Tuesday, October 4, 2016

Why might machine learning be unfair?

Hear Aaron Roth speak on this at Penn Law School, starting around minute 7:30 (you can control the video from under where the slides appear, and you can also speed it up--1.5x is still quite intelligible):

What is Machine Learning? And Why Might it be Unfair? at the Optimizing Government Workshop

Here's a slide I liked from minute 35, about why a simple classifier rule might be a better judge of a majority population than of a minority population, simply because there are a lot more data points for the majority. (+'s are people who paid back their loans, -'s did not, the trick is to predict who will pay based on observables, in this case number of credit cards and SAT scores. The orange population is in fact more credit worthy, but has overall lower SAT scores. If you can only use one classifier, the best one is the blue line: and it denies credit to all the orange folks. If you could take group membership into account and use two lines, you could also distinguish the credit worthy oranges...)


No comments: