{"id":"80a38323-1d39-4cd8-9a52-de8dbaf4a865","arxiv_id":"2507.19211","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"HFGF improves preservation of functional and logical dependencies in synthetic tabular data by generating independent features and reconstructing dependent features from predefined mapping rules.","lead":"Standard AI generators often break the rules that link columns in a data table. This paper's HFGF wrapper generates only the independent columns and then rebuilds the dependent ones from predefined rules, improving rule preservation in synthetic data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The measured improvement is largely tautological: dependent features are reconstructed from the same mapping rules used to build the benchmarks and detected by FDTool and the Q-function, so the experiments do not test whether HFGF helps generative models preserve dependencies.","rationale":"Why this concern is load-bearing: the paper's only experiments are on four synthetic benchmarks whose dependencies are generated from the same config dictionary that HFGF uses for reconstruction. There is no experiment in which dependencies are unknown a priori, inferred from data, and then used by HFGF, even though the method section says real-data dependencies can be extracted with FDTool. The exclusion of constraint-aware baselines such as CuTS, KAMINO, C3TGAN, or LLM-TabFlow further weakens the comparative claim. I do not question the authors' honesty; the code is public and the framework's limitation is stated in the concluding section ('rather than inherently learning these complex relationships'). But the central claim, as phrased in the abstract, implies that HFGF improves dependency preservation as a method; the evidence only shows that deterministic post-hoc reconstruction preserves dependencies when the underlying category coverage is adequate. This supports a CONDITIONAL verdict: the framework may be practically useful, but the claims need to be narrowed and tested on real data with inferred dependencies and fair baselines. Since the reader already issued CONDITIONAL, no verdict change is needed.","tokens_in":12559,"tokens_out":4399,"duration_ms":46075,"concrete_test":"Take a real-world tabular dataset (e.g., one of the five public datasets used in ref [4]) and infer FD/LD rules from the real data with FDTool and the Q-function. Run HFGF with those inferred rules, and compare FD/LD preservation against (a) bare CTGAN/TVAE/GReaT and (b) an oracle-reconstruction baseline that uses the real independent-feature values with the same inferred rules. Run each condition for at least 5 seeds and report mean and spread. If HFGF does not clearly beat (b), and its margin over (a) is reproduced by simply re-attaching deterministic mapping rules, then the reported improvement is an artifact of evaluation design rather than evidence that HFGF improves dependency-aware generation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence is circular. In Section 3 (Algorithm 1), each benchmark dataset is built by drawing independent features from config and then setting dependent features via fi.mapping[category] (or a probabilistic one-to-many sample). HFGF's second step applies exactly those same mapping rules to the generated independent features, and the third step scores preservation with FDTool and the Q-function, which detect exactly those dependencies. Consequently, the FD/LD scores in Tables 1-3 measure essentially whether the generative model reproduced the needed categories of the (only four) independent features; the dependent-feature reconstruction is deterministic and not learned. The authors effectively concede this in Section 5's 'Failure cases' paragraph: if a class is missing from generated independent features, the dependency is lost. Thus the abstract's claim that HFGF 'improves the preservation of FDs and LDs across six generative models' is supported only in a setting where the dependency rules are supplied as oracle inputs and the generative task is reduced to a lower-dimensional marginal. The finding is true by construction, not an empirical discovery about dependency-aware generation, and it does not generalize to settings where dependencies must be discovered or involve numerical features (a limitation stated at the end of Section 5).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Hierarchical Feature Generation Framework (HFGF), which separates features into independent and dependent sets based on known functional and logical dependencies, generates the independent features with a standard generative model, and then reconstructs the dependent features by applying the predefined mapping rules. The authors create four synthetic benchmark datasets with controlled dependency structure and evaluate six generative models (CTGAN, CTABGAN+, TVAE, NextConvGeN, TabuLa, GReaT) with and without HFGF, reporting the percentage of FDs and LDs preserved as measured by FDTool and a Q-function. The stated conclusion is that HFGF consistently improves FD/LD preservation across all six models.","tokens_in":12739,"tokens_out":6229,"duration_ms":60158,"significance":"If the claim held as stated, the framework would offer a simple, model-agnostic way to enforce known categorical dependencies in synthetic tabular data, which is relevant for privacy-sensitive domains such as healthcare. The paper also contributes controlled benchmark datasets with known dependencies and makes code available. However, the central evaluation is largely circular: the dependent features in the benchmark are generated from the same mapping rules that HFGF uses for reconstruction, and the evaluation tools detect exactly those rules. The experiments therefore do not test whether generative models can learn dependencies, but only whether the independent features are generated with sufficient fidelity. This substantially limits the significance of the empirical results, although the benchmark resource and the explicit separation of independent and dependent features may still be useful to the community.","major_comments":[{"comment":"The central evaluation is circular. In Algorithm 1, the benchmark datasets are constructed by setting dependent features using fi.mapping rules (one-to-one, many-to-one, one-to-many). HFGF then reconstructs dependent features in the synthetic data by applying exactly those same mapping rules to the generated independent features, and preservation is measured with FDTool and the Q-function, which detect exactly those rules. Consequently, the high FD/LD preservation scores under HFGF are a definitional consequence of the construction rather than an empirical finding about the generative models' ability to preserve dependencies. The results in Tables 1-3 essentially test only whether the generative model reproduces the needed categories of the few independent features. To make the claim meaningful, the authors should either explicitly frame HFGF as an oracle-based post-processing wrapper for known dependencies or add experiments that break this circularity, for example by deriving dependencies from a real dataset with FDTool and then applying HFGF, or by comparing against a base generative model that attempts to model all features jointly while evaluated on the same dependencies.","section":"Section 3, Algorithm 1 and Section 4, Tables 1-3"},{"comment":"The claim that HFGF \"consistently improves\" preservation is contradicted by the paper's own tables. In Table 1, TVAE LD preservation drops from 60 to 39 (Case 1); in Table 2, TVAE LD preservation drops from 71 to 33 (Case 3); in Table 3, NextConvGeN FD preservation drops from 100 to 75 (Case 1) and from 100 to 86 (Case 3). While Section 4 acknowledges two of these decreases, the abstract, introduction, and conclusion repeat the unqualified claim. The manuscript should state the actual scope, e.g., \"improves in most cases\" or \"improves on average\", and should analyze why the decreases occur, since the text currently attributes only the TVAE LD drops to mode collapse and does not reconcile these counterexamples with the headline claim.","section":"Abstract, Introduction, and Section 4, Tables 1-3"},{"comment":"All reported preservation percentages are point estimates from a single run per model and case, with no error bars, standard deviations, or number of seeds. Since the six generative models are stochastic and two of the datasets have only 100 rows, the observed jumps (e.g., CTGAN FD preservation from 0 to 75) and the decreases (e.g., NextConvGeN from 100 to 75) could be dominated by seed variance. The authors should report results over multiple random seeds (at least five, ideally ten) with mean and standard deviation, or justify why the effect sizes are robust to seed choice. Without this, the quantitative claims in Tables 1-3 are not statistically grounded.","section":"Section 4, Tables 1-3"},{"comment":"The manuscript does not specify how \"preserved LDs (%)\" is computed from the Q-function scores. The Q-function returns a continuous value in [0,1], where 0 indicates a functional dependency and intermediate values indicate logical dependency, but the text does not define the threshold or matching rule that converts scores into a binary \"preserved\" decision for the LD set. Similarly, it is unclear how the set of LDs in the benchmark data is enumerated from Q-scores and how the comparison with synthetic data is performed. Please provide an explicit operational definition of LD preservation, as this is essential for interpreting Tables 1-3.","section":"Section 4 and Figure 2"},{"comment":"The framework assumes that the dependency structure is known in advance and expressible as categorical mapping rules, and the authors state in the Failure cases paragraph that \"only four independent features are generated, which proves to be tractable for the models even with limited training data.\" This indicates that the evaluated scenario is low-dimensional and that the generative task reduces to modeling a few uncorrelated marginals. The paper should test the approach in a more demanding setting, for instance with many independent features (more than ten) or with numerical dependent features, to assess scalability and to support the claim that HFGF can be applied generally in clinical settings. As written, the limitations section acknowledges the constraint but the conclusions still generalize beyond the evidence.","section":"Section 5, Limitations and Future Directions"}],"minor_comments":[{"comment":"The author names are inconsistent between the title page (\"Chaithra Umesha\", \"Kristian Schultza\") and the CRediT statement (\"Chaithra Umesh\", \"Kristian Schultz\"); please ensure the names match throughout.","section":"Title page and CRediT statement"},{"comment":"The Q-function definition in Equation (1) has garbled typesetting: the fraction and the condition are compressed into one line. Please rewrite with proper display format and clearly define the symbols A, B, T, and the relation ~T.","section":"Equation (1)"},{"comment":"The caption states \"Incorporation of HFGF consistently improves dependency preservation across all models,\" but Tables 1-3 contain counterexamples (NextConvGeN Cases 1 and 3; TVAE Cases 1 and 3). The caption should be revised to reflect the actual, qualified results.","section":"Figure 2 caption"},{"comment":"The claim that \"models with HFGF not only improve dependency preservation but also align better with real data distributions\" is not fully supported by the Peacock test p-values: for TVAE the p-value decreases from 0.077 without HFGF to 0.017 with HFGF, and for CTGAN the increase is from 0.006 to 0.063, which is borderline. Please provide a more systematic comparison of the p-values or a formal statistical test over multiple embeddings.","section":"Section 5, Figure 3"},{"comment":"The statement that in one-to-one FDs \"either feature may be treated as independent\" is useful but should be justified or accompanied by a short example, since it affects the determination of the independent set in Algorithm 1.","section":"Section 1, one-to-one FDs"}],"recommendation":"major_revision","confidential_remarks":"The core issue is that the benchmark evaluation is circular: the dependent features are reconstructed from the same rules used to generate the benchmarks and to measure preservation. I would encourage the editor to ask the authors to reframe the contribution either as a constrained post-processing wrapper for known dependencies or to add a non-circular evaluation on real-world data with discovered dependencies. The paper's potential value lies mainly in the benchmark datasets and the explicit independent/dependent decomposition, which are useful resources even if the headline claim about 'consistently improves' is not supported as stated. The authors may also consider comparing with constraint-aware generators such as CuTS and Kamino, despite the stated differences in task focus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper describes a simple wrapper—generate the independent columns with a standard model, then rebuild the dependent columns from a hand-specified mapping table. That is a legitimate engineering trick, and the authors are honest about what it does. But the experiments as designed cannot support the headline claim, because the measured improvement is mostly a tautology.\n\nThe stress-test note is correct. In Algorithm 1 the benchmark data is built by drawing independent features and then setting dependent features via fi.mapping[category]. HFGF's second step applies exactly those same mapping rules, and the third step scores with FDTool and the Q-function, which detect exactly those dependencies. So the FD/LD preservation numbers in Tables 1-3 largely measure whether the base model reproduced the needed categories in the four independent features. The reconstruction of dependent features is deterministic and not learned. The authors even acknowledge this in Section 5: if a class is missing in the generated independent features, the dependency is lost. So the improvement is a consequence of restricting the generative task to a lower-dimensional marginal, not an empirical discovery about dependency-aware generation.\n\nThat said, there is real value here. The paper is clearly written, the code and notebooks are available, and the idea of constructing benchmarks with controlled FD/LD structure is useful for the community. The discussion of failure cases is honest. And for a practitioner who knows their dependency rules in advance—say, sex and pregnancy, or postal code and city—the wrapper is a cheap and reliable way to enforce them.\n\nThe soft spots, in order of importance. First, the circular evaluation is the load-bearing flaw; the abstract's 'consistently improves' needs to be tempered with 'by construction under known rules.' Second, the claim of consistent improvement is not even true in the tables: TVAE's LD preservation drops in Cases 1 and 3 with HFGF, and NextConvGeN loses FD points in Cases 1 and 3. Third, there are no error bars or repeated-seed results, so we can't tell if any of these differences are noise. Fourth, the exclusion of constraint-aware baselines like CuTS, KAMINO, and LLM-TabFlow is a bit convenient; some of those methods handle dependencies without needing the rules as oracle input.\n\nWho is this for? Someone working on synthetic tabular data who has explicit domain constraints and wants a simple post-processing step. It is a modest engineering contribution, not a research breakthrough. I would not cite it in my own work, but I'd bring it to a reading group to discuss evaluation design in synthetic data papers.\n\nRecommendation: send it to peer review—the idea is coherent and the code is there—but expect major revision. The authors need to rerun with dependencies discovered on a real dataset, include constraint-aware baselines, report variance, and rewrite the claims to match what the method actually does.","headline":"A sensible wrapper for enforcing known FD/LD rules into synthetic tabular data, but the benchmark evaluation is circular and the 'consistent improvement' claim goes beyond the evidence.","tokens_in":13298,"tokens_out":2616,"would_cite":false,"duration_ms":24794,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-step generation framework preserves inter-attribute rules in synthetic tabular data.","keywords":["synthetic tabular data","functional dependencies","logical dependencies","hierarchical feature generation","generative models","CTGAN","TVAE","GReaT"],"falsifier":"Run HFGF on a benchmark in which one independent categorical feature has a rare class that the base generative model never emits; if the preserved FD and LD percentages then fall to roughly the no-HFGF baseline, the improvement is conditional on class coverage rather than intrinsic to the framework.","tokens_in":12329,"feed_emoji":"🧩","tokens_out":6582,"duration_ms":59991,"temperature":0.7,"pith_summary":"The paper claims that synthetic tabular data can preserve inter-attribute rules—functional dependencies (FDs) and logical dependencies (LDs)—much more reliably if the generative model is not asked to reproduce the full joint distribution. The proposed Hierarchical Feature Generation Framework (HFGF) generates only the independent features with a standard generative model, then reconstructs the dependent features from predefined mapping rules. The authors evaluate this on four controlled benchmark datasets with known dependencies, using six generative models, and report that the wrapper improves dependency preservation in nearly all settings. This matters because privacy-sensitive fields such as healthcare need synthetic copies that obey the same deterministic and rule-based constraints as the original records.","feed_headline":"Two-step generation preserves inter-attribute rules in synthetic data","feed_subtitle":"Wrapping six generative models with the new framework lifts preservation of functional and logical dependencies.","key_machinery":"The load-bearing object is a configuration dictionary that encodes each feature's type and, for every dependent feature, a mapping rule from the parent feature's categories to the dependent feature's categories; one-to-one and many-to-one mappings are deterministic functional dependencies, while one-to-many mappings are probabilistic logical dependencies. Algorithm 1 builds the benchmark datasets from this dictionary, and the same dictionary drives the reconstruction of dependent features after the generative model produces synthetic independent features. The framework's effect is measured by FDTool, which mines functional dependencies, and the Q-function, a score between 0 and 1 in which 0 marks functional dependence, 1 marks independence, and intermediate values mark logical dependence.","core_discovery":"The central claim is that dependency preservation is a separation problem rather than a modeling problem: once the features of a table are split into an independent set and a dependent set, any generative model trained only on the independent features, combined with a reconstruction step that applies the known mapping rules, produces synthetic tables in which the known FDs and LDs survive far more often than when the model generates every column jointly. Preservation is measured as the percentage of benchmark dependencies that FDTool and the Q-function rediscover in the synthetic table. In the paper's four benchmark scenarios, every tested model—CTGAN, CTABGAN+, TVAE, NextConvGeN, TabuLa, and GReaT—preserves more FDs with HFGF than without, and in the complex imbalanced case all six also preserve more LDs; two small-case comparisons are called out where TVAE preserves more LDs and NextConvGeN preserves more FDs without the wrapper.","pith_inferences":["A natural next test is to apply HFGF to real-world datasets whose dependencies are inferred by FDTool rather than predefined; the gain would then depend on the completeness and accuracy of those inferred rules.","Because the bottleneck shifts to the independent features, an obvious practical add-on is a coverage check that resamples or conditions until every class needed by the mappings appears.","The same separation idea could extend to numerical-numerical and numerical-categorical dependencies by first discretizing numerical features into bins, an extension the authors name as future work.","The harder problem the framework side-steps is learning the dependency structure itself; a model that inferred FDs and LDs from real data would remove the need for a hand-written configuration dictionary."],"forward_implications":["A user who knows the dependency structure of a table can preserve that structure without modifying the generative model itself, simply by wrapping it in HFGF.","Models that normally require large datasets can keep dependencies intact on tables as small as 100 rows, because the generative step only has to learn a low-dimensional distribution of a few independent features.","Dependency preservation becomes a checkable engineering condition: if every category of each independent feature appears in the synthetic sample, the reconstructed dependent features are correct by construction.","The framework applies only when the table has genuinely independent features and the dependencies are categorical mapping rules; fully mutually dependent tables fall outside its scope."],"supporting_citations":[{"why":"Defines the Q-function for logical dependencies and reports that no prior generative model preserved both FDs and LDs, the gap HFGF targets.","marker":"[4]"},{"why":"Supplies FDTool, used to extract functional dependencies from benchmark and synthetic data during evaluation.","marker":"[6]"},{"why":"A survey documenting that generative models do not preserve intricate inter-attribute relationships, motivating the framework.","marker":"[10]"},{"why":"Introduces NextConvGeN, one of the six generative models wrapped by HFGF.","marker":"[14]"},{"why":"Introduces CTGAN, one of the six base models on which HFGF is evaluated.","marker":"[17]"},{"why":"Introduces CTABGAN+, one of the six base models on which HFGF is evaluated.","marker":"[20]"},{"why":"Describes the variational autoencoder architecture underlying TVAE, one of the six base models.","marker":"[22]"},{"why":"Introduces GReaT, the language-model-based generator wrapped by HFGF.","marker":"[25]"},{"why":"Introduces TabuLa, the LLM-based generator wrapped by HFGF.","marker":"[27]"}],"fun_headline_variants":["Split and reconstruct: preserving data dependencies in synthetic tabular data","Two-step framework keeps functional and logical dependencies in synthetic tables","HFGF: generating dependent features after independent ones to preserve rules","Decoupled generation boosts dependency fidelity across six synthetic data models","Reconstructing dependent features after generative modeling preserves table rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dependency structure must be known in advance and expressible as categorical mapping rules, and the base generative model must produce every category value of the independent features that those mappings require.","fun_headline_variants_meta":{"raw":{"variants":["Split and reconstruct: preserving data dependencies in synthetic tabular data","Two-step framework keeps functional and logical dependencies in synthetic tables","HFGF: generating dependent features after independent ones to preserve rules","Decoupled generation boosts dependency fidelity across six synthetic data models","Reconstructing dependent features after generative modeling preserves table rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000909,"raw_usage":{"total_tokens":3880,"prompt_tokens":894,"completion_tokens":2986,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":2903}},"tokens_in":510,"tokens_out":2986,"duration_ms":20790,"temperature":1.0,"reasoning_tokens":2903,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:57:26.780426+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HFGF on a benchmark in which one independent categorical feature has a rare class that the base generative model never emits; if the preserved FD and LD percentages then fall to roughly the no-HFGF baseline, the improvement is conditional on class coverage rather than intrinsic to the framework.","supporting_citations":[{"cited_title":"Umesh, K","cited_arxiv_id":null,"evidence_quote":"Defines the Q-function for logical dependencies and reports that no prior generative model preserved both FDs and LDs, the gap HFGF targets."},{"cited_title":"Buranosky, E","cited_arxiv_id":null,"evidence_quote":"Supplies FDTool, used to extract functional dependencies from benchmark and synthetic data during evaluation."},{"cited_title":"Convex space learning for tabular synthetic data generation","cited_arxiv_id":"2407.09789","evidence_quote":"Introduces NextConvGeN, one of the six generative models wrapped by HFGF."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces CTGAN, one of the six base models on which HFGF is evaluated."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces CTABGAN+, one of the six base models on which HFGF is evaluated."},{"cited_title":"Vahdat, J","cited_arxiv_id":null,"evidence_quote":"Describes the variational autoencoder architecture underlying TVAE, one of the six base models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces TabuLa, the LLM-based generator wrapped by HFGF."}],"review_version":2}