Adaptive user modeling for content-based music retrieval

Abstract

An approach to adapt a content-based music retrieval system (CBMR system) to the user is presented and evaluated. Accepted and rejected songs are gathered to extract the user's preferences. To compare acoustic characteristics of music files, profiles are introduced. These are based on result lists. Each result list is created by a classifier and sorted accordingly to the similarity of the given seed song. To detect important characteristics, the accepted and rejected songs are clustered with k-means. A score for each candidate song is specified by the distance to the mean values of the obtained clusters. The songs are proposed by creating a playlist, which is sorted by the score. Songs accepted by the listener are used to query the CBMR system for new songs and thus extract additional profiles. It is shown that incorporating relevance feedback can significantly improve the quality of music recommendation. The L 2 distance is suitable to determine similarities between profiles of regarded songs. Introducing more than one query song during the recommendation process can further improve the quality. © 2010 Springer-Verlag Berlin Heidelberg.