Hybrid music recommendation based on different dimensions of audio content and an entropy measure

Our music recommendation system recommends a song to a user, at a certain time, based on the listening history of the user. Based on different sets of audio features (MFCC, MPITCH, BEAT, STFT) of all available songs, different clusterings of songs are obtained. Users are given recommendations from one of these clusterings. The right clustering for a user is determined based on the Shannon entropy of the distribution of songs the user listened in each clustering. Using this content based recommendation scheme, as opposed to a static set of features resulted in upto 60 percent increase in recommendation success. In addition to the audio features (content) of songs user listened, the singers for the songs and also the most popular songs at the time of recommendation are also available. We introduce two recommendation algorithms that decide on the weight of content cluster, singer cluster and popularity adaptively for each user, based on the user history. Our experiments on user session data consisting of 2000 to 500 sessions and of length 5 to 15 indicate that these adaptive recommendation schemes give better recommendation results than using only content based recommendation.