TL;DR
TABFAIRGDT is a new, lightweight framework that generates fair synthetic tabular data using autoregressive decision trees instead of complex deep learning. It achieves a 72% speedup over state-of-the-art models—generating 10,000 samples in just one second on a standard CPU—while improving fairness by roughly 50% with minimal impact on predictive utility.
What Problem Does This Paper Address?
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.
One popular solution is to generate fair synthetic data: creating a new version of the dataset that preserves the statistical relationships of the original but breaks the link between sensitive attributes and the target outcome. However, most current "State-of-the-Art" (SOTA) generators rely on deep learning architectures like GANs or Diffusion models. These models are often "overkill" for tabular data; they require heavy pre-processing, expensive GPUs, and long training times. Worse, they frequently produce "out-of-distribution" (OOD) samples—unrealistic data points like "Male/Wife" or "Female/Husband" in a dataset where those combinations don't exist—which can be disastrous in high-stakes fields like healthcare.
Methods
The authors introduce TABFAIRGDT (Tabular Fair Generative Decision Trees), which swaps deep neural networks for a sequence of Decision Trees (DTs).
- Autoregressive Generation: The model generates features one by one. 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 needing complex encoding or normalization.
- Strategic Ordering: To avoid unrealistic data, the model generates all standard features first, then the sensitive attribute ($S$), and finally the target variable ($Y$). This ensures that the sensitive attribute doesn't "pollute" the generation of other features.
- Soft Leaf Resampling: This is the "secret sauce" for fairness. In the final step (generating the target $Y$), the model identifies leaves in the decision tree that contribute most to bias. Instead of a "hard" relabeling (flipping a 0 to a 1), it uses a "soft" resampling technique. By adjusting the probability distribution within specific leaves based on a user-defined parameter ($\lambda$), the model can finely tune the trade-off between fairness (Statistical Parity) and utility (Accuracy).
Datasets
The researchers evaluated their method on six benchmark fairness datasets:
- Adult Census & KDD Census: Predicting income based on demographic data.
- Dutch Census: Predicting occupation level (notable for being entirely categorical).
- Bank Marketing: Predicting if a client will subscribe to a term deposit.
- ACS-I Utah & ACS-I Alabama: Modern variants of the Adult Census dataset.
Key Insights and Ideas
- Simpler is Often Better: While the field has rushed toward deep learning, this paper proves that Decision Trees are often superior for tabular data. They are faster, require no data scaling, and are much easier to debug.
- The OOD Trap: Deep generative models often "hallucinate" feature combinations that are physically or logically impossible. By generating the sensitive attribute after the other features, TABFAIRGDT ensures the synthetic data respects real-world constraints.
- Localized Intervention: Unlike other models that apply fairness constraints across the entire data distribution (which can degrade overall quality), TABFAIRGDT’s "soft leaf resampling" only intervenes in the specific "pockets" of data where bias is most prevalent.
Key Results
- Efficiency: 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 one second on a standard CPU, whereas deep learning models often require minutes and a GPU.
- Fairness-Utility Trade-off: On average, the method improved fairness (Statistical Parity) by ~50% while losing less than 2% in predictive utility (ROC AUC).
- Data Quality: It achieved the highest data quality scores across the board, with a "detection score" of only 0.54% (meaning a discriminator model almost entirely failed to distinguish between real and synthetic samples).
- Categorical Mastery: It was the only method that successfully improved fairness for the Dutch Census dataset, which consists solely of categorical features—a known pain point for deep learning models.
Limitations and Future Work
While highly efficient, the autoregressive nature of the model means that as the number of features grows into the hundreds or thousands, the sequential generation process will eventually slow down.
The current work also focuses primarily on Statistical Parity (ensuring equal positive outcome rates across groups). The authors acknowledge that future iterations should expand to other fairness definitions (like Equalized Odds), handle regression tasks (continuous targets), and address "intersectional fairness" (where an individual might be discriminated against based on a combination of attributes, like race and gender).
Summary updated