Fast and Efficient MMD-Based Fair PCA via Optimization over Stiefel Manifold

Authors: Junghyun Lee, Gwangsu Kim, Mahbod Olfat, Mark Hasegawa-Johnson, Chang D. Yoo7363-7371

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental comparisons based on synthetic and UCI datasets show that our approach outperforms prior work in explained variance, fairness, and runtime.
Researcher Affiliation Collaboration Junghyun Lee1, Gwangsu Kim*2, Mahbod Olfat3,4, Mark Hasegawa-Johnson5, Chang D. Yoo*2 1 Kim Jaechul Graduate School of AI, KAIST, Seoul, Republic of Korea 2 School of Electrical Engineering, KAIST, Daejeon, Republic of Korea 3 UC Berkeley IEOR, Berkeley, CA, USA 4 Citadel, Chicago, IL, USA 5 Department of Electrical and Computer Engineering, University of Illinois Urbana-Champaign, IL, USA
Pseudocode Yes Algorithm 1: REPMS for Mb F-PCA
Open Source Code Yes Codes are available in our Github repository6. https://github.com/nick-jhlee/fair-manifold-pca
Open Datasets Yes COMPAS dataset (Kirchner et al. 2016), Adult income dataset, and German credit dataset. See Section I of the SP for complete description of the pre-processing steps. For both algorithms, we consider two different hyperparameter settings
Dataset Splits Yes For all experiments, we considered 10 different 70/30 train-test splits.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used for its experiments.
Software Dependencies Yes For FPCA, we use the same Python MOSEK(Ap S 2021) implementation as provided by (Olfat and Aswani 2019). (Reference [Ap S 2021] is 'MOSEK Optimizer API for Python. Version 9.2.36. MOSEK.')
Experiment Setup Yes We’ve set K = 100, ϵmin = 10-6, ϵ0 = 10-1, θϵ = (ϵmin/ϵ0)1/5, ρmax = 1010, θρ = 2, dmin = 10-6.