# Chat-REC: Towards Interactive and Explainable LLMs-Augmented Recommender System

**Authors:** Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, Jiawei Zhang  
**arXiv:** [2303.14524](https://arxiv.org/abs/2303.14524)  
**Date:** 2023-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.14524v2/

## Key points

- The paper proposes a novel paradigm called Chat-Rec (ChatGPT Augmented Recommender System) to address limitations of traditional recommender systems
- Chat-Rec empowers recommender systems with a conversational interface for interactive and explainable recommendations
- It utilizes large language models (LLMs) to convert user profiles and historical interactions into prompts for generating personalized responses
- The framework consists of multiple modules, including recommendation module and candidate set compression using LLMs
- LLMs enhance recommendation reasoning ability and provide more accurate and personalized product recommendations
- Chat-Rec effectively improves top-k recommendations compared to traditional recommender systems in experiments conducted by the authors
- It performs better in zero-shot rating prediction tasks as well
- The proposed framework offers new possibilities for implementing AI-generated content in recommender system research.

## Summary

The paper proposes a novel paradigm called Chat-Rec (ChatGPT Augmented Recommender System) to address the limitations of traditional recommender systems, such as poor interactivity and explainability. The Chat-Rec framework empowers recommender systems with a conversational interface, making interactive and explainable recommendations possible. It utilizes large language models (LLMs) to convert user profiles and historical interactions into prompts, which are used to generate personalized and relevant responses to user queries. The Chat-Rec framework consists of multiple modules. In the recommendation module, the prompt constructor generates a natural language paragraph summarizing the user's query and recommendation information. This intermediate answer is then used to refine the prompt constructor and generate an optimized prompt for further compressing and refining the candidate set. The resulting recommendation, along with a brief explanation, is outputted to the user. To improve the performance of recommender systems, the paper introduces candidate set compression using LLMs. User profiles and historical interactions are converted into prompts that LLMs can learn from in order to summarize user preferences for items in a domain. By utilizing in-context learning, LLMs enhance their recommendation reasoning ability and provide more accurate and personalized product recommendations. Once LLMs have learned the user's preferences, they are provided with the candidate set generated by the recommender system. LLMs further filter and sort this candidate set based on the user's preferences, presenting a smaller and more relevant set of items to increase the likelihood of finding something they like. In experiments conducted by the authors, Chat-Rec effectively improved top-k recommendations compared to traditional recommender systems. It also performed better in zero-shot rating prediction tasks. The proposed framework offers a novel approach to improving recommender systems by incorporating conversational interfaces and presents new practical scenarios for implementing AI-generated content in recommender system studies. Overall, this paper introduces Chat-Rec as an innovative paradigm that addresses limitations of traditional recommender systems by leveraging LLMs for conversational recommendations. The framework enhances interactivity and explainability, improves recommendation performance through candidate set compression, and offers new possibilities for AI-generated content in recommender system research.