Skip to content

Latest commit

 

History

History
20 lines (15 loc) · 2.36 KB

2312.10463.md

File metadata and controls

20 lines (15 loc) · 2.36 KB
Raw

Background

  • Background The paper discusses the impact of personalized news recommendation systems like Google News and Microsoft News on news consumption habits, emphasizing the importance of understanding news content accurately. It explores the potential of Large Language Models (LLMs) such as GPT-4 in news recommendation systems to accurately match news content with individual user preferences and interests.

  • Existing Work Existing research has not fully leveraged the advantages of LLMs in personalized news recommendation, especially in dynamically capturing user interests and providing highly relevant recommendations.

Core Contributions

  • Introduced the RecPrompt framework

    • Challenge 1: Optimizing news recommendation systems to capture user interests more effectively Current systems can't efficiently capture user interests using simple prompts. RecPrompt refines prompts iteratively and manually, combined with LLM evaluation and user behavior feedback, to further optimize the prompt template for improved recommendation performance.

    • Challenge 2: Automating the prompt tuning process to reduce the need for manual refinement Manual tuning of prompts is laborious. RecPrompt introduces another LLM for prompt optimization, which automatically generates better prompt templates by observing samples and instructions to enhance the efficiency of the recommendation system.

Implementation and Deployment

Experimental results comparing GPT-4 and GPT-3.5 using the RecPrompt framework show that both prompt strategies significantly improve recommendation performance. GPT-4 performs better than all deep neural models when using LLM-generated prompt templates. This demonstrates that the choice of LLM greatly affects the effectiveness of news recommendations, with GPT-4 outperforming traditional methods and its predecessor GPT-3.5 in topic extraction and summarization.

Summary

This paper presents the RecPrompt model, which optimizes news recommendation using LLMs. Through an iterative optimization process with manually and LLM-generated prompt templates, the news recommendation performance is significantly improved, particularly under the LLM-generated prompt templates utilizing GPT-4. However, this approach does not always outperform traditional recommendation methods and is significantly impacted by the choice of LLM.