Abstract
The research and development of recommender systems is dominated by models of user's preferences learned from ratings for items. However, ratings have several disadvantages, which we discuss, and in order to address these issues we analyse another way to articulate preferences, i.e., as pairwise comparisons: item A is preferred to item B. We have developed a recommendation technology that, combining ratings and pairwise preferences, can generate better recommendations than a state of the art solution uniquely based on ratings.