Problem
At X5 (Perekrestok app) the recommendation team and the advertising team had separate KPIs and fought over the same slots. Ads sat on a fixed grid in the "You may be interested" shelf on the cart page, mixed with personal recommendations. The business could not control key metrics or place ads by any common logic. Users saw irrelevant items and distrusted the shelf. Advertisers got low purchase conversion and no control over the audience.
Example from the article: every third item in the feed was an ad, and advertisers also wanted to promote items that were not yet on the shelves, for example tulips before March 8.
Solution in brief
A reranker built on top of a two-stage personal recommender puts promoted items into the same ranked list as organic ones. It was developed in four iterations: replace analogs with ads, train the ranker on relevance and revenue, use expected utility with eCPM as a single ad currency, then move to real-time serving. Cold items get embeddings from their description text.
Questions this case answers
- How to place paid promoted items into a personalized ranked feed without hurting the organic relevance metrics
- How to rank promoted and organic items in one model when each ad deal type has a different pricing unit
- How to blend paid promoted items into a personalized recommendation feed without hurting global business metrics
- How to rank promoted items by relevance, margin and advertiser revenue in one model
Tools and methods
- Bert4Rec — candidate generation and item embeddings for hot items
- CatBoost — second-stage ranker, also used to score promoted items
- RuBert-mini-frida — embeddings of item descriptions for cold items
- eCPM — common ad currency for CPD, CPM, CPC and CPA deals
- A/B/C tests — three groups compared with fixed ads
- Triton, Redis — online model serving and embedding storage
Google or a one-time email code. No password.
