Abstract
Music streaming services rely on music recommendation systems to keep users engaged and shape their musical taste. These systems rely on a combination of user and item modeling, and are adept at serving relevant recommendations to users through the analysis of collected data. Streaming services must now focus on combating user feelings of stagnation and listening fatigue associated with not receiving exciting and unique recommendations. This thesis proposes that incorporating elements of groove into a music recommendation system’s features can produce higher quality and more surprising recommendations by being genre agnostic while still recommending tracks based on one of the most important characteristics of music. To accomplish this, a beat tracking and onset detection system was used to analyze two varieties of percussive source separated audio to quantify features of groove. These features were then used to sort items into clusters, which were tested in evaluation sessions to determine if groove could influence quality or expectedness of recommendations. While the clusters had little effect on quality of recommendations, participants were consistently reporting items as unexpected and high quality, showing that recommending items based on features of groove could be useful in producing more serendipitous recommendations.