TIGER
TIGER (Transformer Index for GEnerative Recommenders; Rajput et al., Google, 2023 — “Recommender Systems with Generative Retrieval”, arXiv:2305.05065) applies Generative Retrieval to recommendation. It is the work that popularized RQ-VAE Semantic IDs for items, carrying the DSI idea from document search into recsys.
How It Works
- Build semantic IDs — encode each item’s content into a dense embedding, then quantize it with RQ-VAE into a short sequence of codebook tokens.
- Sequence the session — represent a user’s interaction history as the concatenation of those item semantic IDs.
- Generate the next item — a seq2seq Transformer predicts the next item’s semantic ID token-by-token; beam search yields a ranked candidate list.
Why It Mattered
- Demonstrated that a recommender can be a single generative model rather than a dense-retrieve-then-rank pipeline.
- Showed semantic IDs improve cold-start and long-tail recommendation, because new/rare items get content-derived codes that share prefixes with popular neighbours.
- Established the RQ-VAE-semantic-ID recipe now reused across generative retrieval and ranking (e.g. semantic IDs for ranking at YouTube, arXiv:2306.08121).
Related Concepts
- Semantic IDs — the identifier scheme TIGER builds on
- RQ-VAE — produces TIGER’s item codes
- Generative Retrieval — the paradigm TIGER instantiates for recsys
- Differentiable Search Index — the document-retrieval precursor
Articles
- Semantic IDs for Recommendation Systems — Janu Verma; reproduces a TIGER-style pipeline on Amazon Beauty