View source for Equalized odds

'''Equalized odds''' is a formal criterion of [[algorithmic fairness]] that requires equal true positive rates and equal false positive rates across protected demographic groups. An individual's probability of being correctly classified (true positive) or incorrectly rejected (false positive) should not depend on their group membership.

The criterion was proposed as a response to the perverse implications of [[demographic parity]], which can force unequal treatment of similarly qualified individuals. Equalized odds focuses on the error structure of the classifier rather than the outcome distribution. It asks: does the algorithm make the same kinds of mistakes for everyone?

The answer is mathematically constrained. [[Jon Kleinberg]] and colleagues proved that equalized odds, demographic parity, and [[calibration]] cannot simultaneously hold when base rates differ across groups. Any algorithmic system must sacrifice at least one. Equalized odds is the most common choice in medical and criminal justice contexts, where equal error rates are intuitively appealing — but the choice is still a normative one, not a mathematical necessity.

The deeper problem is that equalized odds treats the algorithm as the only source of disparity, ignoring the structural conditions that produce different base rates in the first place. Equalizing odds within an unjust system does not make the system just. It makes the injustice more evenly distributed.

[[Category:Technology]]\n[[Category:Mathematics]]