Output details

Ceteris Paribus Preference Elicitation with Predictive Guarantees

<22>Automating many tasks requires access to human preferences. Actively querying people for their preferences imposes an inordinate burden, yielding sparse replies and limited, biased information (if not defeating the goal of automation). We present the first practical learning algorithm for conditional preferences (CP-nets). Surprisingly, we demonstrate passive learnability of provably more complex preferences than the (simplest) “swap” preferences, previously only actively learnable (http://dx.doi.org/10.1016/j.artint.2010.04.019). Building on this, Michael and Papageorgiou (IJCAI'2013) and on-going work formalise efficient reasoning with learned preferences, and demonstrate the applicability of our learning algorithm on real-world data through the powerful and practical tools we have implemented.