# Generative Recommender with End-to-End Learnable Item Tokenization

**Authors:** Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, Wayne Xin Zhao  
**arXiv:** [2409.05546](https://arxiv.org/abs/2409.05546)  
**Date:** 2024-09-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.05546v3/

## Key points

- Traditional approaches treat item tokenization and generative recommendation training as separate processes, leading to suboptimal performance
- ETEGRec is an End-To-End Generative Recommender that integrates item tokenization and generative recommendation within a unified framework
- The authors propose a novel recommendation-oriented alignment strategy with two key optimization objectives: sequence-item alignment and preference-semantic alignment
- Mutual enhancement is fostered by tightly coupling the learning processes of the item tokenizer and the generative recommender through these objectives
- An alternating optimization technique ensures stable and efficient end-to-end training of the entire framework
- Extensive experiments show superior performance of ETEGRec compared to traditional sequential recommendation models and existing generative recommendation baselines
- The work has been accepted by SIGIR 2025 Research Track, highlighting its significance in the field
- The code for ETEGRec is openly available on GitHub at https://github.com/RUCAIBox/ETEGRec

## Summary

, , , , 


Generative Recommendation Systems: Advancing with ETEGRec


However, constructing effective item identifiers that align with recommender systems remains a significant challenge. Traditional approaches often treat item tokenization and generative recommendation training as separate processes, resulting in suboptimal performance. To address this issue, Enze Liu, Bowen Zheng, Cheng Ling, Lantao Hu, Han Li, and Wayne Xin Zhao introduce ETEGRec - an innovative End-To-End Generative Recommender that integrates item tokenization and generative recommendation within a unified framework. The authors propose a novel recommendation-oriented alignment strategy that includes two key optimization objectives: sequence-item alignment and preference-semantic alignment. By tightly coupling the learning processes of the item tokenizer and the generative recommender through these objectives, mutual enhancement is fostered. Additionally, an alternating optimization technique is developed to ensure stable and efficient end-to-end training of the entire framework. Extensive experiments showcased the superior performance of ETEGRec compared to traditional sequential recommendation models and existing generative recommendation baselines. The authors' work has been accepted by SIGIR 2025 Research Track, underscoring its significance in the field. For those interested in exploring further, the code for ETEGRec is openly available on GitHub at https://github.com/RUCAIBox/ETEGRec. In summary, Their approach demonstrates remarkable improvements over existing methods and sets a new standard for end-to-end learnable item tokenization in recommender systems.