Enabling Explainable Recommendation in E-commerce with LLM-powered Product Knowledge Graph (2412.01837v1)

Abstract: How to leverage LLM's superior capability in e-commerce recommendation has been a hot topic. In this paper, we propose LLM-PKG, an efficient approach that distills the knowledge of LLMs into product knowledge graph (PKG) and then applies PKG to provide explainable recommendations. Specifically, we first build PKG by feeding curated prompts to LLM, and then map LLM response to real enterprise products. To mitigate the risks associated with LLM hallucination, we employ rigorous evaluation and pruning methods to ensure the reliability and availability of the KG. Through an A/B test conducted on an e-commerce website, we demonstrate the effectiveness of LLM-PKG in driving user engagements and transactions significantly.

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

The paper "Enabling Explainable Recommendation in E-commerce with LLM-powered Product Knowledge Graph" introduces an innovative framework termed LLM-PKG that aims to enhance the recommendation capabilities within e-commerce platforms using LLMs. The work represents a significant advancement in leveraging LLMs for building robust Product Knowledge Graphs (PKGs) that can deliver not only accurate but also explainable recommendations.

Summary of the Approach

The primary focus of the proposed LLM-PKG framework is to distill the knowledge encapsulated in LLMs into a structured PKG, which then serves as a foundation for generating explainable recommendations. The framework consists of two main modules: the offline construction of the PKG and the online serving of recommendations.

Offline Construction: The process begins with the generation of an initial PKG through curated prompts fed to an LLM. The prompts are designed to exploit LLM's understanding of user behavior and product relations in the e-commerce domain. The responses from the LLM are systematically composed into RDF (Resource Description Framework) triples, which form the structure of the PKG. To ensure the reliability of the PKG and to mitigate issues like LLM hallucinations, a validation and refinement process is employed. During this stage, the LLM evaluates nodes and edges for accuracy, guiding a data-driven pruning and refinement process.

Product Mapping: The transition from the abstract LLM-generated concepts to tangible products in the inventory is achieved through a vector search approach. By mapping the PKG nodes to the enterprise's product catalog, the theoretical constructs of LLM are seamlessly integrated into actionable e-commerce scenarios.

Online Serving: The online module involves using LLM-PKG for real-time recommendations, employing a caching strategy for optimal performance. The framework supports both item-based and user-centric recommendation strategies, allowing it to deliver relevant product recommendations aligned with user preferences.

Key Experimental Results

The paper presents empirical evidence of the framework’s efficacy through an A/B test conducted on a live e-commerce platform. The experiment, focusing mainly on the sneaker segment, demonstrated significant improvements across various user engagement metrics when employing LLM-PKG-based recommendations as compared to traditional methods. Notable results include an increase in clicks by 5.19% and transactions by 7.59%, which reflects the model's strong performance in enhancing both user interaction and conversion rates.

Theoretical and Practical Implications

The proposed LLM-PKG framework successfully marries the intuitive language understanding capacities of LLMs with the structured and explainable nature of knowledge graphs. This approach addresses several challenges in contemporary recommender systems, most notably the need for transparency and the ability to provide justifiable recommendations. Such capabilities are not only pivotal for enhancing user trust and satisfaction but also crucial in contexts where regulatory requirements for recommendation transparency are stringent.

Speculations and Future Directions

The implementation of LLM-PKG opens several avenues for further research and application. While the current work focuses on item-based recommendations, expanding the approach to user-centric scenarios can unveil more sophisticated personalization strategies. Additionally, further exploration into optimizing the refinement process for the PKG using more granular enterprise data could enhance the model's precision and explainability further. The intersection of LLMs and KGs as proposed in this framework holds potential for addressing the inherent trade-offs between recommendation accuracy and transparency, providing a viable blueprint for next-generation recommender systems. As the field continues to evolve, integrating real-time feedback loops into the PKG construction and refinement process could also provide a dynamic method to adapt to changing user preferences and trends.