Recommendation by examples

Recommender systems usually rely on user profiles to generate personalized recommendations. We argue here that such profiles are often too coarse to capture the current user's state of mind/desire. For example, a serious user that usually prefers documentary features may, at the end of a long and tiring conference, be in the mood for a lighter entertaining movie, not captured by her usual profile. As communicating one's state of mind to a system in (key)words may be difficult, we propose in this work an alternative method which allows users to describe their current desire/mood through examples. Our algorithms utilizes the user's examples to refine the recommendations generated by a given system, considering several, possibly competing, desired properties of the recommended items set (rating, similarity, diversity, coverage). The algorithms are based on a simple geometric representation of the example items, which allows for efficient processing and the generation of suitable recommendations even in the absence of semantic information.