Leveraging Large Language Models for Context-Aware Product Discovery in E-commerce Search Systems

Abstract

This study presents a new way to improve product discovery in e-commerce research using large-scale language models (LLMs) for content-aware instruction. We propose a new architecture integrating LLMs with tensor factorization techniques to capture user-object-content interactions. Our system employs a multi-faceted context representation, incorporating user demographics, session behavior, and temporal factors. The LLM component facilitates a deep semantic understanding of user queries and product descriptions, enabling more nuanced query expansion and improved matching. We introduce a context-aware ranking algorithm that combines traditional IR features with LLM-generated semantic signals. Extensive testing of large-scale e-commerce data shows the superiority of our method over the state-of-the-art basis, with an improvement of 10.1% in Average Precision and 7.8% in Normalized Discounted Cumulative Gain@10. The system has shown to be particularly effective in solving the cold start problem, with a 22.3% improvement in NDCG@10 for new users. Analysis of user engagement metrics shows significant improvement across multiple products, with an overall 18.7% increase in conversions. Scalability tests confirm the system can handle large volumes while maintaining a 100ms response time. This research contributes to the advancement of personal e-commerce research, providing insight into the effective integration of LLMs and content-aware strategies for product development. Discover.

Keywords: E-commerce search, Large language models, Context-aware recommendation, Personalization