Research on E-Commerce Long-Tail Product Recommendation Mechanism Based on Large-Scale Language Models

As e-commerce platforms continue to extend their product catalogs, ensuring accurate recommendation of long-tail items has become a major objective. This is crucial because, in this way, user experience and platform revenue can be significantly enhanced. A common struggle for these systems is the long-tail problem, where the full breadth of data drives the high level of sparsity and the cold start occurs. Consequently, traditional recommendation algorithms have many issues when recommending long-tail items due to lack of available data. Our work reflects a long-tail product recommendation mechanism that represents a combination of the product text descriptions and the user behavior sequences based on a large-scale language model (LLM) as the first proposal of this paper. This mechanism starts with the application of a pre-trained LLM to convert multimodal texts, e.g., product titles, detailed descriptions, and user reviews, into word embedding formats that are semantically meaningful for those particular entities of the text. This approach is referred to as the semantic visor. Going further, we also set up an attention-based user intent encoder to encode user's latent interests in long-tail products in a collaboration pattern. This way, a hybrid ranking model was designed to combine semantic similarity scores, collaborative filtering scores, and LLM-generated recommendation candidates. Through them, we provide extensive empirical results on a real-world e-commerce dataset, thus illustrating that our method surpasses the static baseline models in recalling, hitting, and using diverse users by 12%, 9%, and 15%, respectively, thereby more effectively conveys exposure and purchase rates of long-tail products. Therefore, our work demonstrates the significant benefits of LLMs for interpreting the product content and user intent, suggesting an innovative technical direction for e-commerce recommendation systems.