The easiest way to install DEMV is from the PyPI repository:
pip install demvClone this repository and then install the following libraries:
pandasnumpyscikit-learn
The source code of DEMV is inside the DEMV folder.
Please cite our papers if you use DEMV in your experiments:
Giordano d’Aloisio, Andrea D’Angelo, Antinisca Di Marco, Giovanni Stilo, Debiaser for Multiple Variables to enhance fairness in classification tasks, Information Processing & Management, Volume 60, Issue 2, 2023, 103226, ISSN 0306-4573, https://doi.org/10.1016/j.ipm.2022.103226
@article{daloisio_debiaser_2023,
title = {Debiaser for Multiple Variables to enhance fairness in classification tasks},
journal = {Information Processing & Management},
volume = {60},
number = {2},
pages = {103226},
year = {2023},
issn = {0306-4573},
doi = {https://doi.org/10.1016/j.ipm.2022.103226},
url = {https://www.sciencedirect.com/science/article/pii/S0306457322003272},
author = {Giordano d’Aloisio and Andrea D’Angelo and Antinisca {Di Marco} and Giovanni Stilo},
keywords = {Machine learning, Bias and Fairness, Multi-class classification, Preprocessing algorithm, Equality},
}d’Aloisio, G., Stilo, G., Di Marco, A., D’Angelo, A. (2022). Enhancing Fairness in Classification Tasks with Multiple Variables: A Data- and Model-Agnostic Approach. In: Boratto, L., Faralli, S., Marras, M., Stilo, G. (eds) Advances in Bias and Fairness in Information Retrieval. BIAS 2022. Communications in Computer and Information Science, vol 1610. Springer, Cham. https://doi.org/10.1007/978-3-031-09316-6_11
@inproceedings{d2022enhancing,
title={Enhancing Fairness in Classification Tasks with Multiple Variables: A Data-and Model-Agnostic Approach},
author={d’Aloisio, Giordano and Stilo, Giovanni and Di Marco, Antinisca and D’Angelo, Andrea},
booktitle={International Workshop on Algorithmic Bias in Search and Recommendation},
pages={117--129},
year={2022},
organization={Springer}
}DEMV is a Debiaser for Multiple Variables that aims to increase Fairness in any given dataset, both binary and categorical, with one or more sensitive variables, while keeping the accuracy of the classifier as high as possible. The main idea behind the proposed method is that to enhance the classifier’s fairness during pre-processing effectively is necessary to consider all possible combinations of the values of the sensitive variables and the label’s values for the definition of the so-called sensitive groups.
We approach the problem by recursively identifying all the possible groups given by combining all the values of the sensible variables with the belonging label (class). Next, for each group, we compute its expected (𝑊𝑒𝑥𝑝) and observed (𝑊𝑜𝑏𝑠) sizes and look at the ratio among these two values. If 𝑊𝑒𝑥𝑝/𝑊𝑜𝑏𝑠 = 1, it implies that the group is fully balanced. Otherwise, if the ratio is less than one, the group size is larger than expected, so we must remove an element from the considered group accordingly to a chosen deletion strategy. Finally, if the ratio is greater than one, the group is smaller than expected, so we have to add another item accordingly to a generation strategy. For each group, we recursively repeat this balancing operation until 𝑊𝑒𝑥𝑝/𝑊𝑜𝑏𝑠 converge to one. It is worth noting that, in order to keep a high level of accuracy, the new items added to a group should be coherent in their values with the already existing ones.
The papers describing our work are available at:
-
round_level : floatTolerance value to balance the sensitive groups
-
debug : boolPrints w_exp/w_obs, useful for debugging
-
stop : intMaximum number of balance iterations
-
iter : intMaximum number of iterations
-
__init__(self, sensitive_vars, round_level=1, stop=10000, verbose=False)Args ---------- sensitive_vars : list List of sensitive variable names round_level : float, optional Tolerance value to balance the sensitive groups (default is 1) stop : int, optional Maximum number of iterations to balance the sensitive groups (default is 10000) verbose : bool, optional Prints w_exp/w_obs, useful for debugging (default is False) -
fit(self, x: pd.DataFrame, y: np.ndarray)Balances the dataset's sensitive groups
Args ---------- x : pd.DataFrame Dataset to be balanced y : array-like Labels of the dataset Returns ------- x: Balanced dataset y: Balanced labels of the dataset -
transform(self, x: pd.DataFrame, y: np.ndarray)Balances the dataset's sensitive groups
Args ---------- x : pd.DataFrame Dataset to be balanced y : array-like Labels of the dataset Returns ------- x: Balanced dataset y: Balanced labels of the dataset -
fit_transform(self, x: pd.DataFrame, y: np.ndarray)Balances the dataset's sensitive groups
Args ---------- x : pd.DataFrame Dataset to be balanced y : array-like Labels of the dataset Returns ------- x: Balanced dataset y: Balanced labels of the dataset -
get_iters(self)Gets the maximum number of iterations Returns ------- int: maximum number of iterations -
get_disparities(self)Returns the list of w_exp/w_obsReturns: list: list of disparities values
In the following we show an example usage of our algorithm:
from demv import DEMV
import pandas as pd
df = pd.read_csv('some_data.csv')
protected_attrs = ['s1','s2']
label = 'l'
demv = DEMV(sensitive_vars = protected_attrs, round_level = 1)
x = df.drop(label, axis=1)
y = df[label]
x_new, y_new = demv.fit_transform(x, y)
print('Maximum number of iterations: ',demv.get_iters())The original paper was written by Giordano d'Aloisio, Giovanni Stilo, Antinisca di Marco and Andrea D'Angelo. This work is partially supported by Territori Aperti a project funded by Fondo Territori Lavoro e Conoscenza CGIL CISL UIL, by SoBigData-PlusPlus H2020-INFRAIA-2019-1 EU project, contract number 871042 and by “FAIR-EDU: Promote FAIRness in EDUcation institutions” a project founded by the University of L’Aquila. All the numerical simulations have been realized mostly on the Linux HPC cluster Caliban of the High-Performance Computing Laboratory of the Department of Information Engineering, Computer Science and Mathematics (DISIM) at the University of L’Aquila.
This work is licensed under AGPL 3.0 license.