Dynamic artist-based embeddings with application to playlist generation

Streaming services have become a central platform for music consuming. Often, they rely on large archives on which playlists are created. While many playlists are handcrafted, automatic playlist generation is a rising star, as it lets people explore new music easily. Playlist generation relays on some underlying song organization, which is regularly based on embedding methods. The reliability and flexibility of the constructed embedding reflects on the quality of the generated playlists. One type of embedding is content-based, which organizes the songs using raw audio features. Content-based embeddings typically result in a static representation that captures pairwise song similarity.This paper proposes a content-based embedding process that builds a dynamic, instead of a static embedding that is centered around an artist of interest. This embedding mutually organizes seed artist songs into several clusters as well as the rest of the songs in the dataset. The radio is dynamically generated using the distances in the embedded space, while also taking diversity into account. Moreover, it is scalabile in terms of computational complexity.Empirical results are demonstrated on a public dataset named AcousticBrainz, and evaluated by utilizing the Melon dataset, a novel public dataset for playlists. We show that the proposed dynamic embedding and radio generation algorithm yields a more accurate and diverse sequence when compared to other embedding techniques. The proposed methods may be utilized for other recommendation systems tasks that aim to offer the users a diverse recommendation list based on their initial choice.