Recommender Systems with Generative Retrieval (TIGER) — NeurIPS 2023

This entry from zhichai.net's awesome list (Sequential Recommendation section) indexes the NeurIPS 2023 paper "Recommender Systems with Generative Retrieval", which proposes TIGER (Transformer Index for GEnerative Recommenders).

Metadata

| Field | Content | |-------|---------| | Title | Recommender Systems with Generative Retrieval | | Method | TIGER (Transformer Index for GEnerative Recommenders) | | Venue | NeurIPS 2023 | | Link | NeurIPS 2023 proceedings | | Section | Sequential Recommendation |

One-line summary

TIGER reformulates recommendation as a generative retrieval problem: a Transformer directly generates semantic item identifiers to retrieve items, replacing the classic retrieve-then-rank pipeline.

Background and motivation

Traditional recommender pipelines separate retrieval, ranking, and generation stages, which limits scalability and the ability to exploit semantic understanding of items and user intent. TIGER addresses this by unifying retrieval and generation in a single sequence-to-sequence framework, in the spirit of generative information retrieval for the recommendation setting.

Core idea

• Items are encoded as semantic IDs — compact tuples of codebook tokens obtained by quantizing item content embeddings (e.g., via RQ-VAE style residual vector quantization).
• A Transformer encoder-decoder is trained on user interaction histories; the decoder autoregressively generates the semantic IDs of the next items.
• This approach improves generalization to cold-start items and reduces cascading errors inherent in multi-stage pipelines. The paper reports state-of-the-art results on public recommendation benchmarks; consult the original PDF for exact numbers.

Related entries in this list

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

Engineering considerations

| Concern | Guidance | |---------|----------| | Index updates | Semantic-ID codebooks must be versioned; retraining affects ID stability | | Latency | Autoregressive generation (often beam search) requires latency budgeting vs. ANN-based retrieval | | Evaluation | Compare offline metrics (Recall@k, nDCG@k) against online A/B results before adoption | | Safety | Output filtering still needed; generative retrieval can produce invalid IDs |

Takeaway

TIGER is a foundational reference for generative recommendation (Gen-Rec). Readers interested in the retrieval → ranking → generation pipeline shift, LLM-era recommendation architectures, and item-ID indexing strategies should read it alongside P5, BERT4Rec, and EAGER.