Enhancing E-Commerce Product Recommendations Using LLMs and Transformer-Based Deep Learning Architectures

The integration of large language models with deep learning architectures brought an evolutionary revolution in the context of product recommendation systems because of several limitations of traditional methods, such as collaborative and content-based filtering. In this paper, a novel framework for product recommendation is proposed by integrating large language models jointly with deep learning. The contribution of large language models is that it adds semantic understanding capability to the predictive power provided through neural networks. Domain ontologies will be used in this hybrid model to enhance the accuracy and personalization of recommendations, considering complex user preferences and product attributes in e-commerce platforms. It takes a state-of-the-art pre-trained LLM, such as Llama-3, as input and generates personalized embeddings of users based on history, item descriptions, and contextual information. In this work, the Transformer architecture has been used in re-fining and ranking the products for relevance using attention mechanisms that select the most important features in each recommendation task. Besides, knowledge distillation will be used to conduct the small and efficient student model training process. The distilled model receives soft predictions that involve the teacher LLM, which greatly reduces computational overhead but preserves high recommendation accuracy. Eventually, the framework will be evaluated on a real-world e-commerce dataset to explore how it increases the click-through rate, purchase rate, and user's engagement compared to the traditional systems.