It is a popular way that makes use of users' rating data to recommend products or items to users. Currently, more and more users have contributed their reviews to recommender system for better online shopping experiences. Researchers have become interested in using review texts as extra information to improve recommendation quality. It is argued that reviews written by a user implicitly represent his/her preferences. In this study, a preference guidance recommendation approach is proposed that simultaneously learns latent factors from rating data and latent topics from review texts. More specifically, the learned latent topics are assumed to be positively correlated with both of the corresponding user factors and item factors, which can further improve the accuracy of recommendation prediction. The proposed approach has two advantages. One is that in order to capture such a dependent correlation, a transformation function is used for simultaneously learning latent features, i.e., latent factors and latent topics. The other is that the predicted ratings of items are influenced by the implicit tastes of users, i.e., the latent topics from review texts. Experiments are conducted on the data from Amazon consisting of 28 categories. Experimental results show that the proposed approach obtains 3.32% improvement than the recent TopicMF approach in some category dataset and the average improvement is 0.92% in terms of mean square error.