A novel similarity calculation for collaborative filtering

Abstract

Collaborative filtering, one of the most successful technologies for automated product recommendation, is widely used in electronic commerce. One notable task in practical systems is to compute the similarities between users (items) which can be represented with rating vectors. There has been a variety of similarity methods according to distance and vector-based similarity computing. However, those methods, such as the Pearson correlation method and Cosine similarity method, have never been questioned about the rationality behind those original results. In this paper, we propose a new concept named fluctuation factor which refers to the count of the common rated items between two rating vectors. In addition, one feasible way is presented to remove the influence of different fluctuation factors by z-score method. Finally, 4 kinds of similarity measurements, in both user-based and item-based collaborative filtering algorithm, are combined with the concept to check the effect. After the comparison of the experiment, results demonstrate that those methods can lead to a better recommendation quality when the influence of different fluctuation factors is removed.