Content-Based Music Recommender Systems: Beyond simple Frame-Level Audio Similarity Dissertation zur Erlangung des akademischen Grades

This thesis aims at improving content-based music recommender systems. Besides a general introduction to music recommendation and an in-depth discussion of evaluation methods of content-based music recommender systems, improvements on two different abstraction levels are considered in this thesis: The first and most obvious way to improve a content-based music recommender system is to improve the underlying music similarity measure that is used to generate the recommendations. State-of-the-art frame-level audio similarity algorithms are analyzed and improvements and limitations are discussed. Then a novel blocklevel feature extraction framework and a set of novel block-level features are introduced. To generate recommendations based on these block-level features two approaches are presented. One approach is based on directly estimating similarities based on the block-level features, and the other approach performs a mapping onto a semantic tag space before estimating pairwise song similarities. Finally, it is proposed to combine these two approaches. This combination approach ranked first in the MIREX 2010 Audio Music Similarity and Retrieval task. The second way of improving content-based music recommenders considers a more abstract level of a music recommendation system. At this higher level of abstraction a music recommender system is interpreted as a recommendation network. Based on the analysis of the emerging recommendation network it will be shown that a straightforward top-N recommendation approach can significantly impact the reachability of songs in a music recommendation network. This can radically reduce the usability of music recommender systems. Two strategies to alleviate this issue are presented and evaluated. It will be shown that both strategies can significantly improve the reachability of songs within a music recommendation network.