Problem
VKontakte recommends communities ("authors") to users. Its algorithms work well for communities with more than 1,000 subscribers but are not built for small ones, which have few interactions to learn from. Yet small communities matter: those with fewer than 1,000 subscribers get 36.7% of feed views among their own subscribers. Users no longer search for communities themselves, so a small community that never enters recommendations does not grow.
Example from the article: in the main recommendation block, new small communities lost to large ones in ranking and were not shown at all.
Solution in brief
Add a separate candidate source for small communities. A CatBoost model picks the ones likely to grow, a CLIP classifier removes junk, and a content network predicts each community's collaborative vector. The source is shown in its own block to active recommendation users only. This raised subscriptions to new authors by 31% without lowering total subscriptions.
Questions this case answers
- How to recommend small new items with little interaction data by predicting their collaborative vector from their content
- How to pick which cold items with growth potential to promote before they have interaction history
- How to recommend small new creators or communities that have little interaction data
- How to pick small authors with real growth potential before showing them to users
Tools and methods
- CatBoost — ranks communities by predicted follower growth, and ranks users' candidates in the main system.
- CLIP classifier (fine-tuned) — filters clickbait and 18+ content and assigns VK topics.
- Word2Vec, ALS — collaborative vectors for large communities; Word2Vec vectors are the training target for the content network.
- Content network — TF-IDF, image aggregation and meta-features to a 100-dimensional vector.
- A/B tests — three rollout variants for the new source.
Google or a one-time email code. No password.