Method to Control Embedded Representation of Piece of Music in Playlists

This paper proposes a new method to control embedded representation of a piece of music provided as a set of playlists. To recommend an appropriate piece to a user, numerical representation of music has been studied so far. This study does not focus on explicit representation like signal data, rather implicit representation called embeddings obtained from users’ playlists in order to avoid issues around copyright. In the previous work, naive method was proposed and raw dataset is provided to learn the model of embedding for pieces of music, however, it is still not clear the raw dataset is appropriate for the model of music recommender system. Actually, this paper shows there is bias in a raw dataset and it makes the representation tends to provide trivial result, i.e., clearly same element that can be implied only by shallow knowledge like that pieces of music composed by the same artist are similar each other. This study shows the problem quantitatively and proposes a new method to reduce self-evident feature from embedded representation.