"Artist-on-the-Map"

Abstract

Artists have their own way of expressing their own songs so that when we think of an artist we can understand which music the artist mainly performs and distinguish it from other artists. The fact that artists have their own unique characteristics means that artist can be used as a categorical label to classify songs. This artist label is more specific than genre which is often too broad or in its scope and definition, whereas it is more inclusive than a single song or album. Therefore, many music streaming services provide artist-based music recommendation and search. However, the interface is still based on a simple list of selected artists and their songs that exposes very limited information.

There have been many visualization methods to effectively show a large number of artist information at a time. They are usually based on similarity among artists and displayed them in a two-dimensional or three-dimensional space. When there are too many artists, the number of artists on the space should be summarized due to the spatial limitation. Many of previous methods filtered them using musical elements. However, this was not intuitive for users who do not have much musical knowledge. On the other hand, popularity is one of the information that an artist has, and it does not require any musical knowledge in order to understand it. It is possible to find an artist who can be a representative in similar artists through popularity. Therefore, popularity can be used as a basis for displaying artist information.

In this thesis, we propose a visualization system using artist embedding vectors and popularity for artist search. We extend the study of learning audio features using artist labels and attempt to implement an intuitive search method for artist data by utilizing popularity data.