Individually Fair Gradient Boosting

Authors: Alexander Vargo, Fan Zhang, Mikhail Yurochkin, Yuekai Sun

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also demonstrate the efficacy of our algorithm on three ML problems susceptible to algorithmic bias.
Researcher Affiliation Collaboration Alexander Vargo Department of Mathematics University of Michigan ahsvargo@umich.edu Fan Zhang School of Information Science and Technology Shanghai Tech University zhangfan4@shanghaitech.edu.cn Mikhail Yurochkin IBM Research MIT-IBM Watson AI Lab mikhail.yurochkin@ibm.com Yuekai Sun Department of Statistics University of Michigan yuekai@umich.edu
Pseudocode Yes Algorithm 1 Fair gradient boosting
Open Source Code No No explicit statement of releasing code or a link to a code repository for the described methodology was found.
Open Datasets Yes The German credit data set (Dua & Graff, 2017) contains information from 1000 individuals; the ML task is to label the individuals as good or bad credit risks. The Adult data set (Dua & Graff, 2017) is another common benchmark in the fairness literature. We study the COMPAS recidivism prediction data set (Larson et al., 2016).
Dataset Splits Yes For the baseline GBDT and projecting, we select hyperparameters by splitting 20% of the training data into a validation set and evaluating the performance on the validation set.
Hardware Specification No Section C.6 mentions the number of CPUs and GPUs (e.g., "4 CPUs", "1 GPU") used for experiments but does not provide specific models, types, or memory details of the hardware.
Software Dependencies Yes We use the default parameters in the Ridge CV class from the scikit-learn package, version 0.21.3 (Pedregosa et al., 2011). We implement the adversarial debiasing methods (in the adversarial_debiasing class from the IBM’s AIF360 package, version 0.2.2 (Bellamy et al., 2018)).
Experiment Setup Yes Table 4: Optimal XGBoost parameters for German credit data set. For Bu DRO, we also used a pertubation budget of ϵ = 1.0. Table 6: Optimal XGBoost parameters for the Adult data set. Table 8: Optimal XGBoost parameters for COMPAS data set.