47 cases from Pinterest (Social, Adtech, E-commerce), each digested from a public talk or engineering post and linked to the original. 28 were checked in an online A/B test; 16 report results in production.
Tasks: Retrieve and rank from millions of items, Reduce ML cost, Recommend from what the user just did, Balance relevance and diversity, Evaluate models and 18 more. Methods: Multi-task learning, Two-tower model, Deep & Cross Network, Knowledge distillation, Two-stage recommender, Sequential recommendation and 80 more. Tools: PyTorch, Apache Airflow, Apache Spark, Ray, Qwen, Apache Iceberg and 24 more.
Tech blog: medium.com/@Pinterest_Engineering