As one of the most widely used algorithms in recommender system, the traditional collaborative filtering algorithm faces serious data sparseness problem in the big data trend, which leads to the ineffective in nearest neighbor selection, and restricts the performance of the algorithm. To address this problem, this paper proposes a collaborative filtering algorithm based on user partial feature(UPCF). In our method, it first rates the missing values based on rating bias and item popularity; and then clusters the items in the filled matrix with a K-means clustering algorithm of meliorated initial center. At last, it uses the user-based collaborative filtering algorithm with the user feature in item class to get the recommendations. The MAE measures on the MovieLens dataset shows that compared with the current popular algorithms, the performance of our UPCF algorithm improves about 10% without any increase of algorithm complexity.