Xây dựng hệ thống khuyến nghị ứng dụng ngôn ngữ lớn

Tóm tắt

In the context of the rapid development of artificial intelligence and the explosion of big
 data, recommendation systems have become an indispensable tool for personalizing user
 experiences in e-commerce, online entertainment, and education, as well as reducing
 information overload.
 Recommendation systems using traditional deep learning models have achieved
 many positive results in recent times. However, these systems still face challenges in
 inference with limited or no data, dealing with new users or new items, interpretability,
 scalability, natural language interaction, and understanding real-world context.
 With the strong development of large language models (LLMs) and their humanlike reasoning abilities in language tasks, this thesis explores the applicability of LLMs
 in enhancing the performance of a recommendation system with natural language conversational data. This thesis proposes a new approach, inheriting and improving upon
 the research of Paul Covington et al. (2016) [19], Karl Higley et al. (2022) [7], Tingting
 Liang et al. (2024) [32], and Zhankui He et al. (2023) [18]. The proposed approach involves constructing a multi-step recommendation system that leverages LLMs through
 two main modules: LLM as Retrieval and LLM as Reranking, with seven processing
 steps. This architecture utilizes the natural language understanding capabilities of LLMs
 to capture user preferences, behavior, and context in a more comprehensive manner, addressing data scarcity issues while avoiding the resource-intensive process of fine-tuning
 LLMs.
 Based on the research results, the proposed architecture demonstrates superior performance compared to previous studies using traditional deep learning models applied
 to data and model training for customer recommendations. Specifically, the Recall@50
 metric reaches 0.405, which is higher than the range of 0.32 to 0.376 reported in previous studies. For research using LLMs, the proposed architecture also shows advantages,
 achieving 0.405 in Recall@50 and 0.12 in Recall@5, outperforming studies that only relied on information extraction from LLMs without data training.
 However, these results have not yet reached the level of effectiveness observed in
 studies that applied fine-tuning and training of LLMs on experimental datasets, with results of 0.493 for Recall@50 and 0.194 for Recall@5. Nevertheless, the proposed architecture still holds great potential, thanks to its effective utilization of existing knowledge
 from LLMs without the need for complex training and fine-tuning processes, significantly saving resources.
 This thesis contributes knowledge on the application of LLMs in building a personalized recommendation system and provides a solid foundation for developing more
 intelligent systems in the future.