PaperlayerView exact revision on arXiv

Source-linked AI summary

Creation begins with understanding: LLMs as strategy designers for privacy-preserving tabular data synthesis

Jinmeng Li, Quan Zhang, Hangting Ye, He Zhao, Firas Laakom, Dandan Guo, Jürgen Schmidhuber

arXiv:2608.29674v1cs.LG

TL;DR

Privacy regulations constrain tabular-data sharing, while existing deep generative and record-serializing LLM methods are costly, difficult to audit, or structurally mismatched to tables. TabSSD instead uses an LLM to design synthesis procedures from tree-derived dependence summaries, executing and evaluating the resulting programs locally. Across twelve datasets, it balances fidelity, predictive utility, and empirical privacy risk while reducing computation and token use.

  • Problem

    Privacy regulations constrain tabular-data sharing, while deep generative models are costly to train and audit and serialized-record LLM methods obscure tabular structure and can expose sensitive data.

  • Method

    TabSSD gives an LLM tree-derived summaries of variable dependence, which it translates into locally executable Python synthesis programs for evaluation and selection.

  • Results

    Across twelve datasets, TabSSD achieves a favourable balance among statistical fidelity, predictive utility, and empirical privacy risk while substantially reducing computational and token costs.

  • Takeaways & Limitations

    TabSSD lowers expertise and infrastructure barriers by producing inspectable synthesis procedures without requiring user-designed deep models, LLM fine-tuning, extensive hyperparameter tuning, or local GPU training.

  • Takeaways & Limitations

    TabSSD depends on adequate tree-derived summaries and reliable LLM outputs, while chained trees may miss weak or higher-order dependence and context windows can limit scalability.

Abstract

from arXiv · show

Sharing tabular data in high-stakes domains is constrained by privacy regulations. Synthetic data offer a promising alternative, but deep generative models are costly to train and difficult to audit, while LLM-based methods often serialize records as text, obscuring tabular structure and exposing sensitive data. We introduce Tabular Synthesis Strategy Designer (TabSSD), which uses an LLM to design synthesis procedures rather than directly generate records. TabSSD provides the LLM with tree-derived summaries of variable dependence rather than raw records, which produces Python programs for local execution and evaluation. Across twelve datasets, TabSSD strikes a favourable balance among statistical fidelity, predictive utility, and empirical privacy risk, achieving the best average rank across six metrics among ten methods. Moreover, it substantially reduces local computation and token consumption relative to the compared methods. By enabling human-guided refinement and eliminating user-side model tuning, TabSSD lowers the expertise and infrastructure barriers to transparent tabular data synthesis.

Introduction

TabSSD addresses the structural and privacy challenges of tabular synthesis by using an LLM to design inspectable synthesis strategies rather than generate records directly. It combines tree-derived dependence summaries with local execution and validation to balance fidelity, utility, privacy, and practical cost.

  • Tabular synthesis must preserve distributions and variable relationships while minimizing memorization and re-identification risks.
  • Deep generative models offer modelling flexibility but encode data distributions implicitly within neural network parameters, making them difficult to inspect.
  • Serialized-record LLM methods obscure variable dependence, impose artificial order, and can expose raw data during adaptation or prompting.
  • TabSSD uses chained classification and regression trees to summarize dependence, then has an LLM select sampling strategies and generate executable Python code.The resulting synthesis logic can be audited, diagnosed, and revised without computationally expensive retraining.
  • Across twelve heterogeneous datasets, TabSSD balances statistical fidelity, downstream utility, and privacy while reducing computational and token costs relative to representative methods.A clinical case study also evaluates minority-outcome augmentation and preservation of clinically meaningful phenotype distributions.

Results

Across twelve datasets, TabSSD balances fidelity, predictive utility, privacy risk, and efficiency by having an LLM design inspectable synthesis programs from tree-derived dependence summaries. It also supports local execution and human-guided code editing without raw-record prompts or user-side model tuning.

  • Framework: TabSSD extracts variable dependence with chained trees, prompts an LLM with compact summaries, and locally executes and evaluates multiple Python synthesis strategies.Evaluation covers statistical fidelity, empirical privacy risk, and computational efficiency.
  • Overall performance: Across twelve datasets, TabSSD combines low statistical errors, competitive predictive utility, and a favourable empirical privacy profile.The comparison spans deep generative, diffusion, serialized-LLM, tree-based, energy-based, and interpolation methods.
  • Statistical fidelity: 3.80% mean marginal distribution error and 3.89% mean pairwise correlation error place TabSSD near the leading methods on statistical fidelity.The marginal error is second-lowest, while the pairwise correlation error is lowest among the compared methods.
  • Overall performance: 2.17 average rank across six metrics makes TabSSD the strongest overall ranking profile among the evaluated methods.The six metrics cover marginal fidelity, pairwise fidelity, utility, adjusted DCR, α-precision, and δ-presence.
  • Efficiency: 199.6 seconds average end-to-end runtime makes TabSSD faster than ARF’s 332.6 seconds and the other baselines requiring 3.4 to 38.4 times its runtime.TabSSD uses CPU-only local execution, whereas deep generative baseline runtimes were measured with GPU acceleration.
  • Efficiency: Up to 130.3 times higher token consumption for serialization shows why compact chained-tree rules improve high-dimensional context management.For DARWIN’s 451 variables, TabSSD needs rules from 54 trees and avoids replicating the entire dataset in the prompt.
  • Clinical case study: 0.578 mean balanced accuracy and 0.033 tumour-category total-variation distance show strong clinical augmentation results under the Thoracic Surgery case study.TabSSD achieved the highest balanced accuracy and most closely reproduced the observed tumour-category distribution among compared methods.

Discussion

TabSSD repositions LLMs as designers of inspectable synthesis procedures rather than record-level generators. Across benchmark datasets and case studies, it balances utility, fidelity, privacy, and efficiency while supporting local verification and human intervention.

  • Framework: TabSSD separates local dependence extraction, LLM strategy design, synthesis execution, and validation into an inspectable workflow.Chained trees summarize dependence; the LLM generates executable programs, which are run and evaluated in the data holder’s environment.
  • Benchmark findings: Across 12 benchmark datasets, TabSSD balances statistical fidelity, downstream predictive utility, empirical privacy risk, and computational efficiency.The paper reports a favourable balance across these four dimensions.
  • Applications: Executable synthesis code enables targeted fairness editing and application-specific validation in case studies.The strategy logic can be inspected and modified rather than being embedded only in learned parameters or token-by-token generation.
  • Practical implications: TabSSD lowers expertise and infrastructure barriers by automating dependence extraction, strategy design, local validation, and candidate selection.It avoids requiring deep generative-model design, LLM fine-tuning, extensive hyperparameter tuning, or local GPU training.
  • Limitations: TabSSD depends on adequate tree-derived summaries and reliable LLM outputs, which may miss higher-order dependence or produce invalid, inefficient, or low-fidelity routines.The authors mitigate these risks through multi-candidate generation, local execution, and code inspection.
  • Limitations: Ultra-high-dimensional tables can exceed LLM context windows, while formal privacy mechanisms and human-centred audit studies remain future needs.The paper also identifies information leakage reduction and practitioner understanding as open directions.

Methods

TabSSD extracts variable dependence with chained trees, gives compact structural summaries to an LLM, and locally executes and filters the resulting synthesis strategies.

  • Problem setup: The workflow models tabular data as features and, when applicable, a target, then seeks synthetic data preserving distributions and dependence without compromising privacy.For unsupervised data, the target and target tree are omitted without changing the workflow.
  • Variable dependence extraction: Chained trees predict each feature from the remaining features and predict the target from the features, capturing relationships without parametric assumptions.The resulting trees provide the dependence structure used for strategy design.
  • Variable dependence extraction: Split variables are ordered by their minimum tree depth, with root-near variables treated as more influential predictors.The method records each tree’s split-variable set A(Xj).
  • Dependence encoding: Tree goodness-of-fit summaries use R2 and RMSE for continuous variables and prediction accuracy for categorical variables.These statistics are encoded alongside variable types, ranges, and split variables.
  • Dependence encoding: Compact dependence signals replace full decision rules to limit prompt overhead, especially for high-dimensional datasets.For the target tree, split rules and leaf-level target distributions are retained.
  • Prompt design: The structured prompt combines task, metric, requirement, dependence, and output components to guide executable strategy generation.Metric guidance encourages reasoning about fidelity, utility, and empirical privacy trade-offs rather than directly optimizing them.
  • Candidate generation: Multiple LLM queries generate candidate strategies that may differ in generation order, subspace partitioning, and model choice.Because outputs are stochastic, candidates can have different fidelity, utility, privacy risk, and execution time.
  • Validation and selection: Candidates are executed locally, invalid or non-executable programs are discarded, and survivors are filtered by runtime and δ-presence risk before fidelity-based selection.The statistical error E is the average of marginal-distribution error and pairwise-correlation error.

A.1 Statistics of the evaluated datasets

The evaluation covers real-world classification and regression datasets varying substantially in dimensionality, sample size, and class imbalance.

  • Dataset scope: 13 real-world datasets span classification, regression, small-sample settings, and high-dimensional feature spaces.The dataset collection is used to evaluate the method across varied tabular conditions.
  • Dataset scope: The datasets contain between 4 and 450 features and between 174 and 53,940 samples.Each dataset is divided into training and test sets for downstream machine-learning evaluation.
  • Dataset statistics: The evaluated dataset statistics distinguish numerical and categorical predictors, training and test samples, task type, and majority-to-minority class ratio.The supplementary table uses Num, Cat, and Imb. for these dataset characteristics.

A.2 Implementation details

The evaluation protocol standardizes data splits, candidate execution, privacy-risk measures, predictive utility, and fidelity metrics across synthesis methods.

  • All methods are fit only on the real training split and generate synthetic tables with the same number of rows as that split.
  • Candidate strategies are generated through repeated LLM queries, required to use a unified executable interface, and filtered when they fail execution or violate requirements.
  • Adjusted DCR measures excessive proximity of synthetic records to real training records relative to the train–test split expectation; lower values indicate lower potential memorization risk.
  • δ-presence measures the worst-case real-to-synthetic count ratio across shared k-means partitions and retained granularities, with lower values preferred.
  • Downstream utility uses train-on-synthetic, test-on-real evaluation with AdaBoost, Random Forest, and XGBoost, averaging normalized predictive performance relative to real-data training.
  • Statistical fidelity is assessed through marginal distribution error and pairwise correlation error, using distributional, correlation, and contingency similarities for numerical, categorical, and mixed-type variables.

B.4 δ-presence results

The reported evaluation materials define δ-presence as an empirical re-identification-risk measure, while the supplied dataset-level tables provide broader fidelity, utility, and privacy metrics. The passages include dataset-level results for marginal errors, pairwise correlation errors, α-precision, predictive performance, DCR, and δ-presence.

  • Statistical fidelity: Supplementary Table 2 reports dataset-level marginal distribution errors, where lower values indicate better fidelity.Results are reported as mean ± s.d. across 20 runs, with an average row across dataset-level mean errors.
  • Statistical fidelity: Supplementary Table 3 reports pairwise correlation errors, where lower values indicate better preservation of feature-pair dependence.Results are reported as mean ± s.d. across 20 runs, with an average row across dataset-level mean errors.
  • Statistical fidelity: Supplementary Table 4 reports α-precision scores, where higher values indicate better statistical fidelity.The average row reports the mean of dataset-level mean scores.
  • Downstream utility: Supplementary Table 5 reports downstream predictive performance using mean AUROC for classification and mean RMSE for regression.Higher AUROC and lower RMSE indicate better performance; results are averaged across AdaBoost, Random Forest, and XGBoost.
  • Privacy evaluation: Supplementary Table 6 reports DCR values, where values closer to the train-set proportion indicate lower empirical memorization risk.Results are reported as mean ± s.d. across 20 runs where available.
  • δ-presence results: δ-presence scores quantify empirical re-identification risk, with lower values indicating lower risk.The average row reports the mean of dataset-level mean scores.

D Prompt and generated strategy for the Adult dataset

The Adult-dataset prompt combines task and metric requirements with compact summaries of tree-derived feature dependence and a complete label-tree structure. It asks the LLM to interpret these signals, route samples into suitable subspaces, and generate an executable synthesis strategy rather than directly sample from tree rules.

  • Prompt construction: The prompt is organized into task specification, evaluation criteria, core strategy requirements, feature inter-dependency information, and complete label-tree structure.The feature-dependence and label-tree components are represented separately in the prompt construction.
  • Evaluation criteria: The prompt evaluates synthetic data with machine-learning efficiency, statistical fidelity, and privacy criteria.Downstream utility is assessed with the TSTR protocol, while pairwise and column-wise metrics assess statistical properties and adjusted DCR assesses memorization risk.
  • Tree inputs: Multiple decision trees predict individual features from the other features, while the final decision tree predicts the original label.The feature-tree outputs are restricted to structured summaries, and the label tree is provided completely.
  • Feature-dependence summaries: Feature summaries encode variable type, value range or class information, associated splitting features, and predictive performance.The associated-feature order follows the top-down splitting order of the original tree.
  • Strategy-generation requirements: The LLM must infer feature dependencies and distribution characteristics, construct a routing function, and fit suitable models within each routed subspace.The instructions explicitly prohibit treating decision-tree rules as direct record-sampling rules.
  • Synthesis objective: The prompt’s operational goal is to produce high-quality synthetic samples whose downstream-model performance is close to models trained on the original data.Within routed subspaces, the strategy should learn real-data distribution characteristics and generate samples.

E Ablation results

The ablation studies test dependence extraction, dependence representation and ordering, prompt-guidance components, LLM backbones, and query budget. Overall, the supplied passages support a balanced advantage for the proposed choices, while also identifying context-window and backbone-sensitivity constraints.

  • Experimental scope: The ablation analysis covers eight datasets because complete feature-tree prompts exceed the LLM context window for remaining high-dimensional datasets.Restricting the analysis to common evaluable datasets preserves a consistent comparison protocol.
  • Dependence extraction: Chained-tree dependence extraction provides a stronger overall balance of marginal fidelity, pairwise-dependence fidelity, and downstream utility than linear-model and mutual-information alternatives.The complete label tree and all other prompt components remain unchanged across these variants.
  • Representation and ordering: The proposed dependence representation and ordering provide a favourable balance across evaluated metrics and datasets, although individual variants can win on particular datasets.Compared variants retain full feature trees, condense the label tree, or shuffle associated-feature order.
  • Prompt guidance: The full prompt provides more balanced synthesis performance than prompt variants that remove metric, routing, or reference-analysis guidance.Removing one component can occasionally improve a particular metric on a particular dataset.
  • LLM backbone sensitivity: TabSSD remains applicable across Gemini-2.5-Pro, GLM-5, and Qwen3-Max under the same prompt, query count, and selection protocol.Synthesis performance still varies across backbones and datasets.
  • Query budget: The query-budget analysis selects the lowest best-so-far average statistical error among the first k candidate strategies on Adult, Default, and Blood.Average statistical error is the mean of marginal distribution error and pairwise correlation error.

F Detailed runtime of different algorithms on various datasets

The runtime analysis decomposes TabSSD into tree extraction, LLM strategy generation, and synthesis execution, and compares total and parallel runtimes with baseline algorithms. The supplied passages define the reported runtime columns and notation but do not include the numerical dataset-level values.

  • TabSSD decomposition: TabSSD total runtime is decomposed into feature-CART extraction, selected-strategy generation, and selected-strategy execution.The total is the sum of Tree, Query, and Synth.
  • Runtime reporting: Supplementary Table 12 reports runtime across datasets, including single-run and 10-run parallel TabSSD runtimes in seconds.The table also reports total runtime for each baseline.
  • Table notation: A dash indicates that a method could not be successfully executed on the dataset.NRGB, TDDPM, and TSYN abbreviate NRGBoost, TabDDPM, and TABSYN.
  • Parallel execution: For parallel candidate runs, parallel query and parallel synthesis denote the maximum single-query and single-synthesis times across runs.Parallel total is defined using the corresponding parallel components.

G Case study of failure modes

TabSSD can generate unreliable synthesis strategies whose failures span fidelity, privacy, and efficiency. Inspectable code and multi-objective evaluation allow these failures to be traced and excluded before deployment.

  • LLM-generated strategies can fail across fidelity, privacy, and efficiency dimensions.The framework presents representative failure cases and excludes unreliable strategies through multi-objective evaluation.
  • Repeated single-sample generation with a fixed random state can collapse strongly dependent continuous variables onto a few recurring value pairs.The affected variables are feature_1 and feature_2; routing attributes may also leave their assigned subspaces.
  • 28.80% marginal distribution error and 23.82% pairwise correlation error characterize one reported failure.
  • Resampling an observed row with probability 0.98 causes most synthetic records to be copied from the training support rather than newly generated.Laplace smoothing changes sampling probabilities but cannot prevent exact matches.
  • A 144-feature strategy can incur substantial overhead by training separate conditional boosting models for nearly every feature in multiple subspaces.Synthesis also generates each record feature by feature, adding nested training and sampling overhead.

H Detailed predictive results for clinical data augmentation

The clinical augmentation analysis reports predictive utility under a 2× augmentation setting, evaluating balanced accuracy and minority-class accuracy on a real test set.

  • Predictive utility is evaluated under the 2× augmentation setting.
  • Balanced accuracy and minority-class accuracy are measured on the real test set.
  • Supplementary Table 14 organizes the reported predictive results by method and the two accuracy metrics.

I Tailored data synthesis strategies

TabSSD produces executable, dataset-tailored synthesis strategies by routing records into subspaces and fitting local generative models. The supplied examples show decision-tree routing, subspace-specific modeling, sampling, and output procedures.

  • The generated strategies are executable Python programs tailored to selected datasets, with DARWIN shown in abridged form.
  • Decision-tree logic routes records into subspaces using feature thresholds and leaf identifiers.Examples include routing based on f0, f1, f2, and f3, and a Blood strategy with 20 leaf identifiers.
  • Within each subspace, strategies fit local models or distributions for features and labels.The implementations include conditional boosting models, copula correlations, label distributions, and feature distributions conditioned on labels and subspaces.
  • Synthesis samples are generated from fitted subspace models, with continuous values clipped to routing bounds and labels sampled from learned probabilities.
  • The generated records are assembled with feature values and labels, then written to CSV after adjusting the sample count to n_sample.
  • The DARWIN strategy combines a feature-dependence graph, topological generation order, label-based routing, and subspace-specific models.Listed local models include linear regression with residual sampling and empirical distributions.
Loading 2608.29674v1…