Similarity measures for collaborative filtering recommender systems

Collaborative filtering recommender systems evaluate users' ratings in order to give them better recommendations. One of the popular ways to make rating predictions is by using neighborhood-based models which rely on calculating the similarities between users, and use the concept that similar users will tend to rate the same items similarly. Different similarity measures were proposed in previous studies. In this paper, we present a clear study of the most used similarities (PCS, CVS, MSD, SRC, FPC, WPC and DSim) by implementing them on the same dataset, and taking into consideration different samples from this dataset. Then we evaluate these similarities using the same metrics, in order to have a better comparison and to choose the similarity measure that shows the best accuracy of prediction.