# How to Index Item IDs for Recommendation Foundation Models

**Authors:** Wenyue Hua, Shuyuan Xu, Yingqiang Ge, Yongfeng Zhang  
**arXiv:** [2305.06569](https://arxiv.org/abs/2305.06569)  
**Date:** 2023-05-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.06569v1/

## Key points

- The paper explores the use of large language models (LLMs) in recommendation systems
- Proposes a recommendation foundation model that utilizes LLMs to convert recommendation tasks into natural language tasks
- Simplifies the recommendation pipeline by reducing it from multi-stage filtering to single-stage filtering
- Highlights the importance of creating LLM-compatible item IDs for recommendation foundation models
- Examines the problem of item indexing for these models using P5 as the representative backbone model
- Discusses trivial item indexing methods such as independent indexing, title indexing, and random indexing and highlights their limitations
- Proposes four simple yet effective solutions: sequential indexing, collaborative indexing, semantic (content-based) indexing, and hybrid indexing
- Conducts a reproducibility study of P5 and demonstrates how different item indexing methods significantly influence performance
- Validates proposed solutions on real-world datasets and shows their effectiveness in improving recommendation outcomes

## Summary

The paper titled "How to Index Item IDs for Recommendation Foundation Models" by Wenyue Hua, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang explores the use of large language models (LLMs) in recommendation systems. The authors propose a recommendation foundation model that utilizes LLMs to convert recommendation tasks into natural language tasks. Unlike traditional recommendation models that calculate ranking scores for candidate items, this generative recommendation model directly generates the item(s) to recommend. This simplifies the recommendation pipeline by reducing it from multi-stage filtering to single-stage filtering. To ensure that excessively long text is not generated when deciding which item(s) to recommend, the authors highlight the importance of creating LLM-compatible item IDs for recommendation foundation models. In this study, they systematically examine the problem of item indexing for these models using P5 as the representative backbone model. They replicate P5's results with various indexing methods to emphasize the significance of item indexing. The paper discusses several trivial item indexing methods such as independent indexing, title indexing and random indexing and highlights their limitations. To address these issues, the authors propose four simple yet effective solutions: sequential indexing, collaborative indexing, semantic (content-based) indexing and hybrid indexing. The authors conduct a reproducibility study of P5 and demonstrate how different item indexing methods significantly influence the performance of the model. They validate their proposed solutions on real-world datasets and show their effectiveness in improving recommendation outcomes. Overall, this paper provides valuable insights into optimizing recommendation foundation models by addressing the item indexing problem. The proposed solutions offer practical approaches to enhance the performance of generative recommendation systems based on LLMs.