Feature Learning for Graph-based Contextual Music Recommendations
Recommender systems are applied in different fields such as movie, friend or music recommendations and help users to cope with choice overload. In this work, we are primarily interested in music recommender systems. To compute personalized recommendations, as much information as possible about the items, users and the relation between these are necessary. To collect this information, we rely on different techniques: we rely on a set of Spotify playlists and use these to construct a graph containing all information about users, playlists, songs, artists and genres (including audio features, etc.). From the resulting graph, we aim to extract features using the node2vec learning algorithm to represent the relation of items within the graph. Furthermore, we apply the doc2vec algorithm to be able to extract further features from song lyrics. We use these extracted features in a hybrid recommender approach in which content-based and collaborative filtering features are used to suggest items to a user. For a combination of the features a factorization machine is used to compute the final recommendations.