Music Recommendation by Modeling User’s Preferred Perspectives of Content, Singer/Genre and Popularity

As the amount, availability and use of online music increase, music recommendation becomes an important field of research. Collaborative, content-based and case-based recommendation systems and their hybrids have been used for music recommendation. There are already a number of online music recommendation systems. Although specific user information, such as, demographic data, education and origin have been shown to affect music preferences, they are usually not collected by the online music recommendation systems, because users would not like to disclose their personal data. Therefore, user models mostly contain information about which music pieces a user liked and which ones s/he did not and when. We introduce two music recommendation algorithms that take into account music content, singer/genre and popularity information. In the entropy-based recommendation algorithm, we decide on the relevant set of content features (perspective) according to which the songs selected by the user can be clustered as compactly as possible. As a compactness measure, we use entropy of the distribution of songs a user listened to in the clustering. The entropy-based recommendation approach enables both a dynamic user model and ability to consider a different subset of features appropriate for the specific user. In order to improve the performance of this system further, we introduce the content, singer/genre and popularity learning algorithm. In this algorithm, we first evaluate the extent to which content, singer/genre or popularity components could produce successful recommendations on the past songs listened to by the user. The number of songs in the final recommendation list contributed according to each component is chosen according to the recommendation success of each component. We perform experiments on user session data from a mobile operator. There are 2000 to 500 sessions and of length 5 to 15 songs. Our experiments indicate that the entropy-based recommendation algorithm performs better than simple content-based recommendation. Content, Music Recommendation by Modeling User’s Preferred Perspectives of Content, Singer/Genre and Popularity Zehra Cataltepe and Berna Altinel Istanbul Technical University Computer Engineering Department Ayazaga Campus, Maslak, Sariyer, Istanbul, 34469, Turkey Final Version, July 12, 2008 Book Chapter for the Book Collaborative and Social Information Retrieval and Access: Techniques for Improved User Modeling