Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MADREC: A Multi-Aspect Driven LLM Agent for Explainable and Adaptive Recommendation

Jiin Park, Misuk Kim

TL;DR

MADRec tackles the challenge of scalable, explainable, and personalized recommendations by coupling unsupervised multi-aspect profile extraction from reviews with an autonomous LLM-based agent. The architecture combines memory, specialized tools (Aspect Extraction, Aspect Summary, Re-Ranking), and a Self-Feedback loop to iteratively refine recommendations and explanations across direct, sequential, and explanation-generation tasks. Empirical results across three domains demonstrate superior accuracy and enhanced explainability relative to traditional and other LLM-based baselines, with human evaluators rating the explanations as more persuasive. This approach offers a practical path toward adaptable, domain-agnostic recommender systems that can justify recommendations through structured, aspect-based reasoning.

Abstract

Recent attempts to integrate large language models (LLMs) into recommender systems have gained momentum, but most remain limited to simple text generation or static prompt-based inference, failing to capture the complexity of user preferences and real-world interactions. This study proposes the Multi-Aspect Driven LLM Agent MADRec, an autonomous LLM-based recommender that constructs user and item profiles by unsupervised extraction of multi-aspect information from reviews and performs direct recommendation, sequential recommendation, and explanation generation. MADRec generates structured profiles via aspect-category-based summarization and applies Re-Ranking to construct high-density inputs. When the ground-truth item is missing from the output, the Self-Feedback mechanism dynamically adjusts the inference criteria. Experiments across multiple domains show that MADRec outperforms traditional and LLM-based baselines in both precision and explainability, with human evaluation further confirming the persuasiveness of the generated explanations.

MADREC: A Multi-Aspect Driven LLM Agent for Explainable and Adaptive Recommendation

TL;DR

MADRec tackles the challenge of scalable, explainable, and personalized recommendations by coupling unsupervised multi-aspect profile extraction from reviews with an autonomous LLM-based agent. The architecture combines memory, specialized tools (Aspect Extraction, Aspect Summary, Re-Ranking), and a Self-Feedback loop to iteratively refine recommendations and explanations across direct, sequential, and explanation-generation tasks. Empirical results across three domains demonstrate superior accuracy and enhanced explainability relative to traditional and other LLM-based baselines, with human evaluators rating the explanations as more persuasive. This approach offers a practical path toward adaptable, domain-agnostic recommender systems that can justify recommendations through structured, aspect-based reasoning.

Abstract

Recent attempts to integrate large language models (LLMs) into recommender systems have gained momentum, but most remain limited to simple text generation or static prompt-based inference, failing to capture the complexity of user preferences and real-world interactions. This study proposes the Multi-Aspect Driven LLM Agent MADRec, an autonomous LLM-based recommender that constructs user and item profiles by unsupervised extraction of multi-aspect information from reviews and performs direct recommendation, sequential recommendation, and explanation generation. MADRec generates structured profiles via aspect-category-based summarization and applies Re-Ranking to construct high-density inputs. When the ground-truth item is missing from the output, the Self-Feedback mechanism dynamically adjusts the inference criteria. Experiments across multiple domains show that MADRec outperforms traditional and LLM-based baselines in both precision and explainability, with human evaluation further confirming the persuasiveness of the generated explanations.
Paper Structure (26 sections, 1 equation, 9 figures, 9 tables)

This paper contains 26 sections, 1 equation, 9 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. MADRec Framework
  4. Memory
  5. Tools
  6. Tasks
  7. Self-Feedback
  8. MADRec-Based Recommendation Pipeline
  9. Experimental Setup
  10. Datasets
  11. Evaluation Metrics
  12. Baselines
  13. Training Details
  14. Experimental Results
  15. Results on Recommendation Tasks
  16. ...and 11 more sections

Figures (9)

  • Figure 1: Multi-aspect user profiles and explainable recommendations grounded in aspect-based reasoning.
  • Figure 2: The structure of the MADRec framework. The system consists of Memory, Tools, Tasks, and Self-Feedback.
  • Figure 3: Examples of user and item profiles constructed with our aspect-based framework. The texts highlighted in teal indicate aspect categories, and the following sentences are the summary statements generated for each category.
  • Figure 4: Overall pipeline of the proposed MADRec framework. The system proceeds in four stages: (1) Generating a Multi-Aspect User Profile, (2) Multi Factor Re-Ranking, (3) Tasks, and (4) Self-Feedback.
  • Figure 5: Performance comparison (HR@10) across Beauty, Sports, and Toys domains for four model variants and ChatGPT. RR + SF (ours) consistently outperforms all baselines, while ChatGPT exhibits limited effectiveness, particularly in sequential recommendation.
  • ...and 4 more figures