SYSTEM Cited by 1 source
Pinterest PinnerSage¶
Definition¶
PinnerSage (KDD 2020) is Pinterest's multi-modal, multi-embedding user representation framework. Its key idea: represent each user with multiple embeddings by clustering their engagement history and using cluster medoids as retrieval queries — a significant step beyond single-embedding user models, because it lets the system capture a user's diverse interests simultaneously rather than averaging them into one vector (Source: sources/2026-07-27-pinterest-pinner-progression-better-use-case-representation-driving-weekly-active-user-growth).
Role in the lineage¶
PinnerSage is the conceptual ancestor of User Interest Clusters (UIC):
- PinnerSage — cluster engagement history → medoids as retrieval queries (multiple embeddings per user).
- OmniSage — multi-entity graph representation with functional-utility closeness.
- UIC — personalized clustering over engaged content, dynamic cluster count, and stateful lifecycle metadata layered on top.
The "cluster the user's history, use medoids" mechanic in UIC is directly inherited from PinnerSage; UIC's additions are per-user personalization, dynamic k, and lifecycle state.
Seen in¶
- sources/2026-07-27-pinterest-pinner-progression-better-use-case-representation-driving-weekly-active-user-growth — cited as the multi-embedding foundation UIC builds on.