External Features for Song Classification
Most approaches to song classification problems (genre detection, hit song prediction, mood detection, ...) primarily use so-called internal features to train their models. Internal features are features that are inherent to the song itself - i.e., audio descriptors. Another type of features are external features, which describe the context of a song. Examples for such features include how often the song's artist performs live, how much the record label invested in the marketing campaign for a given album or whether the artist has an active Twitter account.
Those features have the potential for improving existing models, but are hard to obtain. The goal of this thesis is to devise methods for automatically collecting such features for a given set of songs, for example by making use of linked open data sources.