Reviews: Paradoxes in Fair Machine Learning

Originality: somehow high since those concepts have not been analyzed before together. Quality: The claims are correct, they formalize somewhat a limitation of the equalized odds (it cannot provide partition independence, it can behave strangely as new points are added). Clarity: The paper is clear, with the exception that the motivation or implications (beyond theoretical curiosity) might be simply hard to comprehend. Significance: On the one hand, it's good to clarify but one could argue that the premise on testing equalized odds (or equal opportunity, or statistical parity) on those axioms was doomed to begin with. Partition independence (refered to as consistency here) states that you could divide people and run the algorithm among those newly formed groups separately and come to the same conclusion. But, precisely, equalized odds is meant to be sensitive to the composition of the applicants pool.