Problem
The Tinkoff personalization team runs a content feed in the app. Its cards are heterogeneous (restaurant bookings, movie tickets, articles, courses, stories) and share only a title, a description and an image. Carousels (collections of same-type cards) were made by editors and ranked by boosting. Only about 60 carousels existed, so users kept seeing the same ones. Collaborative filtering did not fit: cards appear and expire constantly, and it does not guarantee that a carousel is thematically uniform.
Example from the article: a user who clicks on cinema cards should get a carousel built only of movies, not a mix of one movie, one investment story and one article.
Solution in brief
Embed every card from its text, embed the user in the same space (content-based SASRec), and score cards by cosine similarity. Build candidate carousels around anchor cards and pick the one whose anchor and center are both close to the user. In an A/B test this beat editor collections: engagement rate +15%, swipe rate +21%, CTR +18%.
Questions this case answers
- How to assemble thematically uniform recommendation carousels from heterogeneous cards using text embeddings, so that new cards are recommended without interaction history
- How to build a user vector in the same embedding space as the item content vectors from a sequence of clicks
- How to assemble personalized content carousels that beat editor-curated collections
- How to recommend cold items in a feed of heterogeneous cards using content embeddings
Tools and methods
- BERT-like text encoder — turns title plus description into a 312-dimensional card vector.
- Content-based SASRec — transformer that produces the user vector from the sequence of clicked card embeddings.
- MeanRec — baseline user vector: average of clicked card embeddings.
- Optimal anchor — carousel assembly with a relevance and homogeneity score.
- TD3+BC — offline RL, being tested at the time of the talk.
Google or a one-time email code. No password.
