Evaluation of a recommendation system for musical contents

This paper considers the evaluation of music recommendation systems. These systems aim at providing those pieces of music most likely to fit userspsila preferences. The main limits of existing systems are their lack of precision and the high level of resources necessary for musical characterisation and comparison. In this study, we propose an approach which has a good performance-resources ratio. The evaluation process shows that the recommendation system is able to predict the userspsila preferences with reasonable accuracy. By automatically ranking a list of 5985 unknown songs in order of the userpsilas preference, our method is able to retrieve a known userpsilas preferred song in the first 3.5% positions of the ranked list.