Experimenting with music taste prediction by user profiling

• Authors: Marco Grimaldi;Pádraig Cunningham
• Venue: Proceedings of the 6th ACM SIGMM international workshop on Multimedia information retrieval
• Year: 2004

In recent years many research projects have been published in the area of multimedia information retrieval (MIR). The requirement is to enable access to multimedia data with the same ease as textual information. A distinctly new branch in the MIR research area is categorizing music items by user preference. Some experiments proposed and published by different authors, showed that machine learning techniques can be applied to the problem. This work tries to extend the use of signal approximation and characterization from genre classification to recognition of user taste. The idea is to learn music preferences by applying instance based classifiers to user profiles. The audio signal (item) is characterized by features sensitive to music genres (Rock, Jazz, Classical, Techno). Two different classifiers are explored in order to determine the generalization accuracy of the system: k-NN and feature sub-pace based ensembles (FSSE). Feature selection techniques are explored to boost the accuracy of the predictor. The evaluation shows that this kind of problem can be solved to some extent. When the user taste is driven by a certain genre preference, the system shows reasonable accuracy