Design User and Item Embeddings from Long Histories
Company: Pinterest
Role: Machine Learning Engineer
Category: ML System Design
Difficulty: medium
Interview Round: Onsite
Overview: Design user and visual-item embeddings for retrieval or recommendation when histories can contain 160,000 events. Address time-correct training data, multimodal cold start, efficient long-history summaries, incremental freshness, versioned nearest-neighbor serving, privacy, and deletion.