Online social networks and e-commerce platforms generate massive data that facilitates the modeling of user preferences. However, the data often contain noise and sparsity, which complicates accurate recommendation generation.Social Semantic Mining and Denoising (SSMD) model that combines social trust information, semantic similarity, and deep learning–based denoising to enhance recommendation accuracy. The model uses graph representation and denoising autoencoders to filter irrelevant data and learn meaningful user–item relationships. Experimental evaluation on Ciao dataset demonstrates that the model achieves better accuracy and robustness compared to traditional recommendation algorithms. Despite its improved recommendation performance, the SSMD model relies on centralized data collection, which introduces potential privacy risks. To address this drawback, the proposed system introduces a privacy-preserving enhancement that protects sensitive user information during the recommendation process while maintaining high accuracy and reliability. This improvement ensures secure, efficient, and trustworthy personalized recommendations.