Title: Model Cards: From Exceptional Model Mining to Robust Machine Learning Systems
Abstract: It is necessary to know how a machine learning model would perform under different situations before deploying it in high-stake applications. In this talk, I would like to discuss this problem with a framework called Exceptional Model Mining. In this framework, all the work are built upon a core concept called subgroup. I will present different techniques on quantitatively evaluating the performance of models on subgroups, including the benefits of causal modelling and ensembles of subgroups to build robust machine learning systems.