Large Language Models pour Les Systèmes de Recommandation

Abstract

With the growing importance of information personalization, it has become essential to develop recommendation systems capable of fully leveraging the richness of available textual content. This study proposes the use of Large Language Models (LLMs) in this context, through the development of a book recommendation system based on semantic similarity modeling between metadata and generated summaries. To achieve this goal, five pre-trained models from the Sentence-Transformer family were employed, all of which were fine-tuned to better adapt to the specific task context. Evaluation results showed that the all-MiniLM-L6 model achieved remarkable performance, reaching a recall score of approximately 99.69% when recommending ten books. These findings highlight the strong potential of LLMs in building intelligent, interactive, and scalable recommendation systems.