Talk(DMRN+8): Adapting music similarity models to geographic user groups

We presented first experiments on adapting music similarity models to geographic user groups poster at DMRN+8 at Queen Mary University.

Authors:Daniel Wolff, Tillman Weyde and Andrew MacFarlane
Music Informatics Group, City UniversityLondon.

Abstract:
We present first results of experiments using localised music similarity data for user group-specific similarity prediction. Music similarity is a key component in many MIR disciplines such as music recommendation, but also a topic of research by its own in music psychology and ethnomusicology.

Today’s search and recommendation services already heavily rely on personalised query results and to this end there is a strong interest in adaptable similarity metrics. This study evaluates the feasibility of adapting similarity measures to location-specific subsets of user data. We use information on the country where the data was provided. Apart from directly learning distance metrics from the localised data, we employ a modification of the ITML algorithm to allow for the gradual adaptation of a previously learnt general similarity model to the location-specific data.

The data was collected in the online Game With A Purpose “Spot The Odd Song Out” and is based on the MagnaTagATune dataset and Million Song dataset. Players were presented a series of questions, where they were asked to choose the most dissimilar song out of three. This resulted in relative similarity data describing that the two remaining songs are more similar to each other than each of them is to the odd song out. Our first results show that relative learning with ITML is effective. We also show that training of similarity models to localised data is possible, but our early results do not exceed the performance of a
general model.

Download the slides here.