Real-time Multiattribute Bayesian Preference Elicitation with Pairwise Comparison Queries
Shengbo Guo and Scott Sanner
In: NIPS 2009, 7-12 Dec 2009, Vancouver, Canada.
Preference elicitation (PE) is an important component of interactive decision support systems that aim to make optimal recommendations to users by actively querying their preferences. The PE task consists of (a) querying the user about their preferences and (b) recommending an item that maximizes the user’s latent utility. Of course, a PE system is limited by real-world performance constraints that require phase (a) to be efficient while ensuring phase (b) can make an optimal recommendation with high certainty. Bayesian approaches to PE  have received interest in recent years due to their robust handling of noise in the elicitation process, however, previous work has either relied on expensive sampling methods  or on expensive EM refitting of mixture models  to deal with the lack of a closed-form for the utility belief update. In this work, we propose to avoid both of these problems by adapting the Bayesian ranking approach of TrueSkill  to multiattribute Bayesian PE, which allows us to efficiently maintain and update the belief representation in real-time and naturally facilitates the efficient evaluation of value of information (VOI) heuristics for use in query selection strategies. Our best VOI query strategy is both space- and time-efficient (in contrast to related work) and performs on par with the most accurate (and often computationally intensive) algorithms on experiments with a real-world dataset.