Problem
A home-feed recommender trained on clicks, saves and closeups keeps serving what a user already likes. A user can save ten sourdough recipes today and still churn next month, because nothing helped them find a new reason to return. Dominant interests also win at every stage: they score highest early, fill the candidate pool and crowd out smaller or growing interests. Pinterest found that retention depends on how many durable use-cases a user adopts, and the effect is sharp at the top decile.
Example from the article: a power user planning a wedding, remodeling a kitchen and training for a marathon may have fifteen active interests, yet ten distinct "apartment decorating" Pins can fill the feed.
Solution in brief
Cluster each user's last 500 engaged Pins into a variable number of User Interest Clusters (UICs) with complete-linkage hierarchical clustering on item embeddings. Keep them in a shared feature store and use them for retrieval conditions, a duplicate penalty before ranking, state-dependent ranking weights, and a coverage penalty in the final diversity step.
Questions this case answers
- How to represent a user as a variable number of coherent interest groups derived only from that user's own engagement history
- How to keep a multi-stage ranked feed from being dominated by a user's strongest interest when per-item scores favor it at every stage
- How to keep a recommendation feed covering all of a user's active interests, not only the dominant one?
- How to represent a user's several current use-cases so that every stage of a recommender can use them?
Tools and methods
- Complete-linkage agglomerative clustering — per-user interest clusters with no fixed count.
- OmniSage embeddings — Pinterest's item embeddings, close when items are used together.
- Shared feature store — serves UIC features once per request.
- Sliding Spectrum Decomposition (SSD) — final ordering, here with a UIC penalty.
Google or a one-time email code. No password.