Rich Caruana: The Unexpected Success of GlassBox Learning with Tabular Data

Conventional wisdom was that models simple enough to be interpretable must be less accurate. That is, glassbox models must be less accurate than blackbox models. Surprisingly, this does not appear to be true for tabular data. Moreover, models that are simple enough to be interpretable have other advantages as well: despite their transparency they are more likely to protect privacy, they are easier to correct when they make mistakes, and they make bias easier to detect and mitigate. In this talk we'll review a number of unexpected advances in glassbox learning made over the last 30+ years.