TABFAIRGDT: Fast Fair Tabular Generator

• The paper introduces TABFAIRGDT, a method that uses autoregressive decision trees to generate synthetic tabular data with enhanced fairness and high utility.
• It employs a novel soft leaf resampling technique at the target-generation step to achieve statistical parity while limiting accuracy loss.
• Empirical results show TABFAIRGDT exceeds deep generative baselines by reducing fairness disparity, boosting speed by 72%, and preserving synthetic data quality.

TABFAIRGDT, introduced in "TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees" (Panagiotou et al., 24 Sep 2025), is a fairness-aware synthetic data generator for tabular datasets that replaces deep generative architectures with an autoregressive sequence of decision trees and enforces fairness through a soft leaf resampling step in the final target-generation tree. The method is designed to satisfy three requirements simultaneously: synthetic data quality, fairness control, and downstream utility preservation. It assumes binary sensitive and target variables, requires no data pre-processing, is CPU-compatible, and is presented as a non-parametric alternative to deep fair tabular generators. Empirically, it is reported to outperform state-of-the-art deep generative baselines on the fairness–utility trade-off and on several synthetic data quality measures, while achieving a 72% average speedup over the fastest SOTA baseline and generating fair synthetic data for a medium-sized dataset in about one second on a standard CPU (Panagiotou et al., 24 Sep 2025).

1. Problem formulation and design objectives

TABFAIRGDT operates on a labeled tabular dataset

$D \subseteq X \times S \times Y,$

where $X = X_0 \times X_1 \times \cdots \times X_n$ denotes the tabular feature space, $S$ is the sensitive attribute, and $Y$ is the binary target. The paper assumes

$S \in \{0,1\}, \qquad Y \in \{0,1\}.$

The synthetic dataset $$\hat D \subseteq \hat X \times \hat S \times \hat Y$$ is required to satisfy three explicit objectives: $\hat X \sim X,\quad \hat S \sim S$ for data quality,

$\hat Y \perp \hat S$

for fairness control, and

$\hat Y \sim P(Y \mid X)$

for utility preservation (Panagiotou et al., 24 Sep 2025).

These requirements encode a standard tension in fairness-aware synthetic data generation. Reducing the dependence between $S$ and $X = X_0 \times X_1 \times \cdots \times X_n$0 tends to conflict with preserving the predictive structure of $X = X_0 \times X_1 \times \cdots \times X_n$1 from $X = X_0 \times X_1 \times \cdots \times X_n$2. TABFAIRGDT addresses this tension at the data-generation level rather than through downstream constrained learning. In that respect it belongs to the same broad family of data-level fairness interventions as TabFairGAN, which imposes demographic parity through a second-stage fairness penalty in a WGAN-GP generator (Rajabi et al., 2021), but it does so with a markedly different model class and optimization mechanism.

A central design choice is that fairness is enforced only at the final target-generation step. This is important because the method does not attempt to rewrite the entire joint distribution for fairness. Instead, it preserves feature generation as faithfully as possible and localizes fairness control to the generated target, which the paper presents as a way to reduce discrimination while limiting utility degradation (Panagiotou et al., 24 Sep 2025).

2. Autoregressive decision-tree synthesis

The generator models the joint tabular distribution through sequential conditional sampling: $X = X_0 \times X_1 \times \cdots \times X_n$3 followed by

$X = X_0 \times X_1 \times \cdots \times X_n$4

from the synthesized feature set (Panagiotou et al., 24 Sep 2025).

For each feature $X = X_0 \times X_1 \times \cdots \times X_n$5, TABFAIRGDT trains a decision tree to predict $X = X_0 \times X_1 \times \cdots \times X_n$6 from the previously generated columns $X = X_0 \times X_1 \times \cdots \times X_n$7. Each fitted tree induces a leaf partition $X = X_0 \times X_1 \times \cdots \times X_n$8, and every leaf stores an empirical probability distribution over the possible values of $X = X_0 \times X_1 \times \cdots \times X_n$9: $S$0 Sampling proceeds by routing each synthetic prefix $S$1 through the tree, identifying its leaf, and sampling $S$2 from the corresponding leaf distribution. The first feature $S$3 is sampled directly from $S$4, since it has no parents (Panagiotou et al., 24 Sep 2025).

This construction gives TABFAIRGDT several properties emphasized by the paper. It is non-parametric, so it avoids explicit assumptions about the underlying data distribution. It handles mixed numerical and categorical features natively, without normalization, discretization, embedding, or other pre-processing. It is also lightweight and interpretable, because generation is reduced to tree fitting and leaf-level empirical sampling rather than neural density estimation (Panagiotou et al., 24 Sep 2025).

The method’s autoregressive structure is intended to capture dependencies common in real tabular data. The paper notes that if all features were fully independent, the autoregressive mechanism would be less useful, but the empirical premise is that practical tabular datasets usually contain strong conditional structure (Panagiotou et al., 24 Sep 2025).

3. Fairness control via soft leaf resampling

Fairness is imposed when generating the target $S$5. TABFAIRGDT first fits a target tree

$S$6

then modifies selected leaves so that the generated target becomes closer to statistical independence from the sensitive attribute (Panagiotou et al., 24 Sep 2025).

The fairness criterion is statistical parity: $S$7 At the tree level, the paper defines

$S$8

and

$S$9

For a leaf $Y$0, if its label were flipped from $Y$1 to $Y$2, the estimated fairness and accuracy changes are

$Y$3

and

$Y$4

Leaf selection is then posed as a knapsack-like optimization: $Y$5 with

$Y$6

Because exact optimization is NP-hard, the paper uses a greedy strategy based on the discrimination-to-accuracy ratio (Panagiotou et al., 24 Sep 2025).

The paper’s distinctive contribution is that selected leaves are not hard-relabeled. Instead, TABFAIRGDT uses soft leaf resampling. If a leaf has original output probabilities $Y$7, the adjusted probabilities are

$Y$8

with $Y$9. For a leaf with

$S \in \{0,1\}, \qquad Y \in \{0,1\}.$0

the paper gives

$S \in \{0,1\}, \qquad Y \in \{0,1\}.$1

and

$S \in \{0,1\}, \qquad Y \in \{0,1\}.$2

Thus $S \in \{0,1\}, \qquad Y \in \{0,1\}.$3 controls the fairness–utility trade-off continuously: $S \in \{0,1\}, \qquad Y \in \{0,1\}.$4 leaves the tree unchanged, $S \in \{0,1\}, \qquad Y \in \{0,1\}.$5 fully reverses the leaf distribution, and intermediate values provide partial correction (Panagiotou et al., 24 Sep 2025).

This localized intervention is also used to explain why TABFAIRGDT avoids out-of-distribution samples better than some baselines. The paper argues that fairness imposed too early or too broadly can generate unrealistic combinations such as $S \in \{0,1\}, \qquad Y \in \{0,1\}.$6 or $S \in \{0,1\}, \qquad Y \in \{0,1\}.$7. TABFAIRGDT instead generates all features first, then the sensitive attribute, and applies fairness only at the target step, thereby preserving more of the real conditional structure (Panagiotou et al., 24 Sep 2025).

4. Benchmarks, protocol, and evaluation criteria

The empirical study uses six benchmark datasets, all evaluated under a 3-fold protocol with approximately 66% train and 33% test per split. Generative models are trained only on the training portion, real test data remains unseen during generation, and results are averaged over the three folds. Utility and fairness are assessed by training a downstream LightGBM classifier on synthetic data and testing on real data (Panagiotou et al., 24 Sep 2025).

Dataset	Composition	Sensitive attribute / target
Adult Census	45k; 4 numerical / 8 categorical	sex / income
Dutch Census	60k; 0 numerical / 11 categorical	sex / occupation level
Bank Marketing	40k; 7 numerical / 9 categorical	marital status / deposit
KDD Census	95k; 7 numerical / 31 categorical	sex / income
ACS-I Utah	19k; 2 numerical / 7 categorical	sex / income
ACS-I Alabama	24k; 2 numerical / 7 categorical	sex / income

The baselines are all fairness-aware generators: TabularARGN, TabFairGAN, CuTS, FSMOTE, and PreFair (Panagiotou et al., 24 Sep 2025). TabFairGAN is the most directly comparable deep generative baseline in the paper’s discussion: it is a WGAN-GP-based fair generator trained in two phases, with a fairness penalty based on demographic parity (Rajabi et al., 2021). The comparison is therefore not between fairness-aware generation and unconstrained synthesis, but between competing fairness-aware generation strategies.

Evaluation spans three dimensions. Utility is measured by downstream ROC AUC. Fairness is measured by statistical parity on downstream predictions: $S \in \{0,1\}, \qquad Y \in \{0,1\}.$8 Synthetic data quality is measured through detection score (AUC), KS score for continuous features, TV score for categorical features, precision, recall, density, coverage, and DCR (Distance to Closest Record), where DCR ideally is around 1 and values below 1 may suggest overfitting (Panagiotou et al., 24 Sep 2025).

5. Empirical performance, robustness, and efficiency

The paper reports that increasing $S \in \{0,1\}, \qquad Y \in \{0,1\}.$9 improves fairness and usually decreases utility, as expected, but that TABFAIRGDT achieves the strongest overall trade-off. Averaged across datasets, the reported changes are

$$\hat D \subseteq \hat X \times \hat S \times \hat Y$$0

and

$$\hat D \subseteq \hat X \times \hat S \times \hat Y$$1

The paper also summarizes the method as delivering about a 50% average reduction in statistical parity difference with less than 2% average utility loss (Panagiotou et al., 24 Sep 2025).

Several dataset-specific observations are highlighted. On Dutch Census, which is entirely categorical, TABFAIRGDT is reported as the only method achieving notable fairness improvement with only about 2% utility loss. TabFairGAN is reported to fail on high-dimensional or imbalanced datasets such as KDD Census and Bank Marketing. PreFair often improves fairness but with large utility drops. TABFAIRGDT and TabularARGN are described as the only methods succeeding across all datasets without failure (Panagiotou et al., 24 Sep 2025).

On synthetic data quality, TABFAIRGDT is reported as the strongest method overall. The paper gives a detection score around 0.54, close to the ideal 0.5, strong KS, TV, precision, recall, density, and coverage values, and a DCR around 1.02, indicating no clear overfitting (Panagiotou et al., 24 Sep 2025). This is significant because the method’s fairness gains are not presented as arising from a collapse in distributional fidelity.

Runtime is a major part of the contribution. On a dataset with 10 features and 10k samples, TABFAIRGDT is reported to require 0.81 s for fitting, 0.25 s for sampling, and 1.05 s total, compared with 10.84 s for TabularARGN, 71.75 s for TabFairGAN, 882.65 s for CuTS, 29.19 s for FSMOTE, and 45.5 s for PreFair. The paper attributes this to parallel tree fitting, lightweight CPU-only execution, and the avoidance of deep neural training (Panagiotou et al., 24 Sep 2025).

Robustness analyses further emphasize that the method is not fragile to modeling choices. The authors test original feature order, ascending and descending order by correlation with $$\hat D \subseteq \hat X \times \hat S \times \hat Y$$2, and ascending and descending order by correlation with $$\hat D \subseteq \hat X \times \hat S \times \hat Y$$3, and report that feature order has no statistically significant effect on utility, fairness, or data quality. The parameter $$\hat D \subseteq \hat X \times \hat S \times \hat Y$$4 produces the expected monotone fairness–utility trade-off and is described as smoother than that of deep baselines (Panagiotou et al., 24 Sep 2025).

6. Position within the field and stated limitations

TABFAIRGDT occupies a specific place within the fairness literature on tabular machine learning. It is a data-level mitigation method: it attempts to produce fairer synthetic training data, rather than constraining the downstream learner directly or changing the data acquisition process. This distinguishes it from in-processing approaches such as FairGBM, which imposes proxy-Lagrangian fairness constraints during GBDT training (Cruz et al., 2022), and from acquisition-time approaches such as FairBED, which modifies experimental design to reduce information about sensitive attributes in the collected dataset (Hedman et al., 22 Jun 2026). It also differs from recent LLM-based synthesis frameworks such as FairTabGen, which integrates counterfactual and causal fairness into a GPT-4o-driven prompt orchestration pipeline (Nagesh et al., 15 Aug 2025).

Within fair synthetic data generation specifically, TABFAIRGDT’s defining claim is that deep architectures are not necessary to obtain strong fairness-aware performance on tabular data. TabFairGAN uses a neural WGAN-GP with a fairness-augmented generator objective (Rajabi et al., 2021), whereas TABFAIRGDT uses autoregressive decision trees and leaf-level probability manipulation (Panagiotou et al., 24 Sep 2025). This suggests a methodological split between fairness-aware tabular generation by distribution learning in neural latent spaces and fairness-aware generation by explicit conditional sampling with localized tree corrections.

The paper is explicit about several limitations. The current formulation assumes binary sensitive and target variables. The fairness criterion is statistical parity only, so other notions are not supported in the present implementation. The method is expected to be most effective when the data exhibit meaningful feature dependencies; if features were fully independent, autoregression would be less valuable. Sampling remains sequential, so generation time grows with dataset size even if fitting is very fast. The paper also identifies future work on intersectional fairness, regression tasks, continuous sensitive attributes, alternative fairness constraints, and tree depth or overfitting effects (Panagiotou et al., 24 Sep 2025).

A common misconception would be to treat TABFAIRGDT as a general-purpose causal fairness model. The paper does not make that claim. Its fairness mechanism is explicitly statistical-parity-oriented and target-local. Relative to frameworks that reason about transport maps, causal pathways, or counterfactual invariance (Kim et al., 6 Jan 2025, Nagesh et al., 15 Aug 2025), TABFAIRGDT is narrower in fairness semantics but more lightweight in computation and deployment. A plausible implication is that its strongest contribution is operational rather than ontological: it provides a fast, interpretable, and practical generator for settings where statistical parity is the operative requirement and mixed-type tabular fidelity must remain high.