Music Recommendation with High-dimensional Features

Music recommender systems as part of music streaming platforms such as Spotify or Deezer are trained on large datasets containing user-song interaction logs. Additionally, each song can be represented via high-level features to improve the quality of the recommendations. However, there are many possibilities for how to represent a song and which meta-data to include, leading to the curse of dimensionality for feature-rich datasets. This thesis aims to find suitable methods for feature dimensionality reduction including the analysis of feature importance and to benchmark different approaches on a feature-rich music recommender dataset.