| Summary: | Music recommender systems are used by millions of people every day to discover new and exciting music. Central to making recommendations is the representation of each track, which may be used to calculate similarity. Content representations capture the musical and texture facets of each track, and semantic representations describe social and cultural information provided by listeners. This thesis is motivated by an analysis of the strengths and weaknesses of both content and semantic representations. Content representations can be available for all tracks in a collection, but provide poor recommendation quality. Semantic representations suffer from the cold-start problem and are not available for all tracks, but provide good recommendation quality when a strong representation is available. These observations highlight the need to integrate both content and semantic representations, and use the strengths of each to improve music recommendation quality and discovery. A bridge of the gap between content and semantic representations is achieved in this thesis through hybrid representation. Content texture representations are examined, and a new music-inspired texture representation is defined. This content is integrated with semantic tags directly, and through a mid-level pseudo-tag representation. The effect of these approaches is to increase the high quality discovery of tracks, and to allow users to uncover interesting novel recommendations. The challenge of evaluating music recommendations when many tracks are undertagged is addressed. Implicit and explicit feedback provided by users is exploited to define a new ground truth similarity measure, which accurately reflects how different recommendation methods perform. A user study is conducted to evaluate both this measure, and the performance of integrated representations for recommending strong novel recommendations.
|
|---|
