Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RDRec: Rationale Distillation for LLM-based Recommendation

Xinfeng Wang, Jin Cui, Yoshimi Suzuki, Fumiyo Fukumoto

TL;DR

RDRec tackles the gap in LLM-based recommender systems by learning interaction rationales from reviews via a larger LM and distilling them into a compact model. It introduces a two-stage framework with interaction rationale distillation and rationale-aware recommendation, leveraging a prompt-distillation backbone (POD) and a four-task training regime. Empirical results on three Amazon domains show state-of-the-art performance for both sequential and top-N recommendations, with ablations highlighting the benefit of distilling user preferences and item attributes. The work also analyzes limitations such as hallucinations and unfaithful reasoning, and proposes directions for improved prompts and explanation generation.

Abstract

Large language model (LLM)-based recommender models that bridge users and items through textual prompts for effective semantic reasoning have gained considerable attention. However, few methods consider the underlying rationales behind interactions, such as user preferences and item attributes, limiting the reasoning capability of LLMs for recommendations. This paper proposes a rationale distillation recommender (RDRec), a compact model designed to learn rationales generated by a larger language model (LM). By leveraging rationales from reviews related to users and items, RDRec remarkably specifies their profiles for recommendations. Experiments show that RDRec achieves state-of-the-art (SOTA) performance in both top-N and sequential recommendations. Our source code is released at https://github.com/WangXFng/RDRec.

RDRec: Rationale Distillation for LLM-based Recommendation

TL;DR

RDRec tackles the gap in LLM-based recommender systems by learning interaction rationales from reviews via a larger LM and distilling them into a compact model. It introduces a two-stage framework with interaction rationale distillation and rationale-aware recommendation, leveraging a prompt-distillation backbone (POD) and a four-task training regime. Empirical results on three Amazon domains show state-of-the-art performance for both sequential and top-N recommendations, with ablations highlighting the benefit of distilling user preferences and item attributes. The work also analyzes limitations such as hallucinations and unfaithful reasoning, and proposes directions for improved prompts and explanation generation.

Abstract

Large language model (LLM)-based recommender models that bridge users and items through textual prompts for effective semantic reasoning have gained considerable attention. However, few methods consider the underlying rationales behind interactions, such as user preferences and item attributes, limiting the reasoning capability of LLMs for recommendations. This paper proposes a rationale distillation recommender (RDRec), a compact model designed to learn rationales generated by a larger language model (LM). By leveraging rationales from reviews related to users and items, RDRec remarkably specifies their profiles for recommendations. Experiments show that RDRec achieves state-of-the-art (SOTA) performance in both top-N and sequential recommendations. Our source code is released at https://github.com/WangXFng/RDRec.
Paper Structure (21 sections, 1 equation, 3 figures, 7 tables)

This paper contains 21 sections, 1 equation, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. RDRec Framework
  3. Interaction Rationale Distillation
  4. Rationale-aware Recommendation
  5. Model Optimization and Inference
  6. Experiment
  7. Experimental Setup
  8. Experimental Results
  9. Error Analysis of Sequential Recommendation
  10. In-Depth Analysis of RDRec
  11. Conclusion
  12. Appendix
  13. Experimental Details
  14. Effect of Various Sample Ratios
  15. Execution Time
  16. ...and 6 more sections

Figures (3)

  • Figure 1: Illustration of our motivation. (a) denotes the review after a purchase and (b) refers to the rationale of the purchase distilled by LLMs. (c) and (d) indicate the preference and attribute enrichment, respectively.
  • Figure 2: Illustration of rationale distillation with LLMs via the chain-of-thought (CoT) prompting.
  • Figure 3: Illustration of input and output of four tasks by RDRec in the prompt distillation setting.