Empowering Classification of Retail Products with Large Language Models: Multi-level, Multi-label and Multi-consumer Perspectives

Modern retail systems require precise classification to enhance supply chain efficiency, inventory management, and consumer personalization. However, traditional methods struggle with reliance on annotated data, domain adaptability, and dynamic changes in product categories. This study proposes a novel framework leveraging large language models (LLMs) for retail products classification, addressing the limitations of traditional machine learning and deep learning approaches. The study adopts BERT-style LLMs (e.g., BERT and RoBERTa) and GPT-style LLMs (e.g., GPT, Llama, and DeepSeek) to conduct comparative experiments from both product-oriented and consumer-oriented perspectives, specifically examining three dimensions: multi-level, multi-label, and multi-consumer classification. Experiments across a four-tier classification task demonstrate that fine-tuned GPT-style LLMs like Llama3-Chinese-ft outperform BERT-style LLMs, achieving 76.83% accuracy at the fourth level, while maintaining over 90% accuracy at coarse-grained levels; in multi-label classification, RoBERTa and the fine-tuned Llama3 achieve the highest accuracy; and in multi-consumer classification, GPT-4o mini demonstrates the strongest predictive capability. The results highlight GPT-style LLMs’ superior adaptability to evolving retail demands, and offer retailers and e-commerce platforms scalable solutions for reducing labor and technical costs and improving consumer satisfaction.