Approved
Music recommendations with deep learning
Carl Rynegardh (2012)
Start
2017-09-04
Presentation
2019-06-14 14:00
Location:
Sal E:3139, E-huset, LTH
Finished:
2019-09-07
Master's thesis:
Abstract
In this thesis, we approach the problem of recommending music when there is no user data. Instead, we simulate users by making assumptions on their behavior. Recommendations are made sequentially, and after every recommendation feedback is generated by the simulated user that is used to generate a new recommendation. Every simulated user has specific feature values it considers as optimal. To find the optimal feature values for the user, deep reinforcement learning is applied, and to represent music as a numeric vector a pretrained deep learning model is used, where features are extracted from the different layers. The results show the difficulty of representing music with only a few dimensions, but that the deep reinforcement learning system finds the optimal feature values.
Supervisor: Fredrik Edman (EIT)
Examiner: Jan Eric Larsson (EIT)