The existing social recommendation algorithms do not consider the influence of trusted users on the deep preference of target users. Aiming at this problem, a hybrid recommendation model based on deep learning is proposed, which uses the denoising auto-encoder to learn the rating preference of users and their trusted users, uses the weighted hidden layer to balance the importance of these representations for users, and effectively model potential interactions between users. On this basis, user clu- stering and personalization weights are used to distinguish the impact of different types of users on their trusted users. The experimental results on the public dataset show that the proposed method is superior to the existing social recommendation algorithms. Compared with the main recommendation models SoRec, RSTE, SocialMF, and TrustMF algorithm, the proposed algorithm significantly reduces the mean absolute error (MAE) and the root mean square error (RMSE), thus obtaining better recommendation effect effectively.