Recommender Systems with Generative Retrieval
Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan H. Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, Maheswaran Sathiamoorthy
NeurIPS 2023)
cs.IR, cs.LG
2023-05-09
TIGER quantizes item text into hierarchical Semantic IDs and autoregressively decodes the next item, lifting Beauty NDCG@5 29% over SASRec and enabling cold-start retrieval.
Industrial retrieval usually means dual encoders: pack user history and candidate items into one vector space, then take top hits with ANN / MIPS. At a billion items the embedding table grows linearly with the catalog. New items with no interactions get a random ID and zero signal. Similar items cannot share parameters.
TIGER (Transformer Index for GEnerative Recommenders) drops the ANN. The Transformer is the index: it writes the identifier of the next item, token by token. Generating raw atomic IDs (the P5 path) is weak. The move that matters is giving every item a hierarchical Semantic ID derived from content.
Two stages.
Semantic IDs. Sentence-T5 embeds a sentence built from title, price, brand, and category into a 768-d content vector. RQ-VAE then applies three residual quantization levels, each with a codebook of 256 and a 32-d latent. Residuals shrink at each level, so early codes capture coarse category and later codes cut finer. Collisions get a fourth token, yielding a length-4 integer tuple per item. Codebooks are k-means initialized to avoid collapse, trained for about 20k epochs, with codebook usage at least 80%.
Generative retrieval. A user's chronological item sequence is rewritten as a stream of Semantic ID tokens. An encoder-decoder (T5X, four layers each side, six heads, about 13M parameters) predicts the four codes of the next item. The vocabulary is 256×4 = 1,024 item codes plus 2,000 hashed user tokens. Inference is beam search, one code at a time.
Three Amazon review subsets, users with at least five reviews. Beauty has 22,363 users and 12,101 items, mean sequence length 8.87.
TIGER beats the sequential baselines of the time on all three sets.
| Method | Beauty Recall@5 | Beauty NDCG@5 | Sports Recall@5 | Toys Recall@5 |
| SASRec | 0.0387 | 0.0249 | 0.0233 | 0.0463 |
| S3-Rec | 0.0387 | 0.0244 | 0.0251 | 0.0443 |
| TIGER | 0.0454 | 0.0321 | 0.0264 | 0.0521 |
On Beauty, NDCG@5 is 29% above SASRec and Recall@5 is 17% above S3-Rec. NDCG@5 gains on Sports and Toys are 13% and 21%.
How IDs are built sets the ceiling. Beauty Recall@5: random IDs 0.0296, LSH 0.0379, RQ-VAE 0.0454. Random codes fall back toward weak baselines, so being generative is not enough; the codes need content structure. Qualitatively, the first code tracks coarse categories such as makeup versus hair; the second code cuts finer.
Two capabilities fall out. Cold start: 5% of test items are removed from training, and matching the first three generated codes retrieves unseen items; at ε=0.1, Recall@K beats KNN in the raw semantic space. Diversity: raising decode temperature from 1.0 to 2.0 lifts Entropy@10 on Beauty from 0.76 to 1.38, and the same query spills into neighboring categories.
Invalid IDs happen, but only about 0.1%–1.6% of top-10 beams; a larger beam plus filtering clears them. The 256^4 code space is about 4 trillion entries against 10k–20k items, and the model almost always lands on a valid code.
This is the paper that moved generative retrieval from document IDs into recommender systems, and the common ancestor of OneRec, GRID, and later SID work. Hierarchical Semantic IDs fix three things at once: the embedding table no longer grows linearly with the catalog, similar items share prefixes, and unseen items can be decoded from content codes. Temperature sampling and prefix matching give diversity and cold-start knobs that dual encoders do not expose so directly.
The 29% on Amazon 5-core is real. It does not automatically transfer to a billion-item, long-session setting. You still pay for tokenizer training, collision handling, and invalid-code filtering.
The three Amazon subsets are small and short, with mean sequences under nine steps, not industrial sessions. The P5 baseline used a modified preprocess, so that comparison is softer. RQ-VAE trains for a long time and is sensitive to init; GRID later showed residual K-Means can beat it. Invalid codes are patched by larger beams; prefix matching is left as future work. User tokens occupy 2,000 hash slots; GRID later found removing them is better, and this paper does not ablate that. The cold-start ε is a hand-set mix of seen and unseen items, not a ratio the model chooses.