Fair Regression: Quantitative Definitions and Reduction-Based Algorithms

Authors: Alekh Agarwal, Miroslav Dudik, Zhiwei Steven Wu

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we evaluate our method on several standard datasets, on the tasks of least-squares and logistic regression under statistical parity, with linear and tree-ensemble learners, and compare it with the unconstrained baselines as well as the technique of Johnson et al. (2016). Our method uncovers fairness accuracy frontiers and provides the first systematic scheme for enforcing fairness in a significantly broader class of learning problems than prior work.
Researcher Affiliation Collaboration 1Microsoft Research, Redmond, WA 2Microsoft Research, New York, NY 3University of Minnesota, Minneapolis, MN.
Pseudocode Yes Algorithm 1 Fair regression with statistical parity; Algorithm 2 Fair regression with Bounded Group Loss
Open Source Code No The paper does not provide an explicit statement or link for the open-sourcing of the code for its described methodology.
Open Datasets Yes Adult: The adult income dataset (Lichman, 2013); Law school: Law School Admissions Council s National Longitudinal Bar Passage Study (Wightman, 1998); Communities & crime: The dataset contains socio-economic, law enforcement, and crime data about communities in the US (Redmond & Baveja, 2002)
Dataset Splits No Thus we ended up with a total of five datasets, and split each into 50% for training and 50% for testing.
Hardware Specification No The paper does not provide any specific details about the hardware used for the experiments.
Software Dependencies No We considered two variants of LS and LR oracles: linear learners from scikit-learn (Pedregosa et al., 2011), and tree ensembles from XGBoost (Chen & Guestrin, 2016). The paper mentions software by name but does not provide version numbers.
Experiment Setup Yes We ran Algorithm 1 on each training set over a range of constraint slack values ˆ", with a fixed discretization grid of size 40: Z = {1/40, 2/40, . . . , 1}.