Collaborative Filtering is effective to provide customers with personalized recommendations by analyzing the purchase pattens. Matrix factorization, e.g. Singular Value Decomposition, is another successful technique in recommendation system. We implemented Singular Value Decomposition algorithm to achieve the least total squared errors. Based on the result, item-feature Collaborative Filtering was further designed and implemented in P-Tree data structure. The purpose of item-feature Collaborative Filtering is to achieve the local optimization. Our experiment on Netix Prize data suggests Singular Value Decomposition is good at global optimization and item-feature Collaborative Filtering is good at local optimization. Blending of two algorithms achieves better RMSE score.