Re-weighting Based Group Fairness Regularization via Classwise Robust Optimization

Authors: Sangwon Jung, Taeeon Park, Sanghyuk Chun, Taesup Moon

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that Fair DRO is scalable and easily adaptable to diverse applications, and consistently achieves the state-of-the-art performance on several benchmark datasets in terms of the accuracy-fairness trade-off, compared to recent strong baselines.
Researcher Affiliation Collaboration 1 Department of Electrical and Computer Engineering, Seoul National University 2 NAVER AI Lab 3 ASRI/INMC/IPAI/AIIS, Seoul National University
Pseudocode Yes Algorithm 1: Fair DRO Iterative Optimization
Open Source Code No The paper does not provide an explicit statement or link for open-source code availability for the methodology described.
Open Datasets Yes Two tabular datasets, UCI Adult (Dua et al., 2017) (Adult) and Pro Publica COMPAS (Julia Angwin & Kirchner, 2016) (COMPAS), are used for the benchmark. We also evaluate Fair DRO on UTKFace (Zhang et al., 2017), a face dataset with multi-class and multi-group labels. Civil Comments-WILDS (Koh et al., 2021). The description and results on another vision dataset, Celeb A (Liu et al., 2015), are given in Appendix D.2.
Dataset Splits No The paper mentions evaluating models on "separate test sets" but does not explicitly provide details about training/validation/test dataset splits, percentages, or specific counts for a validation set.
Hardware Specification Yes Experiments are performed on a server with AMD Ryzen Threadripper PRO 3975WX CPUs and NVIDIA RTX A5000 GPUs.
Software Dependencies No The paper states, "We used Py Torch (Paszke et al., 2019)", but it does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes For tabular and vision datasets, we train all models with the Adam W optimizer (Loshchilov & Hutter, 2019) for 70 epochs. We set the mini-batch size and the weight decay as 128 and 0.001, respectively. The initial learning rate is set as 0.001 and decayed by cosine annealing technique (Loshchilov & Hutter, 2017). For the language dataset, we fine-tune pre-trained BERT with the Adam W optimizer for 3 epochs. We set the mini-batch size and the weight decay as 24 and 0.001, respectively. The initial learning rate is set as 0.00002 and adjusted with a learning rate schedule using a warm-up phase followed by a linear decay.