Abstract

Disparate impact of machine learning algorithms is often caused by imbalanced datasets, in particular, the scarcity of data from certain subpopulations. One obvious solution to this problem is to collect more data - and one natural way to do this is to continue collecting data after an algorithm is deployed. In this sense, the use of exploration in an online learning framework is a natural way to "collect more data" over time to address this problem. I will talk about some recent work on fairness in exploration, as well as some open directions in this setting.

Video Recording