Accessing Music Collections Via Representative Cluster Prototypes in a Hierarchical Organization Scheme

This paper addresses the issue of automatically organizing a possibly large music collection for intuitive access. We present an approach to cluster tracks in a hierarchical manner and to automatically find representative pieces of music for each cluster on each hierarchy level. To this end, audio signal-based features are complemented with features derived via Web content mining in a novel way. Automatic hierarchical clustering is performed using a variant of the Self-Organizing Map, which we further modified in order to create playlists containing similar tracks. The proposed approaches for playlist generation on a hierarchically structured music collection and finding prototypical tracks for each cluster are then integrated into the Traveller’s Sound Player, a mobile audio player application that organizes music in a playlist such that the distances between consecutive tracks are minimal. We extended this player to deal with the hierarchical nature of the playlists generated by the proposed structuring approach. As for evaluation, we first assess the quality of the clustering method using the measure of entropy on a genre-annotated test set. Second, the goodness of the method to find prototypical tracks for each cluster is investigated in a user study.