- Bias Mitigation Methods for Binary Classification Decision Systems: Research and Recommendations (arXiv)
Authors: Madeleine Waller, Odinaldo Rodriguez, Oana Kokarasuk
Abstract: Given the increasing importance of designing fair machine learning processes that are fair and do not discriminate against individuals or groups on the basis of protected personal traits, methods for debiasing binary classification decision-making systems. has been extensively studied. In this paper, we systematically review the research status of debiasing methods, report on their advantages and limitations, and provide recommendations for the development of future debiasing methods for binary classification.
2. Binary Classification Surrogate Risk Adversarial Consistency (arXiv)
Authors: Natalie Frank, Jonathan Niles Weed
Abstract: We study the consistency of surrogate risks for robust binary classification. It is common to learn robust classifiers by adversarial training. This is intended to minimize the expected 0-1 loss when each example can be maliciously corrupted within a small ball. A simple set of surrogate loss functions that are \emph{consistent}, i.e., for any data distribution, can replace losses between 0 and 1 without affecting the original adversarial risk minimization sequence. and complete characterization. We also prove a quantitative version of adversarial consistency for ρ margin loss. Our results revealed that the class of adversarially consistent surrogates is significantly smaller than the standard setting with which many common surrogates are known to be consistent.
