15 December 2016: Comparing predictive model performance on reconviction data

Nikolaj discussed his 2014 paper “Which method predicts recidivism best?: a comparison of statistical, machine learning and data mining predictive models”, published in the Journal of the Royal Statistical Association, Series A. For those who missed Nikolaj’s interesting talk, he presented a similar talk for the RSS journal club, shown below.

The meeting discussed the use of accuracy measures such as AUC, ACC, and RMSE for risk assessment models. We also discussed the ethical problem of building algorithms that discriminate against certain groups, and how that might be prevented.

A relevant recent book on this topic is “Weapons of Math Destruction” by Cathy O’Neil, a former Wall Street mathematician who criticizes, among other things, the use of scoring algorithms to decide on paroles in the U.S. See https://weaponsofmathdestructionbook.com/

For those of you who understand German, an excellent discussion of this book and topic was given by Frauke Kreuter and Christof Horn in their data science podcast digdeep (also on iTunes): https://digdeepweb.wordpress.com/2016/10/23/folge-10-math-destruction/

[youtube https://www.youtube.com/watch?v=S6hVYgZmfuk?start=2148]