The thesis addresses the problem of automatic music playlist continuation,
which is essential for enhancing the user experience in music streaming services
such as Spotify. Due to the large amount of data and the implicit nature
of user feedback, the thesis presents the use of the LightGCN architecture,
a simplified version of graph convolutional neural networks, which enables
efficient learning of embeddings of playlists and songs. The model was evaluated
on the publicly available Million Playlist Dataset, where it achieved
competitive results compared to the existing literature. In particular, Light-
GCN outperforms the best approaches in the literature with respect to the
recall metric.
|
