Learning and visualizing user preferences over sets

In previous work, we introduced a representation language, DD-PREF, that balances preferences for particular items with preferences about the properties of the set. Specifically, DD-PREF permits the expression of preferences for depth (i.e., preferences for specific attribute values over others) and diversity of sets (i.e., preferences for broad vs. narrow distributions of attribute values). We have also shown how preferences represented in DD-PREF can be learned from training data. In this paper, we present a new experimental methodology, and give results in a Mars rover image domain. We also provide new visualizations of the learned preferences in this domain. Finally, we describe a Chinese restaurant menu domain for which we are currently gathering user data.