A Stochastic Optimization Framework for Fair Risk Minimization

Authors: Andrew Lowy, Sina Baharlouei, Rakesh Pavan, Meisam Razaviyayn, Ahmad Beirami

TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that FERMI achieves the most favorable tradeoffs between fairness violation and test accuracy across all tested setups compared with state-of-the-art baselines for demographic parity, equalized odds, equal opportunity. These benefits are especially significant with small batch sizes and for non-binary classification with large number of sensitive attributes, making FERMI a practical, scalable fairness algorithm. The code for all of the experiments in this paper is available at: https://github.com/optimization-for-data-driven-science/FERMI.
Researcher Affiliation Collaboration Andrew Lowy EMAIL Department of Mathematics University of Southern California Sina Baharlouei EMAIL Department of Industrial & Systems Engineering University of Southern California Rakesh Pavan EMAIL Department of Computer Science University of Washington Meisam Razaviyayn EMAIL Department of Industrial & Systems Engineering, Computer Science, and Electrical Engineering University of Southern California Ahmad Beirami EMAIL Google Research
Pseudocode Yes Algorithm 1 FERMI Algorithm 1: Input: θ0 Rdθ, W 0 = 0 Rk m, step-sizes (ηθ, ηw), fairness parameter λ 0, iteration number T, minibatch sizes |Bt|, t {0, 1, , T}, W := Frobenius norm ball of radius D around 0 Rk m for D given in Appendix D. 2: Compute b P 1/2 s = diag(ˆp S(1) 1/2, . . . , ˆp S(k) 1/2). 3: for t = 0, 1, . . . , T do 4: Draw a mini-batch Bt of data points {(xi, si, yi)}i Bt 5: Set θt+1 θt ηθ |Bt| P i Bt[ θℓ(xi, yi; θt) + λ θ bψi(θt, W t)]. 6: Set W t+1 ΠW W t + 2ληw i Bt h W t E[by(xi, θ)by(xi, θ)T |xi] + b P 1/2 s E[siby T (xi; θt)|xi, si] i 7: end for 8: Pick ˆt uniformly at random from {1, . . . , T}. 9: Return: θˆt.
Open Source Code Yes The code for all of the experiments in this paper is available at: https://github.com/optimization-for-data-driven-science/FERMI.
Open Datasets Yes In Section 3.1, we showcase the performance of FERMI applied to a logistic regression model on binary classification tasks with binary sensitive attributes on Adult, German Credit, and COMPAS datasets. All of the following datasets are publicly available at UCI repository.
Dataset Splits Yes We generate ten distinct train/test sets for each one of the German and COMPAS datasets by randomly sampling 80% of data points as the training data and the rest 20% as the test data. Adult dataset contains the census information of individuals including education, gender, and capital gain. The assigned classification task is to predict whether a person earns over 50k annually. The train and test sets are two separated files consisting of 32, 000 and 16, 000 samples respectively.
Hardware Specification No No specific hardware details (GPU/CPU models, processor types, or memory amounts) are mentioned in the paper for running experiments. It mentions using 'neural network function approximation' and 'large-scale problems' which implies powerful hardware, but no specifics.
Software Dependencies No The paper mentions implementing models like logistic regression and convolutional neural networks, which implies the use of programming languages and libraries (e.g., Python, PyTorch/TensorFlow), but specific versions for these are not provided.
Experiment Setup Yes The trade-offcurves for FERMI are generated by sweeping across different values for λ [0, 10000]. The learning rates ηθ, ηw is constant during the optimization process and is chosen from the interval [0.0005, 0.01] for all datasets. Moreover, the number of iterations T for experiments in Fig. 1 is fixed to 2000. ... We use a mini-batch of size 512 when using the stochastic solver.