The aim of collaborative filtering is to make predictions for active user by utilizing the rating information of likeminded users in a historical database. But previous methods suffered from problems: sparsity, scalability, rating bias etc. To alleviate those problems, this paper presents a novel approach—Iterative Rating Filling Collaborative Filtering algorithm (IRFCF). Firstly, based on the idea of iterative reinforcement process, object-pair similarity is computed iteratively, and average rating and rating range are introduced to normalize ratings in order to alleviate rating bias problem. Then missing ratings are filled from user and item clusters through iterative clustering process to solve the sparsity and scalability problems. Finally, the nearest neighbors in the set of top clusters are selected to generate predictions for active user. Experimental results have shown that our proposed collaborative filtering approach can provide better performance than other collaborative filtering algorithms.