Abstract
Many recommender systems rely on item ratings to predict users' preferences and generate recommendations. However, users often express preferences by referring to features of the items, e.g., "I like Tarantino's movies". But, it has been shown that user models based on feature preferences may lead to wrong recommendations. In this paper we cope with this issue and we introduce a novel prediction model that generate better item recommendations, especially in coldstart situations, by exploiting both item-based and featurebased preferences. We also show that it is possible to optimize the combination of the two types of preferences when actively requesting them to users.