How we improved the 'Unfamiliar' setting recommendation algorithm in Yandex Music

Savva Stepurin25 May 2024

Problem

Yandex Music runs "My Wave", an endless personalized stream. Its "Unfamiliar" setting shows only tracks the user has not yet heard, and users react actively: likes and skips from the first seconds. The standard pipeline fits this setting badly. Its candidate selectors lean on history and likes, which the setting excludes, and its ranker is trained on regular-wave behavior. The team wanted more rare content and more serendipity (a track the user likes and follows up with the artist's other songs) without losing retention.

Example from the article: a track enters "Unfamiliar" only if it was never shown to the user, the artist was shown at most 20 times, and the user has no likes on the artist or album.

Solution in brief

Add a candidate selector that takes popular tracks of artists similar to artists of random user likes. Train a separate CatBoost ranker on the full event pool, with "Unfamiliar" examples and likes weighted higher. In A/B tests the share of rare tracks rose 35% and retention 1%.

Questions this case answers
  • How to make a recommendation mode for never-seen items surface rarer items and more serendipity when the usual history-based candidate selection is excluded
  • How to train a ranker for a recommendation mode that has few training examples and must respond within a strict runtime budget
  • How to make a recommendation mode for new content (items the user has never seen) surface rarer items and feel more serendipitous
  • How to choose candidate selectors for a mode that must exclude the user's history and likes

Tools and methods

  • Candidate selectors — each returns a subset; the union gives ~10k tracks
  • Similar-artist selector — popular tracks of artists similar to those of random likes
  • CatBoost with PairLogit — pairwise ranking of positive vs negative events
  • ALS, transformers, DSSM, two-tower — embeddings and ranking features
  • HNSW, k-means — nearest-neighbor search
  • A/B testing — rare-track share, diversity, likes, serendipity, retention
Continue reading — it's free

Google or a one-time email code. No password.