Method

Hard negative mining

Selecting difficult negatives to sharpen retrieval or classification.

14 cases from 11 companies apply Hard negative mining. 5 were checked in an online A/B test; 5 report results in production.

Typical tasks: Retrieve and rank from millions of items, Learn user and item embeddings, Rank search results, Match entities and 6 more. Companies: Yandex, VTB, Thumbtack, Pinterest, Rabota.ru, Ramp and 5 more. Often combined with: Two-tower model, In-batch negatives, Multimodal fusion, Triplet loss, DSSM and 40 more.