AI Agent-Driven Framework for Automated Product Knowledge Graph Construction in E-Commerce
By Dimitar Peshevski, Riste Stojanov, Dimitar Trajanov
DOI https://doi.org/10.48550/arXiv.2511.11017
Abstract
The rapid growth of e-commerce platforms has led to an overflow of unstructured product data, which poses challenges for information retrieval, recommendation systems, and data analytics. Knowledge Graphs, which are structured representations of data, are crucial for organizing this information. However, constructing product-specific Knowledge Graphs is often a manual and complex task. This paper presents an automated framework powered by Artificial Intelligence agents to create Knowledge Graphs using unstructured product descriptions. The proposed method is divided into three stages—ontology creation and expansion, ontology refinement, and Knowledge Graph population—utilizing Large Language Models. The evaluation on a dataset of air conditioner descriptions shows the framework’s high effectiveness, achieving over 97% property coverage and demonstrating its scalability for intelligent product data integration.
Introduction
E-commerce and retail platforms are generating significant amounts of unstructured product information, such as descriptions, specifications, and reviews. To utilize this data for applications like product recommendations and analytics, it must be structured into a machine-readable form. Knowledge Graphs help achieve this by representing entities (like products) and their relationships in a graph format.
Despite their utility, creating Knowledge Graphs is typically a manual and labor-intensive process that requires domain-specific knowledge. This paper introduces an automated framework utilizing AI agents to construct Knowledge Graphs specifically for product domains. By employing Large Language Models, the framework automates the creation and refinement of product ontologies and directly generates Knowledge Graphs from product descriptions.
Methodology
The framework consists of three major stages:
Ontology Creation and Expansion: The process starts by sampling product descriptions to identify essential ontology elements, like product classes and attributes, and organizing them into a structured format. This stage iteratively incorporates more product samples to expand the ontology by adding new classes or properties.
Ontology Refinement: This stage enhances the initial ontology using the capabilities of Large Language Models. It addresses any issues of redundancy, generality, or clarity within the ontology to improve its usability and flexibility across different product types.
Knowledge Graph Population: The last stage involves populating the Knowledge Graph with specific product data derived from the descriptions. This step generates RDF (Resource Description Framework) triples, which represent the relationships and attributes of products. The framework ensures the accurate representation of data without generating incorrect information.
Evaluation
The authors evaluated the framework on a dataset consisting of 291 product descriptions for air conditioners. The evaluation focused on three key areas:
- Ontology Coverage: It measured how completely the ontology captured product classes, attributes, and relationships.
- Ontology Quality: This involved a qualitative assessment of coherence, generality, and usability.
- Knowledge Graph Population: They assessed the number of generated RDF triples and how many properties from the ontology were instantiated in the Knowledge Graph.
The results showed that the framework constructed a modular and comprehensive ontology covering 42 classes and 69 properties. It processed 282 of the 291 descriptions, achieving a property coverage of 97.1%, demonstrating the framework’s effectiveness and robustness.
Conclusion and Future Work
The proposed AI agent-driven framework represents a significant advancement in automating the construction of Knowledge Graphs for e-commerce. It effectively eliminates the need for manual processes, allowing faster adaptability to new products.
Future enhancements could include integrating various types of data (like images and user reviews) to enrich the Knowledge Graph further. Additionally, efforts could be directed towards improving the accuracy of data extraction and expanding the framework’s application to other domains, such as finance or healthcare.
The framework promises to lay a strong foundation for advanced applications in e-commerce, such as improved product recommendations and search functionality.