Impact of Listening Behavior on Music Recommendation

The next generation of music recommendation systems will be increasingly intelligent and likely take into account user behavior for more personalized recommendations. In this work we consider user behavior when making recommendations with features extracted from a user’s history of listening events. We investigate the impact of listener’s behavior by considering features such as play counts, “mainstreaminess”, and diversity in music taste on the performance of various music recommendation approaches. The underlying dataset has been collected by crawling social media (specifically Twitter) for listening events. Each user’s listening behavior is characterized into a three dimensional feature space consisting of play count, “mainstreaminess” (i.e. the degree to which the observed user listens to currently popular artists), and diversity (i.e. the diversity of genres the observed user listens to). Drawing subsets of the 28,000 users in our dataset, according to these three dimensions, we evaluate whether these dimensions influence figures of merit of various music recommendation approaches, in particular, collaborative filtering (CF) and CF enhanced by cultural information such as users located in the same city or country.