Method
Two-tower model
Separate user and item encoders trained for retrieval, incl. DSSM.
60 cases from 23 companies apply Two-tower model. 30 were checked in an online A/B test; 18 report results in production. Includes variants: DSSM.
Typical tasks: Retrieve and rank from millions of items, Rank search results, Recommend to new users and new items, Learn user and item embeddings and 9 more. Companies: Yandex, Pinterest, LinkedIn, Meta, Avito, eBay and 17 more. Often combined with: DSSM, Contrastive learning, Multi-task learning, In-batch negatives, Approximate nearest neighbors and 105 more.



