tabfairgdt 0.1.0

Fast Fair Tabular Data Generator using Autoregressive Decision Trees

TabFairGDT

TabFairGDT — Fast Fair Tabular Data Generator using Autoregressive Decision Trees.

Generates synthetic tabular data with built-in fairness constraints via leaf-relabeling of CART models.

Installation

pip install tabfairgdt

Quick Start

import pandas as pd
from tabfairgdt import TABFAIRGDT

df = pd.read_csv("data.csv")
df["income"] = df["income"].astype("category")
df["sex"] = df["sex"].astype("category")

dtype_map = {col: str(df[col].dtype) for col in df.columns}
# Map pandas dtypes to tabfairgdt types
dtype_map = {
    col: "category" if df[col].dtype.name in ("category", "object", "bool")
    else "float" if df[col].dtype.kind == "f"
    else "int"
    for col in df.columns
}

generator = TABFAIRGDT(
    protected_attribute="sex",
    target="income",
    criterion="dp",          # demographic parity
    dtype_map=dtype_map,
    seed=42,
)
generator.fit(df, lamda=0.5)   # lamda=0: no fairness, lamda=1: max fairness
synthetic_df = generator.generate(k=1000)

Parameters

Parameter	Default	Description
protected_attribute	required	Column name of the sensitive/protected attribute
target	required	Column name of the target variable
dtype_map	required	Dict mapping column names to dtypes ('int', 'float', 'category', 'bool', 'datetime')
criterion	"dp"	Fairness criterion: "dp" (demographic parity)
lamda	0.5	Fairness–utility tradeoff passed to fit(): 0 = no fairness, 1 = max fairness
acc_threshold	-1	Max allowed accuracy drop as a fraction (e.g., 0.05 = 5%); -1 = no limit
seed	None	Random seed for reproducibility
parallel	True	Use parallel fitting across columns
re_order	False	Reorder features by correlation ("corr_asc_target", "corr_desc_target", etc.)
smoothing	False	Apply kernel smoothing to continuous columns ("density" or column dict)
proper	False	Bootstrap training data during fitting

Citation

@inproceedings{panagiotou2025tabfairgdt,
  title     = {TABFAIRGDT: A Fast Fair Tabular Data Generator using Autoregressive Decision Trees},
  author    = {Panagiotou, Emmanouil and Ronval, Benoît and Roy, Arjun and Bothmann, Ludwig and Bischl, Bernd and Nijssen, Siegfried and Ntoutsi, Eirini},
  booktitle = {IEEE ICDM 2025},
  year      = {2025},
  url       = {https://arxiv.org/abs/2509.19927}
}