Method
Contrastive learning
Representation learning with positive/negative pairs (InfoNCE, NT-Xent).
48 cases from 28 companies apply Contrastive learning. 18 were checked in an online A/B test; 14 report results in production. Includes variants: InfoNCE.
Typical tasks: Retrieve and rank from millions of items, Learn user and item embeddings, Rank search results, Recommend to new users and new items and 7 more. Companies: Yandex, Pinterest, LinkedIn, Criteo, VK, Skoltech and 22 more. Often combined with: Two-tower model, InfoNCE, Embedding fine-tuning, Multimodal embeddings, Self-supervised learning and 86 more.





