Recommendation systems are designed to provide personalized predictions for items that are most appealing to individual customers. Among various types of recommendation algorithms, k-nearest neighbor based collaborative filtering algorithm attracts tremendous attention and are widely used in practice. However, the k-nearest neighbor scheme can only capture the local relationship among users and the uniform neighborhood size is also not suitable to represent the underlying data structure. In this paper, we leverage emerging graph signal processing (GSP) theory to construct sparse yet high quality graph to enhance the solution quality and efficiency of collaborative filtering algorithm. Experimental results show that our method outperforms k-NN based collaborative filtering algorithm by a large margin on the benchmark data set.