Research on AI Model of Machine Learning in Intelligent Recommendation System

In this paper, a self-supervised recommendation model VAE-CSRM based on Variational Autoencoder (VAE) and Contrastive Learning strategies is proposed. It is designed to solve key challenges in recommendation systems such as data sparsity, cold start problems, and dynamic evolution of user interests. Using the hidden variable modeling approach of VAE, SRM maps user portraits, item attributes, and contextual information into a low-dimensional latent space, effectively capturing the complex relationships between users and items. Additionally, a contrastive learning approach is incorporated to better uncover hidden relationships in user behavior sequences by strengthening the alignment of positive sample pairs and minimizing the alignment of negative pairs, thereby improving the accuracy of long-term interest prediction. The experiment was carried out on MovieLens-10M dataset, and the results showed that SRM improved the accuracy of Top-10 recommendation by 8.5% compared with the traditional matrix decomposition method (SVD++), and improved by 15.2% in new user scenarios. In the long-term interest forecasting task, NDCG@20 improved 10.1 % from the baseline model. The research in this paper provides theoretical support and technical practical reference for the engineering implementation of intelligent recommendation system. In the future, it will focus on multi-modal data fusion, model interpretability enhancement and integrated optimization of privacy protection mechanism.