Scaling Conditional Learned Retrieval for Pinterest Home Feed

Pinterest engineers scaled conditional learned retrieval for home feed candidate generation, replacing two-tower models with a conditional approach that handles multiple user intents. The system retrieves User-to-Pin matches via approximate nearest neighbor search. Engineers include Devin Kreuzer, Yichi Wang, Sujan Reddy Ale, Zelun Wang, Hongtao Lin, and Piyush Maheshwari.