With a sharp improvement in E-commerce and data, the precise rating prediction of recommended items under user preferences has been a hot research topic in the EC intelligence domain. The rating data Matrix Factorization based methods have been widely used in item rating predictions in e-commerce recommendation systems. However, “cold start” and “data sparsity” have seriously restricted the accuracy of such methods. In addition to the rating data, as side information, the massive reviews posted by users rich in semantic and emotional information express user preferences and item characteristics, and will certainly improve the accuracy of the rating prediction. Accordingly, this paper combines the deep learning for the review text and the matrix factorization method for rating data to predict the rating of the recommended items accurately. Firstly, based on the Deep Learning methods, self-attention mechanism and bi-directional RNN (Recurrent Neural Network) with the core of GRU (Gated Recurrent Unit), the deep nonlinear features of users and items are learned from review texts. Then, these features are introduced as a prior mean into the classical rating-based probability matrix factorization model to obtain the latent factor vectors of users and items with the rating of the recommended item accurately predicted. Finally, adopting MSE and MAE as the indicators, the extensive experiments conducted on four real datasets verify that the proposed model TFRMF (Topical Features Regularized Matrix Factorization) performs better than other classical counterparts. The achievements of this work will provide powerful methods and decision supports for accurate and personalized e-commerce recommendation practices.