FR-Train: A Mutual Information-Based Approach to Fair and Robust Training

Authors: Yuji Roh, Kangwook Lee, Steven Whang, Changho Suh

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide experimental results for FR-Train. For the fairness measure, we use disparate impact, while leaving in the supplementary the results for equalized odds and equal opportunity.
Researcher Affiliation Academia 1School of Electrical Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Korea 2Department of Electrical and Computer Engineering, University of Wisconsin Madison, Madison, Wisconsin, USA.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes 1https://github.com/yuji-roh/fr-train
Open Datasets Yes We use two real datasets: Pro Publica COMPAS (Angwin et al., 2016) and Adult Census (Kohavi, 1996), which have 7,214 and 45,222 examples, respectively.
Dataset Splits Yes To make a validation set, we randomly select clean examples that amount to 10% of the entire training data. For FR-Train and RML, the validation set is 10% of Dtr. We consider a scenario where one first constructs a small (which amounts to 5% of Dtr) validation set based on crowdsourcing
Hardware Specification Yes We use Py Torch (Paszke et al., 2017), and all experiments are performed on a server with Intel i7-6850 CPUs.
Software Dependencies No The paper mentions 'Py Torch (Paszke et al., 2017)' but does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes Here λ1 and λ2 are tuning knobs that play roles to emphasize fair and robust training, respectively. We compute the final example weights as W = R + D(X, Z, ˆY ) (1 R) where R = σ( Lc /Ld C) is a conversion of the loss ratio into a probability using the sigmoid function σ and hyperparameter C.