Problem
Avito's main-page recommendations use a three-stage pipeline: several candidate engines, a ranker, and a blender for business logic. Every candidate engine works inside one category, so a user who viewed phones sees only phones. The team wanted cross-category recommendations, but the obvious route was a collaborative model on a 200 million ads × 50 million users matrix. That means heavy infrastructure changes with no guaranteed gain.
Example from the article: a user who viewed tires 10 times should get tires of different sizes and car parts, not Apple phones or real estate.
Solution in brief
Leave the existing engines alone and add a light engine that predicts categories instead of ads. A category is a leaf of a tree built from item parameters. Ads are then retrieved from the predicted categories by parameter search. The team started with a Markov chain over categories and later replaced it with SASRec, a transformer over the category sequence, behind the same service API.
Questions this case answers
- How to add cross-category candidates to a system whose candidate engines each work inside one category, by predicting categories from a user's sequence and then retrieving items
- How to replace a Markov chain with a sequential transformer behind an unchanged service interface when the item catalog is too large for the model
- How to add cross-category recommendations to a system where every candidate engine works inside one category
- How to recommend categories (not items) from a user's click history and then retrieve items from them
Tools and methods
- Markov chain over categories — baseline: a category-to-category transition matrix trained on two weeks of clickstream.
- SASRec — transformer over the user's category history; replaced the Markov chain.
- ALS — collaborative baseline, used only for offline comparison.
- Category tree from item parameters — about 6000 leaf categories, built from the ad metadata system.
- A/B tests with ranker retraining — online check of the mixed-in candidates.
Google or a one-time email code. No password.