The Many Faces of Users: Modeling Musical Preference
User models that capture the musical preferences of users are central for many tasks in music information retrieval and music recommendation, yet, it has not been fully explored and exploited. To this end, the musical preferences of users in the context of music recommender systems have mostly been captured in collaborative filtering-based approaches. Alternatively, users can be characterized by their average listening behavior and hence, by the mean values of a set of content descriptors of tracks the users listened to. However, a user may listen to highly different tracks and genres. Thus, computing the average of all tracks does not capture the user's listening behavior well. We argue that each user may have many different preferences that depend on contextual aspects (e.g., listening to classical music when working and hard rock when doing sports) and that user models should account for these different sets of preferences. In this paper, we provide a detailed analysis and evaluation of different user models that describe a user's musical preferences based on acoustic features of tracks the user has listened to.