Music recommendation based on adaptive feature and user grouping

We present a music recommendation system that uses different features of audio content for each user based on the user´s listening history. The system is based on the idea that different people may give more importance to certain aspects of music. MFCC, MPITCH, BEAT, STFT feature sets are obtained for all the available songs and then different clusterings of the songs based on each possible feature set is obtained. When a user session is observed, the cluster IDs of songs the user listened in each clustering are obtained. The clustering that has been able group the users´ songs in the best possible way according to Shannon entropy is selected as the right clustering for that user. Using this content based recommendation scheme, as opposed to a static set of features resulted in up to 60 percent increase in recommendation success. Based on the same clustering performance idea, inclusion of the singer information and the most popular songs at the time of recommendation, in addition of the content is also explored and has resulted in performance increase. We introduce a third algorithm that is based on adaptive groupings of users. This algorithm is the best performer among the three algorithms we consider. Experiments are carried out on user session data consisting of 2000 to 500 sessions of lengths 5 to 15.