Fair Wrapping for Black-box Predictions

Authors: Alexander Soen, Ibrahim M. Alabdulmohsin, Sanmi Koyejo, Yishay Mansour, Nyalleng Moorosi, Richard Nock, Ke Sun, Lexing Xie

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

Reproducibility Variable Result LLM Response
Research Type Experimental We exemplify the use of our technique in three fairness notions: conditional valueat-risk, equality of opportunity, and statistical parity; and provide experiments on several readily available datasets.
Researcher Affiliation Collaboration Alexander Soen Australian National University alexander.soen@anu.edu.au Ibrahim Alabdulmohsin Google Research ibomohsin@google.com Sanmi Koyejo Google Research Stanford University sanmik@google.com Yishay Mansour Google Research Tel Aviv University mansour@google.com Nyalleng Moorosi Google Research nyalleng@google.com Richard Nock Google Research Australian National University richardnock@google.com Ke Sun Australian National University CSIRO s Data61 sunk@ieee.org Lexing Xie Australian National University lexing.xie@anu.edu.au
Pseudocode Yes Algorithm 1 TOPDOWN (Mt, t, 0, B)
Open Source Code Yes Implementation public at: https://github.com/alexandersoen/alpha-tree-fair-wrappers
Open Datasets Yes To evaluate TOPDOWN2, we consider three datasets presenting a range of different size / feature types, Bank and German Credit (preprocessed by AIF360 [4]) and the American Community Survey (ACS) dataset preprocessed by Folktables3 [9]. ... We use public datasets.
Dataset Splits Yes Data is split into 3 subsets for black-box training, post-processing training, and testing; consisting of 40:40:20 splits in 5 fold cross validation.
Hardware Specification No The paper mentions that hardware details are in Appendix XI in the ethics checklist, but these details are not provided in the main text of the paper. No specific GPU, CPU models or detailed cloud resources are mentioned.
Software Dependencies No For the black-box, we consider a clipped (Assumption 1 with B = 1) random forest (RF) from scikit-learn calibrated using Platt s method [23]. No specific version numbers for scikit-learn or other software dependencies are provided.
Experiment Setup Yes The RF consists of an ensemble of 50 decision trees with a maximum depth of 4 and a random selection of 10% of the training samples per decision tree. For these experiments, we consider age as a binary sensitive attribute with a bin split at 25... For each of these TOPDOWN configurations, we boost for 32 iterations.