TIGER: Recommender Systems with Generative Retrieval (NeurIPS 2023)

Metadata

| Field | Content | |-------|---------| | Title | Recommender Systems with Generative Retrieval (TIGER — Transformer Index for GEnerative Recommenders), NeurIPS 2023 | | Venue | NeurIPS 2023 | | Source | https://proceedings.neurips.cc/paper_files/paper/2023/hash/20dcab0f14046a5c6b02b61da9f13229-Abstract-Conference.html | | Category | Sequential Recommendation |

Overview

TIGER (Transformer Index for GEnerative Recommenders) reframes recommendation as a generative retrieval problem: rather than scoring and ranking a large candidate pool, the model directly *generates* the identifiers of the items the user is likely to interact with next.

Key Ideas

1. Semantic IDs via RQ-VAE: Item embeddings are quantized into short tuples of discrete codes using a residual quantized variational autoencoder (RQ-VAE). Items with similar content semantics receive similar code tuples, giving the index a semantic structure. 2. Sequence-to-sequence generation: A Transformer encoder-decoder consumes the user's interaction history (as sequences of semantic IDs) and autoregressively generates the semantic IDs of recommended items via beam search. 3. Semantic generalization: Because the item space is represented by content-derived codes rather than random IDs, TIGER generalizes to cold-start and long-tail items better than ID-based sequential models.

Method Pipeline

1. Input & representation: Encode item content into embeddings, then quantize them into hierarchical semantic ID tuples with RQ-VAE. 2. Core module: A pretrained Transformer encoder-decoder maps interaction histories to generated semantic ID sequences. 3. Learning strategy: Supervised training on sequential interaction data, predicting next-item semantic IDs. 4. Inference: Autoregressive beam search over the semantic ID vocabulary; generated IDs are mapped back to items.

Experiments and Evaluation

• Benchmarks: Public sequential recommendation datasets.
• Baselines: Established sequential recommenders (e.g., ID-based Transformer models in the BERT4Rec / SASRec family).
• Metrics: Standard top-K recommendation accuracy metrics (e.g., Recall@K, NDCG@K).
• Findings (per the paper): TIGER achieves state-of-the-art accuracy on most benchmarks and notably improves cold-start recommendation. Exact numbers should be verified against the original PDF.

Takeaways for Search / Rec / Personalization

1. Architecture: Generative retrieval collapses retrieval and ranking into one model; index design (how item IDs are constructed) becomes a first-class modeling decision. 2. Data: Content features feed the semantic ID quantizer, so item metadata quality directly affects cold-start performance. 3. Evaluation: Beyond accuracy, generative recommenders require checking validity of generated IDs, latency of beam search, and index update procedures when the item catalog changes. 4. Engineering: Cascaded pipelines may still be needed for large catalogs; semantic ID vocabularies must be versioned and retrained consistently.

Limitations and Future Work

• Index updates require re-quantization when the item catalog or embeddings change.
• Beam search latency may exceed traditional ANN retrieval budgets in large-scale serving.
• Cross-lingual, multimodal, and fairness/constraint-aware extensions remain open directions.

Related Entries (Cross-references)

• BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
• EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
• Efficient On-Device Session-Based Recommendation (ACM TOIS)
• How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
• LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with LLMs
• Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models

Glossary

| Term | Meaning | |------|---------| | RQ-VAE | Residual Quantized Variational Autoencoder, used to produce hierarchical discrete codes | | Semantic ID | Compact tuple of discrete codes representing an item's content semantics | | Generative Retrieval | Directly generating item/document identifiers instead of scoring a candidate pool | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric | | Gen-Rec | Generative Recommendation, the broader paradigm TIGER belongs to |

References

• Original paper: Recommender Systems with Generative Retrieval — NeurIPS 2023 Proceedings