History-Based Collaborative Filtering for Music Recommendation

In this thesis we present history-based collaborative filtering, a novel approach to recommend unfamiliar music to users which them is nevertheless going to suit, in order to broaden theirs horizon of musical familiarity. We refer to people with similar music taste which experienced musical transitions when generating a new recommendation, and show that the results are accurate in terms of musical direction. Our works builds upon a previously developed map of music, a space of songs with approximately 430’000 vertices. We derive the musical transitions from extracted listening data by analysing way points in the map of music, and by explicitly looking for transitions in users’ listening behaviour. The profiles with most significant musical taste transitions are used as references for producing new recommendations. We evaluate our data set with two evaluation measures.