Learning Community-Based Preferences via Dirichlet Process Mixtures of Gaussian Processes

Bayesian approaches to preference learning using Gaussian Processes(GPs) are attractive due to their ability to explicitly modeluncertainty in users' latent utility functions; unfortunately existingtechniques have cubic time complexity in the number of users, whichrenders this approach intractable for collaborative preferencelearning over a large user base. Exploiting the observation that userpopulations often decompose into communities of shared preferences, wemodel user preferences as an infinite Dirichlet Process (DP) mixtureof communities and learn (a) the expected number of preferencecommunities represented in the data, (b) a GP-based preference model over items tailored to each community, and(c) the mixture weights representing each user's fraction of communitymembership. This results in a learning and inference process thatscales linearly in the number of users rather than cubicly andadditionally provides the ability to analyze individual community preferences and their associated members. We evaluate our approach ona variety of preference data sources including Amazon Mechanical Turkshowing that our method is more scalable and as accurate as previous GP-based preference learning work.