Student Project

Show Me What You Got: Predicting Music Popularity

Abstract: There are multiple factors that affect a song’s popularity that can be both related and unrelated to a song’s musicality, from the key and modality in which it’s written, to the name of the artist, year of release, and type of marketing campaign. We attempted to build machine learning models to predict a song’s popularity based only on its musicality features. Our model included custom and pre-existing features of a song such as its key, modality, loudness, dissonance, and dynamics variation among others. Our model was trained using available song files from the Free Music Archive and tested on 60 sample song clips from Spotify. The test dataset consisted of 3 different genres with 20 popular and 20 unpopular songs within each genre. We defined popularity as the first principal component of track listens, interests and user favourites. Our model found that popularity was best predicted in the Hip-Hop and Jazz genres with features such as “speechiness”, “valence” and “instrumentalness” being the most effective.  The genre that we were least able to predict popularity for was Pop. The most effective overall musical features to determine popularity across all genres were “speechiness”, “acousticness” and “dissonance”.

Last updated:

December 12, 2017