Fair Data Adaptation with Quantile Preservation
Authors: Drago Plečko, Nicolai Meinshausen
JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We describe an implementation of the proposed data adaptation procedure based on Random Forests (Breiman, 2001) and demonstrate its practical use on simulated and real-world data. Keywords: Supervised learning, Fairness in machine learning, Causality, Graphical models, Counterfactual fairness (...) In Section 6 empirical performance of our method is demonstrated both on simulated and real-world data (namely the COMPAS and Adult datasets). |
| Researcher Affiliation | Academia | Drago Plecko EMAIL Nicolai Meinshausen EMAIL Seminar fur Statistik ETH Zurich Zurich, 8092, Switzerland |
| Pseudocode | Yes | Algorithm 1: Population Fairness Adaptation (...) Algorithm 2: Fairness Adaptation |
| Open Source Code | Yes | The software is provided as an R package fairadapt. (...) A full implementation of this method for a general situation is available in the fairadapt package on CRAN. (...) We provide a PyTorch implementation of the PSCF method, which can be found in our Github repository. |
| Open Datasets | Yes | In Section 6 empirical performance of our method is demonstrated both on simulated and real-world data (namely the COMPAS and Adult datasets). (...) UCI Adult. The Adult dataset from the UCI machine learning repository (Lichman et al., 2013) (...) COMPAS dataset. The second real dataset we analyze is the COMPAS dataset (Larson et al., 2016) |
| Dataset Splits | Yes | We run our method ten times, with 5000 training and test samples generated from the given SCMs. (...) For both UCI Adult and COMPAS, we split the dataset into 75% training and 25% testing randomly 20 times. |
| Hardware Specification | Yes | The runtime of fairadapt on a single 2.8GHz CPU is 23 seconds, compared to 396 seconds for a single value of β for PSCF (excluding any hyperparameter search). |
| Software Dependencies | Yes | Roger Koenker. quantreg: Quantile regression. r package version 5.05. R Foundation for Statistical Computing: Vienna) Available at: http://CRAN. R-project. org/package= quantreg, 2013. |
| Experiment Setup | Yes | A logistic regression classifier is used after applying fairadapt. (...) We run our method ten times, with 5000 training and test samples generated from the given SCMs. (...) the reductions approach (...) we again use logistic regression for our classifier that allows sample-weighting and vary the fairness constraint violation parameter ϵ {0.1, 0.01, 0.001} |