Understanding Instance-Level Impact of Fairness Constraints

Authors: Jialu Wang, Xin Eric Wang, Yang Liu

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate with extensive experiments that training on a subset of weighty data examples leads to lower fairness violations with a trade-off of accuracy. (Abstract) and In this section, we examine the influence score subject to parity constraints on three different application domains: tabular data, images and natural language. (Section 6)
Researcher Affiliation Academia Department of Computer Science and Engineering, University of California, Santa Cruz, CA, USA.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We publish the source code at https://github.com/UCSC-REAL/Fair Infl.
Open Datasets Yes Firstly, we work with multi-layer perceptron (MLP) trained on the Adult dataset (Dua & Graff, 2017)., Next, we train a Res Net-18 network (He et al., 2015) on the Celeb A face attribute dataset (Liu et al., 2015)., Lastly, we consider Jigsaw Comment Toxicity Classification (Jigsaw, 2018) with text data.
Dataset Splits No The paper describes training and test splits for the datasets but does not explicitly mention a validation dataset split ratio or strategy.
Hardware Specification Yes For all the experiments, we use a GPU cluster with four NVIDIA RTX A6000 GPUs for training and evaluation.
Software Dependencies No The paper mentions optimizers (Adam) and pre-trained models (BERT) but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes We used the Adam optimizer with a learning rate of 0.001 to train all the models. We used γ = 1 for models requiring the regularizer parameter of fairness constraints. (Section 6.1) and The MLP model is a two-layer Re LU network with hidden size 64. (Section 6.2)