TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees

Emmanouil Panagiotou, Benoît Ronval, Arjun Roy, Ludwig Bothmann, Bernd Bischl, Siegfried Nijssen, Eirini Ntoutsi

Abstract

Ensuring fairness in machine learning remains a significant challenge, as models often inherit biases from their training data. Generative models have recently emerged as a promising approach to mitigate bias at the data level while preserving utility. However, many rely on deep architectures, despite evidence that simpler models can be highly effective for tabular data. In this work, we introduce TABFAIRGDT, a novel method for generating fair synthetic tabular data using autoregressive decision trees. To enforce fairness, we propose a soft leaf resampling technique that adjusts decision tree outputs to reduce bias while preserving predictive performance. Our approach is non-parametric, effectively capturing complex relationships between mixed feature types, without relying on assumptions about the underlying data distributions. We evaluate TABFAIRGDT on benchmark fairness datasets and demonstrate that it outperforms state-of-the-art (SOTA) deep generative models, achieving better fairness-utility trade-off for downstream tasks, as well as higher synthetic data quality. Moreover, our method is lightweight, highly efficient, and CPU-compatible, requiring no data pre-processing. Remarkably, TABFAIRGDT achieves a 72% average speedup over the fastest SOTA baseline across various dataset sizes, and can generate fair synthetic data for medium-sized datasets (10 features, 10K samples) in just one second on a standard CPU, making it an ideal solution for real-world fairness-sensitive applications.

Introduction

Machine learning models are only as fair as the data they are trained on. If a dataset contains historical biases against specific demographic groups based on race, gender, or age, the resulting model will likely inherit and amplify those biases. A popular mitigation strategy is generating fair synthetic data—creating a version of the dataset that preserves statistical relationships while breaking the link between sensitive attributes and target outcomes.

Current state-of-the-art (SOTA) generators often rely on complex deep learning architectures like GANs or Diffusion models. These models can be "overkill" for tabular data, requiring heavy pre-processing, expensive GPUs, and long training times. Furthermore, they frequently produce "out-of-distribution" (OOD) samples—logically impossible feature combinations—which are unacceptable in high-stakes domains.

Methodology: Tabular Fair Generative Decision Trees

TABFAIRGDT (Tabular Fair Generative Decision Trees) addresses these issues by replacing deep neural networks with a sequence of Decision Trees (DTs).

Overview of the TABFAIRGDT framework: dataset structure, autoregressive tree fitting, and fair sampling.

Autoregressive Generation

The model generates features sequentially. Each feature is predicted by a DT conditioned on all previously generated features. This non-parametric approach naturally handles mixed data types (numerical and categorical) without complex encoding or normalization.

Strategic Ordering

To prevent unrealistic data generation, the model follows a specific order: standard features first, then the sensitive attribute ($S$), and finally the target variable ($Y$). This ensures the sensitive attribute does not "pollute" the generation of other features, mitigating OOD issues.

Soft Leaf Resampling

The core mechanism for fairness is "soft leaf resampling." During the generation of the target $Y$, the model identifies leaves in the decision tree that contribute most to bias. Instead of hard relabeling, it adjusts the probability distribution within these leaves based on a user-defined parameter $\lambda$. This allows for fine-tuning the trade-off between fairness (Statistical Parity) and utility (Accuracy).

Experimental Results

The researchers evaluated TABFAIRGDT on six benchmark datasets, including Adult Census, Dutch Census, Bank Marketing, and modern ACS-I variants.

Fairness-utility trade-offs (ROC AUC vs Statistical Parity) across benchmark datasets.

Efficiency and Performance

TABFAIRGDT achieved a 72% average speedup over the fastest SOTA baseline. It can generate fair data for a medium-sized dataset (10 features, 10k samples) in just one second on a standard CPU. In terms of performance, the method improved fairness (Statistical Parity) by approximately 50% while losing less than 2% in predictive utility (ROC AUC).

Data Quality

The method achieved superior data quality scores, with a detection score of only 0.54%, indicating that synthetic samples are nearly indistinguishable from real ones. Notably, it was the only method to successfully improve fairness for the Dutch Census dataset, which is entirely categorical—a typical challenge for deep learning models.

Conclusion

TABFAIRGDT demonstrates that simpler models like Decision Trees can outperform deep learning for fair tabular data generation. By using localized interventions like soft leaf resampling and strategic feature ordering, it avoids the "OOD trap" and provides a highly efficient, CPU-compatible solution. Future work will explore expanding to other fairness definitions like Equalized Odds, handling continuous targets in regression tasks, and addressing intersectional fairness.

Summary updated

Abstract

Ensuring fairness in machine learning remains a significant challenge, as models often inherit biases from their training data. Generative models have recently emerged as a promising approach to mitigate bias at the data level while preserving utility. However, many rely on deep architectures, despite evidence that simpler models can be highly effective for tabular data. In this work, we introduce TABFAIRGDT, a novel method for generating fair synthetic tabular data using autoregressive decision trees. To enforce fairness, we propose a soft leaf resampling technique that adjusts decision tree outputs to reduce bias while preserving predictive performance. Our approach is non-parametric, effectively capturing complex relationships between mixed feature types, without relying on assumptions about the underlying data distributions. We evaluate TABFAIRGDT on benchmark fairness datasets and demonstrate that it outperforms state-of-the-art (SOTA) deep generative models, achieving better fairness-utility trade-off for downstream tasks, as well as higher synthetic data quality. Moreover, our method is lightweight, highly efficient, and CPU-compatible, requiring no data pre-processing. Remarkably, TABFAIRGDT achieves a 72% average speedup over the fastest SOTA baseline across various dataset sizes, and can generate fair synthetic data for medium-sized datasets (10 features, 10K samples) in just one second on a standard CPU, making it an ideal solution for real-world fairness-sensitive applications.
BibTeX
@inproceedings{panagiotou2025tabfairgdt,
  author = {Emmanouil Panagiotou and Benoît Ronval and Arjun Roy and Ludwig Bothmann and Bernd Bischl and Siegfried Nijssen and Eirini Ntoutsi},
  title = {{TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees}},
  booktitle = {ICDM},
  year = {2025},
  url = {https://openreview.net/forum?id=wdoumVssdq},
}
Figure