Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DRE: Generating Recommendation Explanations by Aligning Large Language Models at Data-level

Shen Gao, Yifan Wang, Jiabao Fang, Lisi Chen, Peng Han, Shuo Shang

TL;DR

DRE introduces a non-intrusive, data-level explanation framework for black-box recommender systems that does not require access to internal representations. It leverages in-context reasoning with large language models to align user history and recommendations at the data level, and augments explanations through target-aware user preference distillation using item descriptions and reviews. Experimental results on Amazon-category datasets show DRE achieves higher explanation quality than state-of-the-art latent-level methods and ablations demonstrate the importance of reviews and the distillation process. The approach improves transparency and user engagement by providing accurate, user-centric explanations without modifying the underlying recommender model.

Abstract

Recommendation systems play a crucial role in various domains, suggesting items based on user behavior.However, the lack of transparency in presenting recommendations can lead to user confusion. In this paper, we introduce Data-level Recommendation Explanation (DRE), a non-intrusive explanation framework for black-box recommendation models.Different from existing methods, DRE does not require any intermediary representations of the recommendation model or latent alignment training, mitigating potential performance issues.We propose a data-level alignment method, leveraging large language models to reason relationships between user data and recommended items.Additionally, we address the challenge of enriching the details of the explanation by introducing target-aware user preference distillation, utilizing item reviews. Experimental results on benchmark datasets demonstrate the effectiveness of the DRE in providing accurate and user-centric explanations, enhancing user engagement with recommended item.

DRE: Generating Recommendation Explanations by Aligning Large Language Models at Data-level

TL;DR

DRE introduces a non-intrusive, data-level explanation framework for black-box recommender systems that does not require access to internal representations. It leverages in-context reasoning with large language models to align user history and recommendations at the data level, and augments explanations through target-aware user preference distillation using item descriptions and reviews. Experimental results on Amazon-category datasets show DRE achieves higher explanation quality than state-of-the-art latent-level methods and ablations demonstrate the importance of reviews and the distillation process. The approach improves transparency and user engagement by providing accurate, user-centric explanations without modifying the underlying recommender model.

Abstract

Recommendation systems play a crucial role in various domains, suggesting items based on user behavior.However, the lack of transparency in presenting recommendations can lead to user confusion. In this paper, we introduce Data-level Recommendation Explanation (DRE), a non-intrusive explanation framework for black-box recommendation models.Different from existing methods, DRE does not require any intermediary representations of the recommendation model or latent alignment training, mitigating potential performance issues.We propose a data-level alignment method, leveraging large language models to reason relationships between user data and recommended items.Additionally, we address the challenge of enriching the details of the explanation by introducing target-aware user preference distillation, utilizing item reviews. Experimental results on benchmark datasets demonstrate the effectiveness of the DRE in providing accurate and user-centric explanations, enhancing user engagement with recommended item.
Paper Structure (14 sections, 3 equations, 3 figures, 1 table)

This paper contains 14 sections, 3 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. DRE Methodology
  4. Data-level Alignment
  5. Target-aware User Preference Distillation
  6. Explanation Generation
  7. Experimental Setup
  8. Evaluation Metrics & Dataset
  9. Comparison Methods
  10. Experimental results
  11. Main Results
  12. Ablation Study
  13. Case Study
  14. Conclusion

Figures (3)

  • Figure 1: Comparison between existing latent-level and our proposed data-level recommendation explanation method.
  • Figure 2: Overview of DRE, which firstly align the explanation module and recommender with Data-level Alignment, and then generate the explanation by incorporating details of target from Target-aware User Preference Distillation
  • Figure 3: An example of explanation by DRE and ChatGPT.