Problem
The Yandex TV recommender team ranks movies in personalized carousels on the Kinopoisk home screen. Content has two types with different data richness, and each type had its own model: a neural ranker for one, CatBoost for the other. Two models give two scores with different distributions that must be blended, although both types are movies and series.
Example from the article: Type A (Kinopoisk movies, about 20 thousand items) has 35 million interactions a week. Type B (external movies watched in partner cinemas, about 10 thousand items) has 7 thousand, because the team loses sight of what users do in the partner app.
Solution in brief
Replace both models with one multi-head neural ranker: a shared body and one scoring head per content type, trained in two stages with freezing. Add a second output to each head that is trained without group weights and dropped at inference. The first version lost 2.4% to 5.6% online; the version with the extra output won.
Questions this case answers
- How to replace separate rankers for two item types with very different data volumes by one multi-head model that transfers behavior from the data-rich type
- How to stop a ranker trained on watch-time weights from overfitting to short-term engagement by adding an unweighted auxiliary output
- How to replace separate rankers for two content types with one multi-head neural ranker
- How to train one model on two content types when one has 35 million interactions a week and the other has 7 thousand
Tools and methods
- DeepRanker (DCN V2-style) — shared neural ranker with embeddings
- CatBoost — the original type B ranker
- Two-stage training with frozen layers — fits the data-poor type on top of the data-rich one
- Auxiliary
_nwoutput — trained without weights, not used for ranking - NDCG offline; TVT and DAU online — metrics
Google or a one-time email code. No password.
