Aiming at the typical problem of predicting users' evaluation of items in recommender systems, we propose a neural matrix factorization model based on latent factor learning. Based on the classical latent factor model, the model utilizes the representation ability of deep learning, adds auxiliary features and cross features in the process of vectorization to improve the abstract expression ability of the model to the latent factor vector, and realizes the improvement of the model by auxiliary feature information. Finally, compared with the existing deep learning models, the prediction results are significantly improved in RMSE and MAE metrics.