Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging Memory Retrieval to Enhance LLM-based Generative Recommendation

Chengbing Wang, Yang Zhang, Fengbin Zhu, Jizhi Zhang, Tianhao Shi, Fuli Feng

TL;DR

This work tackles the challenge of leveraging long-term user histories in LLM-based generative recommendations, constrained by limited context windows. It introduces AutoMR, an external memory and a learned retriever that identify and extract the most relevant long-term interactions to augment LLM prompts for next-item generation. The memory stores past interactions as hidden representations, and the retriever is trained to prioritize memory elements that improve prediction perplexity, via KL-divergence-based supervision. Across two real-world datasets, AutoMR consistently outperforms strong baselines, including semantic retrieval methods, demonstrating the value of learning a recommendation-specific memory retrieval mechanism for long-term user interests.

Abstract

Leveraging Large Language Models (LLMs) to harness user-item interaction histories for item generation has emerged as a promising paradigm in generative recommendation. However, the limited context window of LLMs often restricts them to focusing on recent user interactions only, leading to the neglect of long-term interests involved in the longer histories. To address this challenge, we propose a novel Automatic Memory-Retrieval framework (AutoMR), which is capable of storing long-term interests in the memory and extracting relevant information from it for next-item generation within LLMs. Extensive experimental results on two real-world datasets demonstrate the effectiveness of our proposed AutoMR framework in utilizing long-term interests for generative recommendation.

Leveraging Memory Retrieval to Enhance LLM-based Generative Recommendation

TL;DR

This work tackles the challenge of leveraging long-term user histories in LLM-based generative recommendations, constrained by limited context windows. It introduces AutoMR, an external memory and a learned retriever that identify and extract the most relevant long-term interactions to augment LLM prompts for next-item generation. The memory stores past interactions as hidden representations, and the retriever is trained to prioritize memory elements that improve prediction perplexity, via KL-divergence-based supervision. Across two real-world datasets, AutoMR consistently outperforms strong baselines, including semantic retrieval methods, demonstrating the value of learning a recommendation-specific memory retrieval mechanism for long-term user interests.

Abstract

Leveraging Large Language Models (LLMs) to harness user-item interaction histories for item generation has emerged as a promising paradigm in generative recommendation. However, the limited context window of LLMs often restricts them to focusing on recent user interactions only, leading to the neglect of long-term interests involved in the longer histories. To address this challenge, we propose a novel Automatic Memory-Retrieval framework (AutoMR), which is capable of storing long-term interests in the memory and extracting relevant information from it for next-item generation within LLMs. Extensive experimental results on two real-world datasets demonstrate the effectiveness of our proposed AutoMR framework in utilizing long-term interests for generative recommendation.

Paper Structure

This paper contains 17 sections, 6 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. AutoMR
  4. Framework Overview
  5. Retriver Training
  6. Related Work
  7. Experiments
  8. Experimental Setup
  9. Datasets
  10. Baselines
  11. Evaluation metrics
  12. Implementation details
  13. Overall Results (RQ1)
  14. Analysis
  15. Comparison with Random Retrieval (RQ2).
  16. ...and 2 more sections

Figures (3)

  • Figure 1: An overview of the proposed AutoMR, which includes three key components: Memory, Retriever, and LLM-based generative recommender (e.g., BIGRec).
  • Figure 2: Performance comparison of AutoMR and random retrieval. The bottom, middle, and top lines of the box represent the 25th percentile, median (50th percentile), and 75th percentile, respectively.
  • Figure 3: Visualization of the historical interaction times for some samples where incorporating long-term history improves Recall@1. The orange circles represent the long-term history retrieved by AutoMR.