Predicting User Preferences Via Similarity-Based Clustering

This paper explores the idea of clustering partial preference relations as a means for agent prediction of users' preferences. Due to the high number of possible outcomes in a typical scenario, such as an automated negotiation session, elicitation techniques can provide only a sparse specification of a user's preferences. By clustering similar users together, we exploit the notion that people with common preferences over a given set of outcomes will likely have common interests over other outcomes. New preferences for a user can thus be predicted with a high degree of confidence by examining preferences of other users in the same cluster. Experiments on the MovieLens dataset show that preferences can be predicted independently with 70-80% accuracy. We also show how an error-correcting procedure can boost accuracy to as high as 98%.