flâneur

Relative notions of fairness

fairmlbook.org · 14,986 words · saved by 1 readers

In Chapter 3, we considered a range of statistical criteria that help to highlight group-level differences in both the treatments and outcomes that might be brought about by the use of a machine learning model. But why should we be concerned with group-level differences? And how should we decide which groups we should be concerned with? In this chapter, we’ll explore the many different normative reasons we might have to object to group-level differences. This is a subtle, but important, shift in focus from Chapter 2, where we considered some of the normative reasons why individuals might object to decision-making schemes that distribute desirable resources or opportunities. In this chapter, we’ll focus on why we might be concerned with the uneven allocation of resources and opportunities across specific groups and society overall. In particular, we’ll review the normative foundations that ground claims of discrimination and calls for distributive justice. We’ll then try to connect thes

Home Systematic relative disadvantage Six accounts of the wrongfulness of discrimination Intentionality and indirect discrimination Equality of opportunity The narrow view The broad view The middle view Tensions between the different views Merit and desert The cost of fairness Connecting statistical and moral notions of fairness Demographic parity Calibration The similarity criterion Randomization, thresholding, and fairness The normative underpinnings of error rate parity Error rate parity and the middle view of equality of opportunity What to do about error rate disparity Alternatives for re

saved by

related reading