Research on application and optimization of multimodal big data analysis in intelligent recommendation system

Multimodal big data combines various types of information, such as text, image, audio and video, which provides a richer data source for the recommendation system and can fully understand the user's needs and content characteristics. However, how to efficiently fuse multimodal data and optimize the performance of recommendation system still faces many challenges. Firstly, this paper introduces the basic principle and workflow of multimodal generation model, and expounds its advantages in improving the diversity and interpretation of recommendations. Then, aiming at the key factors of performance optimization of multimodal recommendation system, optimization strategies such as dynamic modal weighted fusion, mixed parallel training and multimodal hash recommendation algorithm are proposed. The experimental results show that this method performs well on Amazon, TikTok and self-built e-commerce data sets, which significantly improves the recommendation accuracy, system efficiency and multimodal fusion quality. The experimental results show that the dynamic weighted fusion mechanism effectively improves the correlation between modes, the hybrid parallel strategy greatly improves the training efficiency, and the hash recommendation algorithm significantly reduces the response time and memory occupation while ensuring the recommendation quality. The research in this paper provides new ideas and methods for the application of multimodal big data analysis in intelligent recommendation system, which has important theoretical significance and practical value.