0.0203
SynFD (repair)2,43894.71±0.000.9751±0.00000.9236±0.00000.9830±0.00000.2935±0.0000---0.0779±0.0152
SynFD + Filter2,309100.00±0.000.9748±0.00000.9237±0.00000.9814±0.00000.3056±0.0000---0.0032±0.1433
BeijingTabDiff (raw)37,58180.50±0.080.9833±0.00020.9664±0.00410.9880±0.00180.5600±0.0005--0.6451±0.0343
TabDiff + Filter30,253100.000.9830±0.00040.9679±0.00370.9450±0.00100.5099±0.0017--0.6501±0.0203
SynFD (repair)37,581100.00±0.000.9834±0.00000.9662±0.00000.9893±0.00000.6256±0.0000--0.6395±0.0205
SynFD + Filter37,581100.00±0.000.9834±0.00000.9662±0.00000.9893±0.00000.6256±0.0000--0.6395±0.0205
HospitalTabDiff (raw)3,79344.37±0.510.8801±0.01230.5390±0.01170.4975±0.00420.0020±0.00110.2612±0.0038--
TabDiff + Filter1,683100.000.8417±0.01760.4503±0.01350.7290±0.02650.1383±0.02550.2662±0.0203--
SynFD (repair)3,79399.95±0.000.8910±0.00000.6235±0.00000.4939±0.00000.5995±0.00000.2498±0.0120--
SynFD + Filter3,791100.00±0.000.8910±0.00000.6234±0.00000.4948±0.00000.5539±0.00000.2563±0.0034--
MigraineTabDiff (raw)30190.59±2.000.9804±0.00270.9165±0.01140.9027±0.02140.1743±0.02650.2887±0.0933--
TabDiff + Filter273100.000.9717±0.00130.9048±0.01520.8832±0.01120.1828±0.03200.1775±0.0420--
SynFD (repair)30197.34±0.000.9798±0.00000.9276±0.00000.9377±0.00000.1927±0.00000.2757±0.0251--
SynFD + Filter293100.00±0.000.9761±0.00000.9254±0.00000.9244±0.00000.1975±0.00000.3413±0.0461--
TPC-HTabDiff (raw)21,1414.18±0.060.9490±0.00060.4773±0.00020.9794±0.00560.1360±0.00420.6560±0.0009--
TabDiff + Filter883100.000.8547±0.00070.2411±0.00070.0116±0.00210.7967±0.01630.6610±0.0050--
SynFD (repair)21,141100.00±0.000.9524±0.00000.5547±0.00000.9841±0.00000.2688±0.00000.6618±0.0000--
SynFD + Filter21,141100.00±0.000.9524±0.00000.5547±0.00000.9841±0.00000.2688±0.00000.6618±0.0000--
AverageTabDiff (raw)13,05154.210.95180.75950.86580.22090.40200.2922
TabDiff + Filter6,869100.000.92220.69230.70050.39770.36820.3173
SynFD (repair)13,05198.400.95630.79910.87760.39600.39580.2808
SynFD + Filter13,023100.000.95550.79870.87480.39030.41980.3181
AuthorsOpenReview ↗

Response to Reviewer NGjV(3/3)

The matched FD-valid-size study omits Hospital and TPC-H (W2).

Appendix Table 14 provides matched-size FD-valid comparisons on Airbnb, Beijing, and Migraine. For Hospital and TPC-H, we additionally evaluated TabDiff+Filter under a 50× generation budget, the maximum used in the original paper. On Hospital, TabDiff+Filter retained only 2,399 of 189,650 generated rows (1.265%), below the target of 3,793. On TPC-H, it retained only 1,103 of 1,057,050 rows (0.104%), below the target of 21,141. In contrast, SynFD+Filter retained 3,791 of 3,793 and 21,141 of 21,141 rows, respectively, using only a 1× generation budget. SynFD also achieved higher Trend on Hospital (0.623 vs. 0.496) and TPC-H (0.555 vs. 0.257).

DatasetMethodGeneratedFD-validKeep (%)ShapeTrendα-Prec.β-Rec.F1
HospitalTabDiff + Filter (1×)3793168344.370.8420.4500.7290.1380.266
TabDiff (50×, Raw)189650----0.9280.6130.4970.002--
TabDiff + Filter (50×)18965023991.2650.8750.4960.8980.0780.252
SynFD + Filter (1×)3793379199.950.8910.6230.4950.5540.256
TPC-HTabDiff + Filter (1×)211418834.180.8550.2410.0120.7970.661
TabDiff (50×, Raw)1057050----0.9610.4990.9820.139--
TabDiff + Filter (50×)105705011030.1040.8450.2570.0350.7420.656
SynFD + Filter (1×)2114121141100.000.9520.5550.9840.2690.662

TPC-H shows substantially higher β-Recall for TabDiff+Filter than for SynFD (W2).

Our primary objective is to preserve FD consistency while maintaining distributional fidelity. Although TabDiff+Filter achieves higher β-Recall on TPC-H (0.797 vs. 0.269), it retains only 4.18% of generated samples, compared with 100% for SynFD. Such extensive filtering changes the composition of the retained data and may contribute to the difference in β-Recall. SynFD achieves higher Trend (0.555 vs. 0.241) and α-Precision (0.984 vs. 0.012), while also improving β-Recall over unfiltered TabDiff (0.269 vs. 0.136). Even with a 50× generation budget, TabDiff+Filter retains only 0.104% of generated samples, with lower Trend and α-Precision.

HFGF-style reconstruction has a higher average downstream F1 than SynFD (W2).

As discussed in Section 5.2, our goal is to preserve FD consistency without substantially compromising distributional fidelity. Although HFGF-style reconstruction achieves higher average F1 (0.489 vs. 0.420), SynFD achieves higher Trend (0.799 vs. 0.759) and α-Precision (0.875 vs. 0.825), as shown in Table 4. Reconstructing dependent attributes may preserve predictive information for specific downstream tasks, but can alter the statistical relationships learned through joint generation. SynFD instead selectively repairs FD violations while retaining jointly generated values whenever possible.

FD consistency guarantees and the reported 98.40% retention (Q1).

As discussed in Section 4.1, when multiple FDs impose conflicting assignments on the same dependent attribute, SynFD prioritizes the first applicable FD, so some constraints may remain unsatisfied. The reported average RFD=98.40R_{\mathrm{FD}}=98.40% measures the proportion of repaired tuples that pass FD filtering, not the FD consistency of the final filtered output. The remaining inconsistent tuples are removed, yielding 100% FD-consistent outputs. Compared with the original TabDiff samples, SynFD increases FD-valid retention from 54.21% to 98.40%, substantially reducing the number of tuples discarded by final filtering.

Reviewer_5NwEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies functional dependency (FD) preservation in tabular data synthesis: synthetic data may exhibit high statistical fidelity while violating deterministic constraints across rows. The authors propose SynFD, which first generates a complete synthetic table using a base generator and then repairs it following a topological ordering of an attribute-dependency graph. For determinant configurations with existing mappings, dependent values are retrieved directly. For unseen configurations that have not been changed by an earlier repair, generated values are retained and registered in the mapping. For repair-induced unseen configurations, a separately trained mask-conditioned diffusion model completes the dependent values, and the resulting mappings are reused.

The paper also distinguishes between satisfying deterministic dependencies within the generated relation and agreeing with a reference mapping. Across four real-world datasets and one TPC-H-derived dataset, SynFD reports an average FD filtering retention rate of 98.40% and training-mapping agreement of 99.77%. The evaluation additionally examines distributional fidelity, downstream utility, ablations, and sampling overhead.

Critical strengths

S1. The paper addresses a concrete, practically relevant problem and achieves substantial improvements in structural consistency. In Table 2, compared with the underlying TabDiff generator, average R_FD increases from 54.21% to 98.40%, and R_map increases from 87.39% to 99.77%. On TPC-H, R_FD increases from 4.18% to 100%. Here, R_FD denotes the fraction of tuples retained by the paper’s sequential FD filtering procedure. Although this metric does not capture every aspect of synthetic data quality, these results clearly demonstrate the value of explicit repair for structural constraints. This is the strongest consideration supporting acceptance.

S2. The repair framework has a sensible modular design and accounts for changes in context caused by upstream repairs. Processing dependencies topologically avoids repairing dependent attributes before their determinants are finalized. Distinguishing native unseen configurations from repair-induced unseen configurations also has a clear motivation: in the latter case, the original dependent value may no longer be appropriate for the modified context. Combining deterministic mapping reuse with conditional completion is a promising design for practical use. The authors additionally provide a conditional-completion ablation, an HFGF-style comparison, and some experiments with matched numbers of FD-valid samples, enabling further examination of the design.

Critical weaknesses

W1: 1. Algorithm 2 can produce FD conflicts even when no dependent attribute is constrained by multiple FDs.

There is a coordination gap between steps 16–19 and steps 27–31 of Algorithm 2 on page 14. A repair-induced unseen configuration is first added to the pending set (P). A subsequently processed native-unseen row with the same determinant configuration can then insert a mapping directly. Finally, batch completion overwrites that mapping but updates only the rows in (P).

Consider the following minimal example. Let the attribute order be (A,B,D,C), with FDs (A->B) and ((B,D)->C). Let the training relation be [(0,0,0,0),\ (1,1,1,1),] and let the initial synthetic rows, in processing order, be [(0,1,1,0),\ (0,0,1,1).] After repairing (A->B), both rows have ((B,D)=(0,1)). The first row enters (P), while the second is treated as native unseen and registers (M(0,1)=1). If the conditional model completes the first row with (C=0), the subsequent steps update only that row and the mapping. The final rows therefore have identical determinants but different dependent values, (C=0) and (C=1).

This execution path is permitted by the published pseudocode and differs from the conflicting assignments imposed by multiple FDs that the authors already acknowledge on page 6. It directly affects the central mechanism by which mapping reuse is intended to maintain consistency. The authors should clarify whether the implementation already handles this case and whether correcting it affects the reported results. This analysis alone does not establish that the authors’ implementation contains the same error.

W2: 2. The experiments do not sufficiently isolate the benefits of the additional repair mechanisms.

Appendix D.2 explicitly states that the baseline results underlying Table 3 are evaluated after FD filtering, whereas SynFD is evaluated on its repaired relation. Together with Appendix C.4, this implies differences in both final sample size and final FD satisfaction. For example, SynFD’s R_FD on Airbnb is 94.71%, rather than complete consistency. Consequently, the distributional advantages in the main table cannot be attributed entirely to better repair or conditional generation.

The matched FD-valid sample-size comparisons in Tables 14/15 are a useful control and should be credited. However, they cover only three datasets and TabDiff; Hospital and TPC-H are excluded because of the sampling budget limit. They therefore do not fully establish the overall conclusions across all datasets.

Moreover, the average gains from conditional completion over random assignment are relatively small. In Table 4, Trend increases from 0.793 to 0.799, α-Precision from 0.868 to 0.875, and F1 from 0.412 to 0.420. Table 7 reports conditional completion for only 0.30% of FD-processing instances on Airbnb; this percentage is not the fraction of affected rows. Without complete trigger statistics, a direct comparison against retaining the originally generated value, and the additional model’s training and inference costs, it is difficult to assess the incremental value of conditional diffusion over simpler mapping-based repair. This affects the evidence for the paper’s main methodological contribution, rather than merely the completeness of its reporting.

Questions for authors
  1. Algorithm and implementation: How does the implementation handle the case above, where the same determinant configuration appears in both the pending and native-unseen paths? Please check the counterexample, specify how a unique assignment is maintained across all relevant rows, and clarify whether the reported experiments are affected.

  2. Comparison protocol: Could the authors report the final sample count and final FD consistency for each method in the main table, and provide paired comparisons among the original TabDiff output, repaired SynFD output, and filtered SynFD output? For datasets where the target number of valid samples cannot be reached, could the authors report quality and actual computational cost under a common budget, so that repair benefits can be distinguished from filtering losses?

  3. Incremental value of the additional module: How many rows and FD-processing instances does MCGEN affect on each dataset? Starting from the same initial synthetic table, how large and stable are its gains over retaining and registering the original value or using empirical-distribution assignment, and what is its additional runtime? Please also clarify which stages are independently repeated in the “three independent runs,” and why the main SynFD distributional metrics all report zero standard deviations.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 5NwE (1/4)

We thank the reviewer for the constructive feedback. We address the concerns below and will incorporate the additional results and clarifications into the revised manuscript.

Algorithm 2 can produce FD conflicts even when no dependent attribute is constrained by multiple FDs (W1 / Q1).

We agree that the pseudocode does not specify synchronization between pending and previously processed rows sharing the same determinant. In our implementation, after conditional completion, we update Mr(s)\mathcal{M}_r(\mathbf{s}) and propagate the selected value to all rows with that determinant, rather than only those in the pending set P\mathcal{P}.

In the reviewer's counterexample, after repairing A→BA\rightarrow B, both rows share (B,D)=(0,1)(B,D)=(0,1). If completion selects C=0C=0, both rows are updated accordingly. We will revise Line 31 of Algorithm 2 to specify: "Assign xi,jr←vx_{i,j_r} \leftarrow v for every xi∈Tsyn\mathbf{x}_i \in \mathcal{T}_{\mathrm{syn}} with xi,Sr=s\mathbf{x}_{i,\mathcal{S}_r} = \mathbf{s}."

Appendix D.2 describes different FD-filtering protocols for baselines and SynFD (W2).

This is a wording error in Appendix D.2. Both the baselines and SynFD in Table 3 and Tables 8–12 were evaluated using the same FD-filtering procedure. The reported SynFD results are therefore based on repair followed by filtering. We will clarify this in the revised manuscript.

The final sample counts and FD validity of the evaluated outputs are not explicitly reported (Q2).

The tables below report the final FD-valid sample counts, average retention rates, and final FD consistency for all eight methods across five datasets, alongside the original distributional fidelity and downstream utility metrics. All methods undergo the same FD-filtering procedure, yielding 100% FD-consistent retained outputs. SynFD retains 98.40% of generated tuples on average, compared with 54.21% for TabDiff, while achieving higher Shape (0.9555 vs. 0.9222), Trend (0.7987 vs. 0.6923), and α-Precision (0.8748 vs. 0.7005).

(a) Final FD-valid samples and retention

MethodAirbnbBeijingHospitalMigraineTPC-HAvg. Keep (%)Final FD Cons. (%)
CTGAN + Filter126.7±8.128,493.0±22.61,619.7±19.9218.0±6.0126.7±10.039.35100.00
TVAE + Filter1,144.0±28.234,181.0±43.33,793.0±0.0300.7±0.6581.3±12.568.10100.00
GaussianCopula + Filter116.0±0.028,654.0±0.02,990.0±0.0259.0±0.06.0±0.049.18100.00
TabDDPM + Filter2,404.7±0.628,111.7±87.81,714.3±16.7298.3±0.66.7±0.663.56100.00
TabSyn + Filter1,808.0±14.530,172.0±41.8131.3±2.3296.0±4.4352.3±19.251.58100.00
TabDiff + Filter1,253.3±22.530,253.3±28.71,683.0±19.3272.7±6.0883.3±12.254.21100.00
GATD + Filter10.0±1.736,222.7±57.61,682.3±22.5295.3±1.221.3±4.747.87100.00
SynFD + Filter2,309.0±0.037,581.0±0.03,791.0±0.0293.0±0.021,141.0±0.098.40100.00

(b) Distributional fidelity and downstream utility

MethodShape ↑Trend ↑α-Precision ↑β-Recall ↑F1 ↑R² ↑
CTGAN + Filter0.7500±0.00380.4600±0.01040.7117±0.01070.3150±0.01690.3524±0.0333-0.4158±0.1787
TVAE + Filter0.7170±0.00300.5501±0.00630.4528±0.00800.2176±0.00460.7337±0.1140-0.5075±0.0694
GaussianCopula + Filter0.7812±0.00000.5128±0.00000.4940±0.00000.3055±0.00000.2732±0.0370-0.1566±0.2899
TabDDPM + Filter0.6746±0.00390.4721±0.00460.3009±0.00260.2150±0.02820.3017±0.0287-2.2317±0.2530
TabSyn + Filter0.8795±0.00240.6072±0.00480.5810±0.01070.3470±0.00800.4450±0.02790.2900±0.0666
TabDiff + Filter0.9222±0.00470.6923±0.00790.7005±0.01210.3977±0.01830.3682±0.02240.3173±0.0203
GATD + Filter0.3390±0.01240.3064±0.01330.5072±0.05170.6177±0.03920.5158±0.14990.3764±0.0040
SynFD + Filter0.9555±0.00000.7987±0.00000.8748±0.00000.3903±0.00000.4198±0.01650.3181±0.0819
AuthorsOpenReview ↗

Response to Reviewer 5NwE (2/4)

Additional stage-wise comparison (Q2).

To isolate the contribution of the DAG-guided repair framework, we compare four settings under the same initial generation budget: (i) TabDiff (raw), (ii) TabDiff + Filter, (iii) SynFD (repair), and (iv) SynFD + Filter, as shown in the table below.

Before FD filtering, SynFD increases the average proportion of FD-valid tuples from 54.21% to 98.40% (1.82×) without discarding any synthetic tuples. Trend improves from 0.7595 to 0.7991 and α-Precision from 0.8658 to 0.8776. These improvements are achieved through repair alone, without relying on FD filtering.

After applying the same FD-filtering procedure, SynFD outperforms TabDiff + Filter in average Shape (0.9555 vs. 0.9222), Trend (0.7987 vs. 0.6923), and α-Precision (0.8748 vs. 0.7005).

DatasetMethodFinal nFD-valid (%)Shape ↑Trend ↑α-Prec. ↑β-Rec. ↑F1 ↑R² ↑
AirbnbTabDiff (raw)2,43851.41±0.920.9663±0.00220.8983±0.00220.9614±0.01320.2320±0.0087---0.0607±0.0853
AirbnbTabDiff + Filter1,253.3±22.5100.000.9601±0.00360.8975±0.00640.9336±0.01980.3611±0.0160---0.0155±0.0203
AirbnbSynFD (repair)2,43894.71±0.000.9751±0.00000.9236±0.00000.9830±0.00000.2935±0.0000---0.0779±0.0152
AirbnbSynFD + Filter2,309100.000.9748±0.00000.9237±0.00000.9814±0.00000.3056±0.0000---0.0032±0.1433
BeijingTabDiff (raw)37,58180.50±0.080.9833±0.00020.9664±0.00410.9880±0.00180.5600±0.0005--0.6451±0.0343
BeijingTabDiff + Filter30,253.3±28.7100.000.9830±0.00040.9679±0.00370.9450±0.00100.5099±0.0017--0.6501±0.0203
BeijingSynFD (repair)37,581100.00±0.000.9834±0.00000.9662±0.00000.9893±0.00000.6256±0.0000--0.6395±0.0205
BeijingSynFD + Filter37,581100.000.9834±0.00000.9662±0.00000.9893±0.00000.6256±0.0000--0.6395±0.0205
HospitalTabDiff (raw)3,79344.37±0.510.8801±0.01230.5390±0.01170.4975±0.00420.0020±0.00110.2612±0.0038--
HospitalTabDiff + Filter1,683.0±19.3100.000.8417±0.01760.4503±0.01350.7290±0.02650.1383±0.02550.2662±0.0203--
HospitalSynFD (repair)3,79399.95±0.000.8910±0.00000.6235±0.00000.4939±0.00000.5995±0.00000.2498±0.0120--
HospitalSynFD + Filter3,791100.000.8910±0.00000.6234±0.00000.4948±0.00000.5539±0.00000.2563±0.0034--
MigraineTabDiff (raw)30190.59±2.000.9804±0.00270.9165±0.01140.9027±0.02140.1743±0.02650.2887±0.0933--
MigraineTabDiff + Filter272.7±6.0100.000.9717±0.00130.9048±0.01520.8832±0.01120.1828±0.03200.1775±0.0420--
MigraineSynFD (repair)30197.34±0.000.9798±0.00000.9276±0.00000.9377±0.00000.1927±0.00000.2757±0.0251--
MigraineSynFD + Filter293100.000.9761±0.00000.9254±0.00000.9244±0.00000.1975±0.00000.3413±0.0461--
TPC-HTabDiff (raw)21,1414.18±0.060.9490±0.00060.4773±0.00020.9794±0.00560.1360±0.00420.6560±0.0009--
TPC-HTabDiff + Filter883.3±12.2100.000.8547±0.00070.2411±0.00070.0116±0.00210.7967±0.01630.6610±0.0050--
TPC-HSynFD (repair)21,141100.00±0.000.9524±0.00000.5547±0.00000.9841±0.00000.2688±0.00000.6618±0.0000--
TPC-HSynFD + Filter21,141100.000.9524±0.00000.5547±0.00000.9841±0.00000.2688±0.00000.6618±0.0000--
AverageTabDiff (raw)13,05154.210.95180.75950.86580.22090.40200.2922
AverageTabDiff + Filter6,869100.000.92220.69230.70050.39770.36820.3173
AverageSynFD (repair)13,05198.400.95630.79910.87760.39600.39580.2808
AverageSynFD + Filter13,023100.000.95550.79870.87480.39030.41980.3181
AuthorsOpenReview ↗

Response to Reviewer 5NwE (3/4)

Matched FD-valid sample-size comparisons in Appendix Table 14 exclude Hospital and TPC-H (W2 / Q2).

Appendix Table 14 provides matched-size FD-valid comparisons on Airbnb, Beijing, and Migraine. For Hospital and TPC-H, we additionally evaluated TabDiff + Filter under a 50× generation budget, the maximum used in the original paper, as shown in the table below.

On Hospital, TabDiff + Filter retained only 2,399 of 189,650 generated rows (1.265%), below the target of 3,793. On TPC-H, it retained only 1,103 of 1,057,050 rows (0.104%), below the target of 21,141. In contrast, SynFD + Filter retained 3,791 of 3,793 and 21,141 of 21,141 rows, respectively, using only a 1× generation budget. SynFD also achieved higher Trend on Hospital (0.623 vs. 0.496) and TPC-H (0.555 vs. 0.257).

DatasetMethodGeneratedFD-validKeep (%)ShapeTrendα-Prec.β-Rec.F1
HospitalTabDiff + Filter (1×)3793168344.370.8420.4500.7290.1380.266
HospitalTabDiff (50×, Raw)189650----0.9280.6130.4970.002--
HospitalTabDiff + Filter (50×)18965023991.2650.8750.4960.8980.0780.252
HospitalSynFD + Filter (1×)3793379199.950.8910.6230.4950.5540.256
TPC-HTabDiff + Filter (1×)211418834.180.8550.2410.0120.7970.661
TPC-HTabDiff (50×, Raw)1057050----0.9610.4990.9820.139--
TPC-HTabDiff + Filter (50×)105705011030.1040.8450.2570.0350.7420.656
TPC-HSynFD + Filter (1×)2114121141100.000.9520.5550.9840.2690.662

Contribution of conditional completion (W2 / Q3).

As discussed in Section 4.1, when upstream FD repairs change downstream determinants, the original dependent values may no longer match the repaired context. Retain + Mapping maintains FD consistency through consistent assignments, but cannot adapt these values to the updated attributes. For example, if an upstream repair changes a person's income, retaining the original financial status may no longer reflect the updated income and property information, even if the FD is satisfied.

MCGEN addresses this limitation by generating dependent values conditioned on the repaired tuple and reusing the generated values consistently for rows sharing the same determinant. To evaluate its contribution, we compare Retain + Mapping, Random Assignment, and MCGEN under the same DAG-guided repair framework.

As shown in the table below, MCGEN achieves the highest average Trend (0.799), α-Precision (0.875), and R² (0.318). Compared with Retain + Mapping, it improves Trend from 0.796 to 0.799, α-Precision from 0.865 to 0.875, and R² from 0.266 to 0.318.

VariantShape ↑Trend ↑α-Prec. ↑β-Rec. ↑F1 ↑R² ↑
TabDiff0.9520.7600.8660.2210.4020.292
TabDiff + Filter0.9220.6920.7010.3980.3680.317
w/o Conditional (Random)0.9540.7930.8680.3940.4120.277
Retain + Mapping0.9550.7960.8650.3950.4220.266
HFGF-style Reconstruction0.9590.7590.8250.3760.4890.309
SynFD (MCGEN)0.9560.7990.8750.3900.4200.318

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.