A Sequentially Fair Mechanism for Multiple Sensitive Attributes
Authors: Francois Hu, Philipp Ratz, Arthur Charpentier
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A data-driven estimation procedure for the derived solution is developed, and comprehensive numerical experiments are conducted on both synthetic and real datasets. Our empirical findings decisively underscore the practical efficacy of our post-processing approach in fostering fair decision-making. |
| Researcher Affiliation | Academia | Franc ois Hu1, Philipp Ratz2, Arthur Charpentier2 1Universit e de Montr eal, Montr eal, Qu ebec, Canada 2Universit e du Qu ebec a Montr eal, Montr eal, Qu ebec, Canada |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are provided in the paper. |
| Open Source Code | Yes | 1Code source at: https://github.com/phi-ra/Sequential Fairness. |
| Open Datasets | Yes | To illustrate a possible application of our methodology and showcase its use in a real world use-case, we consider data collected in the folktables package of (Ding et al. 2021). (...) Specifically, we consider synthetic data (X, A, Y ) with the following characteristics: X 2 Rd: Comprises d non-sensitive features generated from a centered Gaussian distribution X Nd(0, σXId), where σX > 0 parameterizes its variance. A = A1:r 2 { 1, 1}r: Represents r binary sensitive features, with Ai 2 B(qi) 1 following a Bernoulli law with parameter qi = P( X > i), where X N(0, σX). Here, = ( 1, . . . , r) is a user-set parameter. Y N(1T X+1T A, 1): Represents the regression task. |
| Dataset Splits | Yes | We generated 10,000 synthetic examples and divided the data into three sets (50% train, 25% test and 25% unlabeled). (...) The data is split into 64% train, 20% test and 16% unlabeled data. (...) As a base model, we opt for a light GBM (Ke et al. 2017) model with early stopping, where the early stopping iterations are optimized using 5-fold cross validation on the training data. |
| Hardware Specification | No | No specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) are provided. |
| Software Dependencies | No | The paper mentions 'scikit-learn in Python', 'light GBM (Ke et al. 2017) model', 'Fairlearn package (Weerts et al. 2023)', and 'Fair Balance from (Yu, Chakraborty, and Menzies 2021)' but does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | Default parameters are set as follows: d = 10, r = 3, σX = 0.15, and = (0, 0.05, 0.1). (...) As a base model, we opt for a light GBM (Ke et al. 2017) model with early stopping, where the early stopping iterations are optimized using 5-fold cross validation on the training data. |