Many prediction domains, such as ad placement, recommendation, trajectory
prediction, and document summarization, require predicting a set or list of
options. Such lists are often evaluated using submodular reward functions that
measure both quality and diversity. We propose a simple,