Adapting Metrics for Music Similarity Using Comparative Ratings

Understanding how we relate and compare pieces of music has been a topic of great interest in musicology as well as for business applications, such as music recommender systems. The way music is compared seems to vary among both individuals and cultures. Adapting a generic model to user ratings is useful for personalisation and can help to better understand such differences. This paper presents an approach to use machine learning techniques for analysing user data that specifies song similarity. We explore the potential for learning generalisable similarity measures with two stateof-the-art algorithms for learning metrics. We use the audio clips and user ratings in the MagnaTagATune dataset, enriched with genre annotations from the Magnatune label. 1. MOTIVATION In the recent years, increased efforts have been made to adapt MIR techniques, especially for music recommendation, to specific contexts or user groups. This is encouraged by developments in machine learning that make more algorithms applicable to accumulated user data, like user preferences or click-trough data for ranked search results, and enable the involvement of crowd wisdom into general classification and distance learning tasks. Moreover, the combination of different information sources has been proven successful for improving music recommendation and for classification into cultural categories such as musical genres. This paper shows the results of some experiments on learning a musical distance metric from user similarity comparisons. Similarity models of mixed acoustic and tag features are trained using comparative user judgent data on song similarities. We derive information of the form ”Song A is more similar to Song B than to Song C”, represented by binary