Master’s Thesis in Computer Science: Recognizing artist ambiguity with machine learning techniques

Spotify is a music streaming service whose ambition is to offer everyone easy access to all of the world’s music. Maintaining metadata quality is a a nontrivial challenge considering that the catalogue size rules out the possibility for manual curation and that the number of content providers and different delivery formats is large. An interesting problem that involves metadata is to have a one-to-one mapping between real-world and Spotify artists: two different artists should be two separate entities on Spotify. Currently, when several albums by artists with the same name are sent to the service, they are arranged under the same artist identifier, assuming that they belong together. This is an issue when two artists with the same name have their albums on Spotify: all the albums are grouped together, so users think they belong to the same artist and fans have a bad user experience when presented with additional clutter besides their favorite artist’s releases. This thesis explores the use of machine learning techniques to detect Spotify artists credited as having produced albums that actually belong to several real-world artists, namely ambiguous artists. Several features for representing the artists are presented, such as the existence of multiple matches between the Spotify artist and external music databases with curated content and the number of countries the artist has registered recordings in. Using every possible combination of the features, the examples are classified with Naive Bayes and logistic regression. Two of the resulting best performant classifiers with low false positive rates are then queried with unseen data sets of random and most popular artists to assess their predictions. We found that the most useful features for the classification were the existence of multiple matches between the Spotify artist and the external music databases, the artist’s name length and the number of languages of the artist’s track names. Logistic regression proved to be superior to Naive Bayes on a test set of random artists. When the classifier has detected the ambiguous artists, after a manual artist separation process, heuristics can be put into place so that incoming albums that could belong to more than one artist will be assigned to the most likely one. This practical solution to the artist ambiguity problem is also briefly discussed.