This paper proposes a two-phase rating predicting framework that fuses co-clustering and non-negative matrix factorization method. First, it uses a novel co-clustering method (BlockClust) to divide the raw rating matrix into clusters much smaller than the original matrix. Then it employs weighted non-negative matrix factorization algorithm to predict the unknown ratings. In virtue of co-clustering preprocessing, this method achieves a higher predicting accuracy and efficiency on these low-dimensional and homogeneous sub-matrices. Moreover, it proposes three update schemes for the corresponding update scenarios in recommender systems. Finally, the proposed method is implemented together with seven types of related CF (collaborative filtering) methods. The comparisons show the efficiency of the proposed method and its potential in large real-time recommender systems.