A collaborative filtering model of non-negative matrix factorization based on diversity of latent factors is effective in predicting high dimension and sparse matrix. This model can recommend personally and utilize the feedback information from other similar users effectively. However, it has the disadvantage of low prediction accuracy. Due to the diversity of ratings for users or items under different circumstances, a collaborative filtering model based on non-negative matrix factorization was proposed. The model considered the diversity of latent characteristic matrices for users and items. The single-element non-negative multiplication update rule and the principle of alternate direction method were integrated in the training of the model, which not only guaranteed the non-negativity of the target matrix, but also improved the convergence rate of the models. Finally, experiments were carried out on real industrial data sets. The experimental results show that the prediction accuracy of the proposed model is higher than that of the classical non-negative matrix factorization model.