Problem
The Yandex Market recommendation team feeds item vectors into two-tower models for ranking and candidate generation. The more information a vector holds, the better these models work. Content-only vectors (bag of words over item names) miss collaborative signal. Trainable ID vectors capture it, but cannot embed new items and may not fit into GPU memory with up to tens of billions of items. Interactions form a natural graph, so the team asked whether graph neural networks add value on top of content and ID vectors.
Example from the article: a user opens a Yandex card holder, sees a block of similar items and moves to a Yandex keychain. Such pairs train the graph item vectors.
Solution in brief
Compare item encodings through reuse: feed a two-tower model's output into the final CatBoost ranker and measure relative NDCG. Content plus hashed ID embeddings beat graph vectors (Twin, PinSage-like), and graphs on top of content plus ID added nothing measurable.
Questions this case answers
- How to encode a very large catalog of entities so that new entities without history still get vectors that carry collaborative signal
- How to decide whether graph-based representations add value on top of content and ID embeddings, measured through the downstream ranking model
- How to encode items for a recommender so that the item vector carries collaborative signal
- How to decide whether graph neural network embeddings are worth building on top of content and ID embeddings
Tools and methods
- Bag-of-words content vector — tokenizer trained on item names
- Multisite unified embedding — hashed ID embeddings
- Twin — transductive link prediction on a user-item graph
- PinSage-like encoder — transformer over random-walk neighbors
- Two-tower model plus CatBoost — evaluation setup
Google or a one-time email code. No password.
