The traditional user-based collaborative filtering (CF) algorithms often suffer from two important problems: Scalability and sparsity because of its memory-based k nearest neighbor query algorithm. Item-Based CF algorithms have been designed to deal with the scalability problems associated with user-based CF approaches without sacrificing recommendation or prediction accuracy. However, item-based CF algorithms still suffer from the data sparsity problems. This paper presents a CF recommendation algorithm, named CFBIS (collaborative filtering based on influence sets), which is based on the concept of influence set and is a hot topic in information retrieval system. Moreover, it defines a new prediction computation method for this new recommendation mechanism. Experimental results show that the algorithm can achieve better prediction accuracy than traditional item-based CF algorithms. Furthermore, the algorithm can alleviate the dataset sparsity problem.