Content-Based Music Retrieval Using Query Integration for Users with Diverse Preferences

This paper proposes content-based music information retrieval (MIR) methods based on user preferences, which aim to improve the accuracy of MIR for users with “diverse” preferences, i.e., users whose preferences range in songs with a wide variety of features. The proposed MIR method dynamically generates an optimal set of query vectors from the sample set of songs submitted by the user to express their preferences, based on the similarity of the songs in the sample set. Experiments conducted on a music collection with subjective user ratings verify that our proposal is effective to improve the accuracy of contentbased MIR. Furthermore, by implementing a two-step MIR algorithm which utilizes song clustering results, the efficiency of the proposed MIR method is significantly improved.