"Enabling Explainable Recommendation in E-commerce with LLM-powered Product Knowledge Graph"

Product recommendations get smarter by turning LLM's knowledge into structured graphs.

LLM-PKG introduces a novel approach to enhance e-commerce recommendations by distilling LLM knowledge into product knowledge graphs, enabling explainable recommendations while maintaining real-time performance and mitigating hallucination risks through rigorous validation.

-----

https://arxiv.org/abs/2412.01837v1

🔍 Original Problem:

Modern e-commerce recommendation systems face challenges in providing explainable recommendations while maintaining real-time performance. Direct LLM integration is impractical due to response time constraints.

-----

🛠️ Solution in this Paper:

→ The paper introduces LLM-PKG, which first builds a Product Knowledge Graph using LLM-generated recommendations and rationales.

→ The system employs rigorous validation through LLM-based scoring and pruning to ensure reliability.

→ Product mappings are done using vector search with fine-tuned BERT embeddings.

→ The final graph is cached in key-value stores for real-time serving.

-----

💡 Key Insights:

→ LLMs can effectively capture product relationships and user intentions that are difficult to mine from traditional e-commerce data

→ Knowledge graph construction can be automated using LLM while maintaining quality through validation

→ Caching strategies enable real-time recommendation serving at scale

-----

📊 Results:

→ 5.19% increase in clicks

→ 7.59% improvement in transactions

→ 8.56% growth in Gross Merchandise Bought

→ 10.84% increase in ads revenue