A large in-situ dataset for context-aware music recommendation on smartphones

Context-based services have received an increasing attention due to the prevalence of sensor-rich mobile devices such as smartphones. The idea is to recommend information that would be of interest to a user according to the user's surround context. Although remarkable progress has been made, relatively little research has been made to contextualize music playback based on a large-scale dataset of real-life listening records. This paper presents our recent endeavor in collecting 5,502 real-life listening records with context annotation using Android smartphones in-situ. The user-provided context annotation contains labels selected from 10 user activity categories and 10 user mood categories. Moreover, we also compute a rich set of sensor features to capture the context at which the users listen to music, encompassing location, time, acceleration, proximity, etc. Our evaluation shows that with such context information we are able to significantly improve the performance of music recommendation, using factorization machine as the recommendation engine.