Poor recommendation quality is one major challenge in collaborative filtering recommender systems. Sparsity of source data sets is one major reason causing the poor quality. The popular singular value decomposition techniques and the agent-based methods to a certain extent are able to alleviate this issue. But at the same time they also introduce new problems. To reduce sparsity, a novel collaborative filtering algorithm is designed, which firstly selects users whose non-null ratings intersect the most as candidates of nearest neighbors, and then builds up backpropagation neural networks to predict values of the null ratings in the candidates. Experiments are conducted based on standard dataset. The results show that this methodology is able to increase the accuracy of the predicted values, resulting in improving recommendation quality of the collaborative filtering recommendation algorithm.