Скачать или смотреть Classification and Discrimination Part 4. Video 41

This is one in a series of videos produced by Professor Seth J. Chandler at the University of Houston Law Center for the course Analytic Methods for Lawyers.

This video, the last in a four-part series, series, builds on the understanding of classifiers and discrimination developed in the prior three videos to examine a very real world use of them: helping to make decisions about whether a prisoner should be released or not. The use of machine learning to make this sort of decision, both in theory and in implementation, has been highly controversial, with the respected investigative body Pro Publica charging that the algorithms discriminate against minorities. The creator of the "COMPAS" algorithm, Northpointe has vigorously denied the charge. This essay confirms that the algorithms behave in a "discriminatory" manner as charged by Pro Publica but that Northpointe's defense of that behavior has considerable force. Virtually all of the discrimination involved arises from the higher incidence of observed parole-breaking among African Americans and very little of it from the algorithm itself. Northpointe is no more guilty of invidious discrimination -- and no less -- than most of those developing machine learning models to work on groups with different incidences. The behavior noted by Pro Publica is generally math, not racism. The essay does suggest, however, that the costs of inaccuracy, which are particularly significant in this area, could be reduced by preserving use of a strong classifier but not insisting on the same thresholds for labeling a prisoner as likely to break parole or observe it. Alternatively, one could cure much of the "discrimination" created by classifiers operating on this data without hugely increasing the costs of inaccuracy by subjecting prisoners to race-specific thresholds.

#Classifiers
#Supervised machine learning
#Discrimination
#COMPAS
#Northpointe
#ConvexHull
#NMinimize
#ClassifierMeasurementsObject
#Inaccuracy costs
#Pro Publica