Fairness Guarantees under Demographic Shift

Authors: Stephen Giguere, Blossom Metevier, Bruno Castro da Silva, Yuriy Brun, Philip S. Thomas, Scott Niekum

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Shifty using the UCI Adult Census dataset (Kohavi and Becker, 1996), as well as a real-world dataset of university entrance exams and subsequent student success. We show that the learned models avoid bias under demographic shift, unlike existing methods. Our experiments demonstrate that our algorithm s high-confidence fairness guarantees are valid in practice and that our algorithm is an effective tool for training models that are fair when demographic shift occurs.
Researcher Affiliation Collaboration Stephen Giguere Department of Computer Science, University of Texas at Austin sgiguere@cs.utexas.edu Blossom Metevier, Yuriy Brun, Bruno Castro da Silva, & Philip S. Thomas College of Information and Computer Sciences, University of Massachusetts Scott Niekum Department of Computer Science, University of Texas at Austin
Pseudocode Yes An overview of Shifty is shown in Figure 1. Algorithm 1 presents high-level pseudocode for classification algorithms that provide high-confidence fairness guarantees under demographic shift.
Open Source Code Yes Our main contributions are: (4) an open-source Shifty implementation and a release of all our data. To support efforts to reproduce our results, all code and data used in this paper will be made publicly available upon publication.
Open Datasets Yes We evaluate Shifty using the UCI Adult Census dataset (Kohavi and Becker, 1996), as well as a real-world dataset of university entrance exams and subsequent student success. The UCI Adult Census dataset is available for download from the UCI Machine Learning Repository (Kohavi and Becker, 1996), and the UFRGS Entrance Exam and GPA dataset is available at https://doi.org/10.7910/DVN/O35FW8 (da Silva, 2019).
Dataset Splits Yes First, the data partitioning step splits the input dataset into two parts, which are used to perform the candidate selection and fairness test steps, respectively (see Appendix C.1). In our experiments, we split the input data evenly between Dc and Df, but we hypothesize that there may be more effective techniques for determining the optimal splitting proportion.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper mentions implementation details but does not specify software dependencies with version numbers (e.g., Python, PyTorch, specific libraries and their versions).
Experiment Setup Yes Additional implementation details are provided in Appendix C. The tolerances for each definition of fairness were set to ϵDP = 0.05, ϵDI = 0.8, ϵEOp = 0.15, ϵEOd = 0.2 and ϵPE = 0.025 for our experiments using the UCI Adult Census dataset, and ϵDP = 0.1, ϵDI = 0.8, ϵEOp = 0.05, ϵEOd = 0.1 and ϵPE = 0.05 for our experiments using the UFRGS GPA dataset.