To solve the problem of complex context and data sparsity, a new algorithm for the tensor decomposition based on context similarity of mobile user was proposed, namely UCS-TF (User-Context-Service Tensor Factorization recommendation). Multi-dimensional context similarity model was established with combining the user context similarity and confidence of similarity. Then, K-neighbor information of the target user was applied to the three-dimensional tensor decomposition, composed by user, context and mobile-service. Therefore, the predicted value of the target user was obtained, and the mobile recommendation was generated. Compared with cosine similarity method, Pearson correlation coefficient method and the improved Cosine1 model, the Mean Absolute Error (MAE) of the proposed UCS-TF algorithm was reduced by 11.1%, 10.1% and 3.2% respectively; and the P@N index of it was also significantly improved, which is better than that of the above methods. In addition, compared with Cosine1 algorithm, CARS2 algorithm and TF algorithm, UCS-TF algorithm had the smallest prediction error on 5%, 20%, 50% and 80% of data density. The experimental results indicate that the proposed UCS-TF algorithm has better performance, and the user context similarity combining with the tensor decomposition model can effectively alleviate the impact of score sparsity.