A personalized content-based music similarity function

Modeling music similarity is useful for a great variety of applications in everyday life: music recommendation, thematic playlists generation and music libraries organization are just some examples. The main goal is to find a method to automatically evaluate how much two generic audio tracks are similar. In this book, we devise an adaptive non-linear method to model the concept of music similarity. Our system collects similarity evaluations from each user, and exploits these data to fit personalized similarity functions. To this end, we exploit the content-based information inherently contained in music tracks, modeled by means of suitable low-level features. We rely on two strongly consolidated techniques: Multidimensional Scaling (MDS), a feature embedding method, and Support Vector Regression (SVR), a robust regression approach that can be made non-linear by using suitable kernel functions. The result is a novel system that is able to learn the similarity concept as it is perceived by each user, thus providing more accurate and user-tailored evaluations.