AI BriefingScaling Conditional Learning Retrieval for Pinterest Home Feed
Key point
Pinterest expanded Conditional Learning Retrieval (CLR), which reflects multiple intents, to the home feed to improve efficiency.
Details
Candidate generation for the Pinterest home feed is a large-scale user-pin retrieval problem. While existing 2-tower models struggled to capture multiple intents with a single user embedding, Conditional Learning Retrieval (CLR) generates embeddings that reflect diverse interests based on explicit retrieval contexts.
CLR evolved from initial interest-based notifications to a multi-embedding retrieval framework for the home feed. It expanded its scope to support LLM-based user interest signals, pin conditions via clustering of recently engaged pins, and board conditions utilizing random walks on the board graph. This replaced legacy heuristic candidate generators and simplified the serving stack while achieving metric improvements.
On the infrastructure side, latency was reduced through GPU-based serving, and an agentic budget adjustment policy leveraging Claude Code was introduced to automatically tune the number of conditions per request. Additionally, the model foundation was strengthened by sharding a 20GB-scale ID embedding table with TorchRec to manage memory and collisions in large-scale ID spaces.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.