Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enriching Semantic Profiles into Knowledge Graph for Recommender Systems Using Large Language Models

Seokho Ahn, Sungbok Shin, Young-Duk Seo

TL;DR

The paper tackles the challenge of constructing rich, transferable user profiles for recommender systems by analyzing profiling along four dimensions and showing the limitations of LLM-only and KG-only approaches. It introduces SPiKE, a hybrid model that uses LLMs to generate semantic profiles for all KG entities and then propagates these profiles through a knowledge graph, with a pairwise alignment loss to harmonize textual and graph signals. The approach is instantiated through three components—entity profile generation, profile-aware KG aggregation, and pairwise profile preference matching—and is validated on three real-world benchmarks, showing consistent gains over state-of-the-art KG- and LLM-based methods. The work demonstrates that integrating semantic profiling with graph propagation yields broader coverage and improved recommendation quality, with practical implications for scalable, knowledge-grounded recommendation systems.

Abstract

Rich and informative profiling to capture user preferences is essential for improving recommendation quality. However, there is still no consensus on how best to construct and utilize such profiles. To address this, we revisit recent profiling-based approaches in recommender systems along four dimensions: 1) knowledge base, 2) preference indicator, 3) impact range, and 4) subject. We argue that large language models (LLMs) are effective at extracting compressed rationales from diverse knowledge sources, while knowledge graphs (KGs) are better suited for propagating these profiles to extend their reach. Building on this insight, we propose a new recommendation model, called SPiKE. SPiKE consists of three core components: i) Entity profile generation, which uses LLMs to generate semantic profiles for all KG entities; ii) Profile-aware KG aggregation, which integrates these profiles into the KG; and iii) Pairwise profile preference matching, which aligns LLM- and KG-based representations during training. In experiments, we demonstrate that SPiKE consistently outperforms state-of-the-art KG- and LLM-based recommenders in real-world settings.

Enriching Semantic Profiles into Knowledge Graph for Recommender Systems Using Large Language Models

TL;DR

The paper tackles the challenge of constructing rich, transferable user profiles for recommender systems by analyzing profiling along four dimensions and showing the limitations of LLM-only and KG-only approaches. It introduces SPiKE, a hybrid model that uses LLMs to generate semantic profiles for all KG entities and then propagates these profiles through a knowledge graph, with a pairwise alignment loss to harmonize textual and graph signals. The approach is instantiated through three components—entity profile generation, profile-aware KG aggregation, and pairwise profile preference matching—and is validated on three real-world benchmarks, showing consistent gains over state-of-the-art KG- and LLM-based methods. The work demonstrates that integrating semantic profiling with graph propagation yields broader coverage and improved recommendation quality, with practical implications for scalable, knowledge-grounded recommendation systems.

Abstract

Rich and informative profiling to capture user preferences is essential for improving recommendation quality. However, there is still no consensus on how best to construct and utilize such profiles. To address this, we revisit recent profiling-based approaches in recommender systems along four dimensions: 1) knowledge base, 2) preference indicator, 3) impact range, and 4) subject. We argue that large language models (LLMs) are effective at extracting compressed rationales from diverse knowledge sources, while knowledge graphs (KGs) are better suited for propagating these profiles to extend their reach. Building on this insight, we propose a new recommendation model, called SPiKE. SPiKE consists of three core components: i) Entity profile generation, which uses LLMs to generate semantic profiles for all KG entities; ii) Profile-aware KG aggregation, which integrates these profiles into the KG; and iii) Pairwise profile preference matching, which aligns LLM- and KG-based representations during training. In experiments, we demonstrate that SPiKE consistently outperforms state-of-the-art KG- and LLM-based recommenders in real-world settings.
Paper Structure (22 sections, 14 equations, 5 figures, 9 tables)

This paper contains 22 sections, 14 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Overview
  3. Profiling in Recommendation
  4. Dimensions for Enriching Profiles
  5. Our design rationale.
  6. Our Approach for Model Selection
  7. Task Definition
  8. Methodology
  9. SPiKE Overview
  10. Entity Profile Generation Using LLMs
  11. Profile-Aware KG Aggregation
  12. Pairwise Profile Preference Matching
  13. Model Training and Inference
  14. Experimental Settings
  15. Experimental Results
  16. ...and 7 more sections

Figures (5)

  • Figure 1: Illustrative flow of SPiKE. SPiKE is built on the idea from §\ref{['sec:02_overview']} that LLMs and KGs each contribute their essential strengths (i.e., semantic profiling and structural propagation). Following a minimal-role design, SPiKE generates entity profiles using LLMs (§\ref{['sec:entity_profiling']}), integrates them into KG aggregation (§\ref{['sec:kg_aggregation']}), and applies pairwise matching (§\ref{['sec:pairwise_matching']}) during training.
  • Figure 2: Examples of entity profile generation prompts used in SPiKE for book recommendation. Each entity in the KG is profiled in the order of item, auxiliary entity, and user, incorporating relevant knowledge bases and preference indicators.
  • Figure 3: Sensitivity analysis of key SPiKE hyperparameters on Books dataset, with final selections marked by gray vertical dashed lines. All choices are made by considering performance across all @$K$ metrics and efficiency trade-offs.
  • Figure 4: Performance comparison under different graph-density settings on Books dataset, controlled by varying the interaction ratio. Each user keeps at least one interaction.
  • Figure 5: Case studies on Books dataset.