Resource-constrained Fairness

Machine learning models are used to make decisions in high-impact areas of our lives such as finance, justice, and healthcare [Mehrabi et al., 2021]. Fair machine learning has emerged in response to the notion that simply making maximally accurate decisions is not enough and that training high-performance classifiers can result in both the transfer of existing biases from data to new decisions, as well as the introduction of new biases [Wachter et al., 2020]. Many studies that focus on improving fairness in machine learning overlook the practical limitations under which these models operate. For example, scenarios including university admissions, healthcare provision, and corporate hiring, are normally constrained by finite resources. Universities have a restricted quota of students to admit annually, healthcare facilities are bounded by available space and staff, and companies have a limited number of positions to fill.