Content-based top-N recommendations with perceived similarity

Abstract

Similarity-based recommender systems can be used to pre-compute distance between item pairs, and then to quickly recommend similar items to users. The content-based approach to similarity uses the item's description, which in movies could mean genre, director or cast. These similarity methods are often built with unsupervised learning, which means the notion of similarity is defined by those who write the method. This notion may not match that of the users, resulting in poor user experience. In this paper we used user-contributed labels representing perceived similarity between movies to build a supervised content-based (CB) model for movie recommendations. Our user study shows that the CB method with human perception factored in was significantly preferred over the CB model without.