Pushing the limits of fairness impossibility: Who's the fairest of them all?

Authors: Brian Hsu, Rahul Mazumder, Preetam Nandy, Kinjal Basu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show experiments demonstrating that our post-processor can improve fairness across the different definitions simultaneously with minimal model performance reduction.
Researcher Affiliation Collaboration Brian Hsu Linked In Corporation Sunnyvale, CA bhsu@linkedin.com Rahul Mazumder Linked In Corporation, Sunnyvale, CA (Massachusetts Institute of Technology, Cambridge, MA) rmazumder@linkedin.com Preetam Nandy Linked In Corporation Sunnyvale, CA pnandy@linkedin.com Kinjal Basu Linked In Corporation Sunnyvale, CA kbasu@linkedin.com
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code No The paper mentions utilizing an "open-source implementation" for a specific transformation method (NMDT) at https://github.com/joaquimg/QuadraticToBinary.jl, but does not state that the code for their own methodology is being released or is available.
Open Datasets No In Table 1, we take each dataset, create a 60/40 train-test split, train a grid-searched random forest model, and score the training data... We discuss the datasets and problem parameters for all experiments in the Appendix ??. The named datasets are ACS Income, ACS Insurance, ACS Mobility, ACS Poverty, ACS Coverage, ACS Travel, Heart Disease, COMPAS. However, no specific access information (links, DOIs, repositories) or formal citations with authors and year for these datasets are provided within the main paper.
Dataset Splits No In Table 1, we take each dataset, create a 60/40 train-test split, train a grid-searched random forest model, and score the training data. This explicitly mentions a train-test split but no validation split.
Hardware Specification No No specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments were found.
Software Dependencies Yes Our MILP solution solved by Gurobi ([15]) against the QCQP problem solved by IPOPT ([21]) and the reference [15] is "Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2022."
Experiment Setup Yes Next, we discretize the scores into bins, parameterize the problem (# bins, ϵ, max movement, window size, solve time) on the scored training data, and compare our MILP solution solved by Gurobi ([15]) against the QCQP problem solved by IPOPT ([21]).