Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM-Powered Explanations: Unraveling Recommendations Through Subgraph Reasoning

Guangsi Shi, Xiaofeng Deng, Linhao Luo, Lijuan Xia, Lei Bao, Bei Ye, Fei Du, Shirui Pan, Yuxiao Li

TL;DR

The paper tackles the difficulty of explaining recommendations when knowledge graphs are noisy or incomplete by introducing LLM-SRR, a framework that reconstructs KG from user reviews, builds a user-centered subgraph through attention-based diffusion, and generates post-hoc explanations with LLMs. The core contributions are a novel LLM-powered KG reconstruction and explanation pipeline, an attentive diffusion subgraph reasoning module for robust user representations, and a scoring mechanism that fuses subgraph context with item profiles to produce interpretable recommendations. Empirically, LLM-SRR achieves state-of-the-art or strong performance gains on four real-world datasets and demonstrates practical utility through a METC cross-selling case study with explicit explanations that reduce recommendation hallucinations. The work underscores the practical value of integrating LLMs with knowledge graphs for transparent, accurate, and business-relevant recommendations, with notable implications for marketing analytics and cross-channel strategies.

Abstract

Recommender systems are pivotal in enhancing user experiences across various web applications by analyzing the complicated relationships between users and items. Knowledge graphs(KGs) have been widely used to enhance the performance of recommender systems. However, KGs are known to be noisy and incomplete, which are hard to provide reliable explanations for recommendation results. An explainable recommender system is crucial for the product development and subsequent decision-making. To address these challenges, we introduce a novel recommender that synergies Large Language Models (LLMs) and KGs to enhance the recommendation and provide interpretable results. Specifically, we first harness the power of LLMs to augment KG reconstruction. LLMs comprehend and decompose user reviews into new triples that are added into KG. In this way, we can enrich KGs with explainable paths that express user preferences. To enhance the recommendation on augmented KGs, we introduce a novel subgraph reasoning module that effectively measures the importance of nodes and discovers reasoning for recommendation. Finally, these reasoning paths are fed into the LLMs to generate interpretable explanations of the recommendation results. Our approach significantly enhances both the effectiveness and interpretability of recommender systems, especially in cross-selling scenarios where traditional methods falter. The effectiveness of our approach has been rigorously tested on four open real-world datasets, with our methods demonstrating a superior performance over contemporary state-of-the-art techniques by an average improvement of 12%. The application of our model in a multinational engineering and technology company cross-selling recommendation system further underscores its practical utility and potential to redefine recommendation practices through improved accuracy and user trust.

LLM-Powered Explanations: Unraveling Recommendations Through Subgraph Reasoning

TL;DR

The paper tackles the difficulty of explaining recommendations when knowledge graphs are noisy or incomplete by introducing LLM-SRR, a framework that reconstructs KG from user reviews, builds a user-centered subgraph through attention-based diffusion, and generates post-hoc explanations with LLMs. The core contributions are a novel LLM-powered KG reconstruction and explanation pipeline, an attentive diffusion subgraph reasoning module for robust user representations, and a scoring mechanism that fuses subgraph context with item profiles to produce interpretable recommendations. Empirically, LLM-SRR achieves state-of-the-art or strong performance gains on four real-world datasets and demonstrates practical utility through a METC cross-selling case study with explicit explanations that reduce recommendation hallucinations. The work underscores the practical value of integrating LLMs with knowledge graphs for transparent, accurate, and business-relevant recommendations, with notable implications for marketing analytics and cross-channel strategies.

Abstract

Recommender systems are pivotal in enhancing user experiences across various web applications by analyzing the complicated relationships between users and items. Knowledge graphs(KGs) have been widely used to enhance the performance of recommender systems. However, KGs are known to be noisy and incomplete, which are hard to provide reliable explanations for recommendation results. An explainable recommender system is crucial for the product development and subsequent decision-making. To address these challenges, we introduce a novel recommender that synergies Large Language Models (LLMs) and KGs to enhance the recommendation and provide interpretable results. Specifically, we first harness the power of LLMs to augment KG reconstruction. LLMs comprehend and decompose user reviews into new triples that are added into KG. In this way, we can enrich KGs with explainable paths that express user preferences. To enhance the recommendation on augmented KGs, we introduce a novel subgraph reasoning module that effectively measures the importance of nodes and discovers reasoning for recommendation. Finally, these reasoning paths are fed into the LLMs to generate interpretable explanations of the recommendation results. Our approach significantly enhances both the effectiveness and interpretability of recommender systems, especially in cross-selling scenarios where traditional methods falter. The effectiveness of our approach has been rigorously tested on four open real-world datasets, with our methods demonstrating a superior performance over contemporary state-of-the-art techniques by an average improvement of 12%. The application of our model in a multinational engineering and technology company cross-selling recommendation system further underscores its practical utility and potential to redefine recommendation practices through improved accuracy and user trust.
Paper Structure (27 sections, 10 equations, 3 figures, 7 tables)

This paper contains 27 sections, 10 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Explainable Recommendation System
  4. Large Language Models and Knowledge Graphs Combination
  5. Subgraph Reasoning
  6. Method
  7. Preliminary
  8. KG Reconstruction and Explanation Generation
  9. LLMs for Information Extraction and Injection
  10. LLMs for Post-hoc Explanation Generation
  11. Attentive Diffusion Subgraph Reasoning
  12. Recommendation Scoring
  13. Experiments
  14. Experiment Settings
  15. Datasets
  16. ...and 12 more sections

Figures (3)

  • Figure 1: Main scenario and task: Arrows of different colors in KGs represent different channels (e-commerce platforms). Different products may not be sold in the same channel. The system needs cross channels to recommend and provide human-acceptable, reliable explanation descriptions.
  • Figure 2: Framework of LLM-SRR: I. Text information is extracted by the LLM and injected into the original knowledge graph by pre-define rules {u is user, p is the property, i is item, r is relation}; II. The attention score is calculated by the neighbours in different diffusion layers and subgraph could be generated. III. The final recommendation score is computed by the similarity function between the user and item, where a explanation path can be generated by LLM in this component.
  • Figure 3: Real cases discovered by our model, each containing a subgraph which end nodes is predicted item by recommendation model. u1 – u4 : METC users, i1 – i4 : Auto air filter related products, R_R1: “review: reliable”, R_R2 : “review: no smell”, O_AW: “order: auto wiper”, S_HD: “search: Home Decoration ”, C_1: “channel: Channel 1 auto parts flagship store”, P_CO: “profile: car owner” , C_2: “channel: Channel 2”, P_PU: “profile: premium user”, C_3: “channel: Channel 3 ”, C_4: “channel: Channel 4”.