VK separated shopping recommendations from its general feed ranking, using CatBoost to prioritize orders while limiting product-post exposure.
VK has built a separate product-recommendation pipeline inside the VKontakte Feed, with Ozon as its first partner. Rather than adding products to the existing content model, the company gave commercial posts their own event logging, candidate selectors, ranking model and allocation within the final feed.
The separation reflects a different objective. The regular feed has long optimized for actions such as likes, comments, subscriptions and time spent, while the commercial system estimates whether showing a particular product to a particular user could lead to an order. VK says the candidate catalog currently contains about 74,000 product posts.
To address the initial shortage of training data, VK increased logging for product candidates from the feed's usual level of about 2% to 100%, without collecting every event across the entire feed. According to the company, this selective approach expanded the product-domain training sample fifteenfold. Before enough orders accumulated, the system also relied on prior commercial-content interactions, segment-level popularity lists and more frequent proxy events, including expanding a post before following its link.
Candidates come from two principal sources. One uses encrypted partner signals covering purchases, product-page openings, cart additions and favorites. The other connects ordinary content interests with related products: engagement with a recipe post, for example, can produce candidates featuring cookware, knives or recipe books. The ranking model then decides whether any of those candidates should appear.
VK implemented multi-target ranking with CatBoost. Orders receive the greatest weight, followed by click-outs to the partner; likes, comments and viewing time provide weaker positive signals. Irrelevant impressions, rapid scrolling and negative reactions are penalized. The baseline probability of inserting commercial content is 1.5%, rising to as much as 5% when the model identifies a highly converting product for that user.
That allocation is a ceiling, not a requirement. Product posts scoring below the relevance threshold are withheld, so the system can leave the commercial slot unfilled. VK also removes a post when the corresponding product is out of stock at the partner and restores it when inventory returns.
Comparing March 1–30 with August 1–30, VK reported a 53% increase in impression-to-created-order conversion. Created orders rose 43.2-fold, partner click-outs 55.3-fold, product-post views 28.3-fold and click-out CTR 95.6%. These are company-reported period comparisons: the source provides neither absolute values nor a control-group description. The order increase therefore reflects both greater exposure and higher reported conversion, rather than ranking quality alone.
| Metric | Change |
|---|---|
| Impression-to-created-order conversion | +53% |
| Created orders | 43.2-fold |
| Partner click-outs | 55.3-fold |
| Product-post views | 28.3-fold |
| Partner click-out CTR from a post | +95.6% |
Sources
Originally published by AI Feed: VK Built a CatBoost Product Recommender for Its Social Feed.
Top comments (0)