Method

Semantic IDs

Discrete item codes from quantized embeddings for generative recommendation (TIGER).

9 cases from 7 companies apply Semantic IDs. 3 were checked in an online A/B test; 4 report results in production.

Typical tasks: Retrieve and rank from millions of items, Recommend to new users and new items, Recommend from what the user just did, Learn user and item embeddings and 5 more. Companies: Pinterest, Instacart, Meta, Spotify, Yandex, Avito and 1 more. Often combined with: Multimodal embeddings, Contrastive learning, Two-tower model, Supervised fine-tuning, LLM-as-a-judge and 29 more.