Enhancing Recommender Systems Through Multimodal Large Language Models and Neural Matrix Factorization
Yizhuo Jiang · 2024
This This study proposes a new recommendation system model that integrates multimodal data, large language models (LLM), and neural matrix factorization techniques. By using Vision Transformer (ViT) and the BERT model, we deeply mine the image and text information of users and products, and combine unstructured data with neural network technology for evaluation prediction and recommendation, showing excellent prediction performance. This study also verified the performance advantages of this method compared with traditional recommendation system methods through various ablation experiments. The results highlight the importance of text data in capturing user preferences and improving recommendation accuracy. In addition, the study found that in practical applications, more detailed segmentation processing of visual data is needed to accurately evaluate the specific contribution of visual data to multi-modal recommendation systems. These findings provide important theoretical and practical guidance for further optimizing recommendation systems.