Philip Tovstogan: Facilitating Interactive Music Exploration
Music recommendation systems are an integral part of modern music streaming services. Usually, they utilize exploit vs explore model from game theory. There is a lot of work being done on music exploitation, but not enough on exploration. We address the issue of strict categorization of music according to classes (genres, moods, themes) by utilizing the latent space that is learned by deep music auto-tagging systems. We introduce the web interface to directly explore latent tag and embedding spaces. We conduct user experiments to understand the semantics of the dimensions of the latent spaces. We also introduce the assistant system that uses reinforcement learning to learn the user's context and goal as well as personalize the process to provide the most relevant exploration directions at the time of interaction. The evaluation of the system is performed based on the experiments involving user interactions with the system with the metrics based on user engagement and novelty.
Methods for Supporting Electronic Music Production with Large-Scale Sound Databases