Beyond Adult and COMPAS: Fair Multi-Class Prediction via Information Projection

Authors: Wael Alghamdi, Hsiang Hsu, Haewon Jeong, Hao Wang, Peter Michalak, Shahab Asoodeh, Flavio Calmon

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

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive numerical comparisons with state-of-the-art benchmarks demonstrate that our approach maintains competitive performance in terms of accuracy-fairness trade-off curves, while achieving favorable runtime on large datasets.
Researcher Affiliation Academia 1John A. Paulson School of Engineering and Applied Sciences, Harvard University 2Department of Computing and Software, Mc Master University
Pseudocode Yes Algorithm 1 : Fair Projection for solving (8).
Open Source Code Yes The data can be found at [INE20], and code for pre-processing the data and the implementation of Fair Projection can be found at https://github.com/Hsiang Hsu/Fair-Projection.
Open Datasets Yes We also evaluate Fair Projection on a dataset derived from open and anonymized data from Brazil s national high school exam the Exame Nacional do Ensino Médio (ENEM) with over 1 million samples. ... The data can be found at [INE20], and code for pre-processing the data and the implementation of Fair Projection can be found at https://github.com/Hsiang Hsu/Fair-Projection.
Dataset Splits Yes All values reported in this section are from the test set with 70/30 train-test split.
Hardware Specification Yes These experiments were run on a machine with AMD Ryzen 2990WX 64-thread 32-Core CPU and NVIDIA TITAN Xp 12-GB GPU.
Software Dependencies No The paper mentions using 'Scikit-learn' and 'Tensor Flow' for implementation, but does not specify any version numbers for these or other software components.
Experiment Setup Yes For Fair Projection (the constrained optimization in (6)), we use cross-entropy (Fair Projection-CE) and KL-divergence (Fair Projection-KL) as the loss function8... For consistency, we used the same fairness metric (MEO, α = 0.01), base classifier (GBM), and train/test split, and each number is the average of 2 repeated experiments.