Music Recommendations; Approximating user distributions to address the cold start problem

Abstract: In today's data driven society the world is at a point of information overload. As people rely on Google for information and other platforms such as Netflix and Spotify for entertainment, the need for relevant filtering of content has never been higher. As a result, recommendation systems have seen a great surge in demand. One can divide the space of recommendation algorithms into primarily two approaches. In the context of music,a collaborative based one where underlying correlations between users dictate the model, and the content based approach which examines the more specific relationship each user has to the songs. This paper aims to highlight the issues many collaborative models face when there is a lack uneven amount of interactions with the songs; this is usually the case for less popular or new items. To address this, a content based approach is suggested based on music feature data with the goal to distinguish unique user distributions based on song characteristics. After evaluating this method against a popularity based baseline model, there was a small but not significant difference in the error. This suggested that there are a lot of room for improvement in the approximation of user distributions, leading to the conclusion that with more elaborative methods one could most likely expand upon this research and build strong recommendations based on the idea of probabilistic user distributions.