Neighborhood based algorithms are one of the most common approaches to Collaborative Filtering (CF). The core element of these algorithms is similarity computation between items or users. It is reasonable to assume that some ratings of a user bear more information than others. Weighting the ratings proportional to their importance is known as feature weighting. Nevertheless in practice, none of the existing weighting schemes results in significant improvement to the quality of recommendations. In this paper, we suggest a new weighting scheme based on Matrix Factorization (MF). In our scheme, the importance of each rating is estimated by comparing the coordinates of users (items) taken from a latent feature space computed through Matrix Factorization (MF). Moreover, we review the effect of a large number of weighting schemes on item based and user based algorithms. The effect of various influential parameters is studied running extensive simulations on two versions of the Movielens dataset. We will show that, unlike the existing weighting schemes, ours can improve the performance of CF algorithms. Furthermore, their cascading capitalizes on each other's improvement.