SYSTEM Cited by 1 source
Pinterest Conditional Learned Retrieval (CLR)¶
Definition¶
Conditional Learned Retrieval (CLR) is Pinterest's learned Home Feed retrieval system that generates candidates by combining user information with a "condition" — a board, Pin, or topic — to ensure semantic relevance between the retrieved Pins and the condition (Source: sources/2026-07-27-pinterest-pinner-progression-better-use-case-representation-driving-weekly-active-user-growth). It is the stage-1 retriever in Pinterest's retrieval → ranking funnel.
From followed-interests to UIC conditioning¶
Previously, a primary CLR condition was followed-interests — the topics a user selected during onboarding. Two problems:
- Skew toward dominant interests — retrieved large volumes of candidates in well-established categories while leaving smaller or emerging interests with little representation.
- Static — the signal did not evolve with a user's changing behavior.
Pinner Progression replaced the followed-interest condition entirely with UIC clusters in a UIC-conditioned CLR. This makes retrieval:
- More personalized — conditions derive from each user's own engagement clusters, not a global interest taxonomy.
- More flexible — the system can control how many candidates are allocated per use-case (e.g. sample 5 of a user's 10 clusters and retrieve candidates for each).
Landmark selection: exploit → explore¶
- Early iterations selected a representative Pin ("landmark") from each cluster using recency + action-type weighting — effective for relevance but inherently exploit-heavy.
- A later iteration introduced frontier sampling: select landmarks at the boundary of a cluster in embedding space, farthest from the medoid — shifting retrieval toward a more exploratory distribution without sacrificing engagement.
Budgeting to avoid overfetch¶
A key production challenge was effective budgeting to avoid overwhelming downstream stages. Because UIC clusters are scoped to a user's currently active interests (recent engagement, above a coherence threshold), CLR avoids retrieving candidates for stale or decayed interests unlikely to convert. This reduced overfetch (broadly relevant but ultimately unused candidates) and translated to meaningful infrastructure cost savings, while keeping total candidate volume sent downstream unchanged.
Seen in¶
- sources/2026-07-27-pinterest-pinner-progression-better-use-case-representation-driving-weekly-active-user-growth — the retrieval integration point for UIC.