An effective similarity measure for collaborative filtering

Collaborative filtering is one of the most successful and widely used methods for automated item recommendation. The most critical component of recommender algorithm is the mechanism of finding similarities among users using item ratings data and so that items can be recommended based on the similarities. The calculation of similarities has relied on traditional vector similarity measures such as Cosine and Pearson's correlation which, however, have some problems and can't exactly express the similarity between users with the data sparsity. This paper presents a new similarity measure called PNR that utilize amended City-Block-Distance expressing the similarity between users, which focuses on improving recommendation performance of collaborative filtering recommender system under data sparsity. Empirical studies on MovieLens datasets show that our new proposed approach consistently outperforms traditional similarity measures.