Meta-Learning for Next-Song Recommendations (Master Informatik/Komedia)


Music Recommendation Systems are well-suited to incorporate user feedback. When recommending the next song in a queue, for instance, the timespan between generation and consumption of recommendations is generally small. Moreover, multiple recommendations can be consumed during a single session as users tend to listen to more than a single song in a row. Enriching the quality of the experience can hence be achieved by allowing users to provide feedback to which the RS should immediately adapt. For instance, the system should recognise if people deviate from recommended paths and adjust its underlying model accordingly.

Meta-Learning has become known in the Machine Learning community as a means to adapt to changing environments with only a few shots. Without the necessity of retraining, the system rather “learns-how-to-learn”. That is, even if the observed patterns of user behaviour differ from what was expected, e.g. when users decide to refuse recommendations and listen to songs of their own choice instead, the RS can quickly learn how to deal with the new situation.

(The thesis can either be written in German or English.)


Tim Donkers