Dynamic Graph-Retrieval Augmented Generation for Reliable and Explainable Consumer Electronics Recommendation

Jaesung Park, DAEKI KIM, JIMAN KIM, jiyoon yoon, Seongwoon Jung

Keywords: LLM, Graph Knowledge, RAG, Vector DB, Graph DB, Consumer Electronics, Recommendation

TLDR: The paper introduces the Dynamic Graph-Retrieval Augmented Generation framework, which integrates a knowledge graph and RAG to overcome LLM limitations and significantly improve recommender system accuracy and reliability.

Abstract: Although Large Language Models (LLMs) have expanded the capabilities of recommender systems, they are hindered by inherent limitations, including a propensity for factual hallucination and reliance on outdated domain knowledge. These constraints pose significant challenges in contexts requiring high-fidelity recommendations, such as consumer electronics purchases where precise, current specifications are critical. To address these issues, this paper proposes a novel framework termed Dynamic Graph-Retrieval Augmented Generation (Graph-RAG), which integrates LLMs, knowledge graphs, and RAG techniques within a multi-agent architecture. The framework dynamically deciphers complex user purchase intents and prioritizes decision-critical factors through a modular communication protocol that enables cross-agent collaboration, a `Specification Vector Index' that resolves semantic disparities between natural language queries and technical attributes, and a graph-based dynamic retrieval engine that facilitates fact-grounded reasoning. Empirical validation forms a pivotal contribution of this work, with rigorous experimental verification confirming the system's efficacy in minimizing hallucinations through structured knowledge grounding. Quantitative metrics demonstrate statistically significant improvements in recommendation accuracy, such as a +22.7% increase in precision, and reliability, while traceable decision pathways enhance operational transparency. This research delivers a foundational architecture for a possible practical recommender systems, validated through real-world deployment scenarios and test dataset, and establishes a benchmark for empirically substantiated innovation in AI-driven recommendation frameworks. By bridging theoretical innovation and practical deployment, this study marks a critical advancement in the field, offering both a new methodology and robust evidence of its real-world applicability.

Submission Number: 20