Learning Antidote Data to Individual Unfairness

Authors: Peizhao Li, Ethan Xia, Hongfu Liu

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through extensive experiments on multiple tabular datasets, we demonstrate our method resists individual unfairness at a minimal or zero cost to predictive utility compared to baselines.
Researcher Affiliation Academia 1Brandeis University 2Cornell University.
Pseudocode Yes Algorithm 1 Anti DRO: DRO with Antidote Data for Individual Fairness
Open Source Code Yes Code available at https://github.com/brandeismachine-learning/Anti Indiv Fairness.
Open Datasets Yes Datasets We involve census datasets Adult (Kohavi & Becker, 1996) and Dutch (Van der Laan, 2000), educational dataset Law School (Wightman, 1998) and Oulad (Kuzilek et al., 2017), and criminological dataset Compas (Angwin et al., 2016) in our experiments.
Dataset Splits Yes Table 4. Dataset Statistics. We report data statistic including sample size, feature dimension, sensitive attribute, as well as the number of positive and negative comparable samples in training / testing set, respectively. ... Adult 30,162 / 15,060 ... Compas 4,626 / 1,541 ... Law School 15,598 / 5,200 ... Oulad 16,177 / 5,385 ... Dutch 45,315 / 6,556
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only mentions using 'neural networks' and 'logistic regression' models.
Software Dependencies No The paper mentions 'Scikit-learn' (Pedregosa et al., 2011), 'Py Torch', and 'XGBClassifier from https://xgboost.readthedocs.io/en/stable/' but does not specify version numbers for these software components.
Experiment Setup Yes We use Adam optimizer for training the generator. We set the learning rate for generator gθ to 2e-4, for discriminator dθ to 2e-4, weight decay for gθ to 1e-6, for dθ to 0. We set batch size to 4096 and training epochs to 500. ... For logistic regression, we set the strength of ℓ2 penalty to 1, and max iteration to 2.048. For neural networks, we set optimization iterations to 10,000, initial learning rate to 1e-1, ℓ2 penalty strength to 1e-2, with SGD optimizer and decrease learning rate by 50% for every 2,500 iterations.